High-content microscopy reveals a morphological signature of bortezomib resistance

Version of Record

Accepted for publication after peer review and revision.

Download
Cite
Share
CommentOpen annotations (there are currently 0 annotations on this page).

Version of Record published: October 18, 2023 (This version)
Accepted Manuscript published: September 27, 2023 (Go to version)
Accepted: September 13, 2023
Received: July 27, 2023
Preprint posted: May 2, 2023 (Go to version)

1. Of interest
Pancreatic cancer symptom trajectories from Danish registry data and free text in electronic health records

Jessica Xin Hjaltelin, Sif Ingibergsdóttir Novitski ... Søren Brunak

Research Article Nov 21, 2023
Further reading

Abstract
Editor's evaluation
Introduction
Results
Discussion
Materials and methods
Data availability
References
Decision letter
Author response
Article and author information
Metrics

Abstract

Drug resistance is a challenge in anticancer therapy. In many cases, cancers can be resistant to the drug prior to exposure, that is, possess intrinsic drug resistance. However, we lack target-independent methods to anticipate resistance in cancer cell lines or characterize intrinsic drug resistance without a priori knowledge of its cause. We hypothesized that cell morphology could provide an unbiased readout of drug resistance. To test this hypothesis, we used HCT116 cells, a mismatch repair-deficient cancer cell line, to isolate clones that were resistant or sensitive to bortezomib, a well-characterized proteasome inhibitor and anticancer drug to which many cancer cells possess intrinsic resistance. We then expanded these clones and measured high-dimensional single-cell morphology profiles using Cell Painting, a high-content microscopy assay. Our imaging- and computation-based profiling pipeline identified morphological features that differed between resistant and sensitive cells. We used these features to generate a morphological signature of bortezomib resistance. We then employed this morphological signature to analyze a set of HCT116 clones (five resistant and five sensitive) that had not been included in the signature training dataset, and correctly predicted sensitivity to bortezomib in seven cases, in the absence of drug treatment. This signature predicted bortezomib resistance better than resistance to other drugs targeting the ubiquitin-proteasome system, indicating specificity for mechanisms of resistance to bortezomib. Our results establish a proof-of-concept framework for the unbiased analysis of drug resistance using high-content microscopy of cancer cells, in the absence of drug treatment.

Editor's evaluation

In this work, the authors have profiled the morphological signatures of the HCT116 cell line and correlated them with bortezomib treatment response, which could provide novel insight into the research of resistance in multiple myeloma from the perspective of morphology. The findings are supported by solid evidence and sufficient experimental validation.

https://doi.org/10.7554/eLife.91362.sa0

Introduction

Targeted cancer therapies often fail due to drug resistance, which makes determining the drug sensitivity of populations of cancer cells requisite for timely and effective treatment (Garraway and Jänne, 2012; Pisa and Kapoor, 2020; Vasan et al., 2019). Resistance is complex and can be categorized as acquired, manifesting in the context of prolonged treatment, or intrinsic, pre-existing in the cancer cell population (Gottesman et al., 2016). Most studies of resistance focus on the drug’s known target(s) or expression levels of drug efflux pumps (Garraway and Jänne, 2012; Gottesman et al., 2016). However, we currently lack unbiased methods of identifying drug resistance in cells especially prior to treatment.

Bortezomib is an anticancer drug commonly used to treat multiple myeloma and nearly half of multiple myeloma patients show no initial response to bortezomib therapy, indicating intrinsic resistance (Chen et al., 2011; Gonzalez-Santamarta et al., 2020; Mitsiades et al., 2004; Hideshima et al., 2001; Vincenz et al., 2013). Bortezomib resistance can be attributed to targeted mechanisms such as mutations in the bortezomib-binding pocket of the proteasome subunit (PSMB5) and overexpression of proteasome subunits (Barrio et al., 2019; Franke et al., 2012; van de Ven et al., 2008; Wacker et al., 2012) as well as non-specific mechanisms, such as upregulation of prosurvival or anti-apoptotic pathways and enhanced cell adhesion to the extracellular matrix (Gonzalez-Santamarta et al., 2020; Hideshima et al., 2007). A priori knowledge of tumor cells’ susceptibility to candidate therapeutics could aid in identifying effective treatment options, resulting in fewer relapses and failed treatments due to resistance. However, current methods for evaluating drug resistance depend on viability assays and sequencing, which may be limited in its usefulness without knowledge of the full spectrum of resistance-conferring genomic alterations (Wheler et al., 2014) or knowing specific mutations or indels in the target that suppress drug activity (Kapoor and Miller, 2017). Methods for determining tumor cell susceptibility prior to therapy are desirable.

A growing literature suggests that specific genomic alterations, treatment response, and prognosis can be predicted from conventional hematoxylin and eosin tissue slides using machine learning (Cifci et al., 2022; Lee and Jang, 2022), indicating that image data holds promise for predicting drug resistance. High-content microscopy, which uses cell-based automated microscopy to capture information-rich images, has successfully categorized small molecule inhibitors by their mechanisms and targeted pathways (Ljosa et al., 2013; Perlman et al., 2004) and shown a relationship between morphological profiles and genetic perturbations (Rohban et al., 2017), including specific mutations associated with lung cancer (Caicedo et al., 2022). This profiling method often uses high-throughput microscopy, generating a large amount of image data from which thousands of quantitative, single-cell morphological features can be extracted to characterize signals that could not be discovered using low-throughput methods and would otherwise be impossible to study by eye. However, high-content microscopy has not been used to examine the features of resistance in the absence of drug treatment.

Here, we used Cell Painting (Bray et al., 2016), a multiplex, fluorescence microscopy assay that labels eight cellular components using six stains imaged in five channels, as an unbiased method to characterize the morphological differences between bortezomib-resistant and -sensitive cancer cells. We applied a reproducible imaging- and computation-based profiling pipeline to process the images and identify a high-dimensional cell morphology signature to predict bortezomib resistance that we evaluated using machine learning best practices. This morphological signature correctly predicted the bortezomib resistance of seven out of ten clones not included in the signature training dataset. Overall, our results establish a proof-of-concept framework for identifying unbiased signatures of drug resistance using high-content microscopy. The ability to identify drug-resistant cells based on morphological features provides a valuable method for characterizing resistance in the absence of drug treatment.

Results

Isolating and capturing Cell Painting profiles for HCT116-based bortezomib-resistant clones

We first isolated and characterized drug-resistant cells (Figure 1A). To isolate drug-resistant clones, we used an approach we have described previously (Kasap et al., 2014; Wacker et al., 2012) and the HCT116 cell line. These cancer cells express multidrug resistance pumps at low levels and are mismatch repair deficient, providing a genetically heterogeneous polyclonal population of cells (Papadopoulos et al., 1994; Teraishi et al., 2005; Umar et al., 1994) allowing for isolation of drug-resistant clones in ~2–3 weeks. We hypothesized that a rapid selection of resistance could favor the isolation of clones with intrinsic resistance. To determine the appropriate drug concentrations to use in order to isolate drug-resistant clones, we performed proliferation assays on HCT116 parental cells with our drugs of interest: bortezomib, ixazomib, or CB-5083 (Figure 1—figure supplement 1A–C and Figure 1—source data 1). We also isolated bortezomib-sensitive (wild-type; WT) clones by dilution of the HCT116 parental cell line and acquired two published bortezomib-resistant clones (BZ clones A and E) both with mutations in PSMB5 identified by RNA sequencing performed in previous work (Figure 1—source data 2; Wacker et al., 2012). We characterized the bortezomib-resistant clones and found that the median lethal doses (LD50s) for bortezomib were ~2.8- to ~9-fold that of HCT116 parental cells (Figure 1—figure supplement 2B). In contrast, bortezomib-sensitive clones had LD50s for bortezomib that ranged from ~0.7- to~1.2-fold that of HCT116 parental cells (Figure 1—figure supplement 2A). Together these methods provided a total of twelve bortezomib-resistant, five ixazomib-resistant, five CB-5083-resistant, and twelve bortezomib-sensitive clones as well as HCT116 parental cells for our experiments.

Figure 1 with 2 supplements see all

Download asset Open asset

Experimental design for using Cell Painting to examine morphological profiles of drug-resistant cells.

(A) Graphic of the experimental workflow: we isolated drug-resistant clones by treating parental HCT116 cells with the desired drug and then expanded them for experiments. We isolated drug-sensitive clones by diluting HCT116 cells and then expanded them for experiments. We then performed proliferation assays on select clones to evaluate them for multidrug resistance. Next, we performed Cell Painting on both drug-resistant and -sensitive clones, using multiplexed high-throughput fluorescence microscopy of fixed cells followed by feature extraction and morphological profiling to search for features that contribute to a signature of drug resistance. (B) One representative field of view of cells labeled with six fluorescent dyes and captured in five channels used for morphological profiling with Cell Painting. Scale bars, 50 μm.

Figure 1—source data 1 LD50s for HCT116 cells. Calculated LD50s from the results of proliferation assays in (Figure 1—figure supplement 1A-C) using HCT116 parental cells and either bortezomib, ixazomib, or CB-5083.: https://cdn.elifesciences.org/articles/91362/elife-91362-fig1-data1-v2.xlsx
Download elife-91362-fig1-data1-v2.xlsx
Figure 1—source data 2 Cell line descriptions. Descriptions of HCT116 cells used in experiments. NA, not applicable; n.d., not determined; * isolated previously (Wacker et al., 2012).: https://cdn.elifesciences.org/articles/91362/elife-91362-fig1-data2-v2.xlsx
Download elife-91362-fig1-data2-v2.xlsx

To screen for multidrug resistance, which might convolute a specific signature of bortezomib resistance, we measured proliferation of the bortezomib-resistant and -sensitive clones in the presence of two drugs with different mechanisms of action: taxol (a microtubule poison) and mitoxantrone (a topoisomerase inhibitor) (Liu, 1989). Bortezomib-resistant and -sensitive clones treated with taxol had LD50s ranging from ~0.6- to~1.9-fold that of HCT116 parental cells (Figure 1—figure supplement 2C and D). Treating cells with mitoxantrone, we found that the bortezomib-sensitive clones (Figure 1—figure supplement 2E) and most of the bortezomib-resistant clones had similar LD50s (Figure 1—figure supplement 2F). There was one exception (BZ06) that had an LD50 nearly 14-fold higher than that of HCT116 parental cells.

We next applied the Cell Painting assay to all these drug-sensitive and -resistant clones. Cell Painting captures signal in five imaging channels from six fluorescent dyes that stain cells for eight cellular components including mitochondria, actin, Golgi, plasma membrane, cytoplasmic RNA, nucleoli, endoplasmic reticulum, and DNA (Figure 1B; Bray et al., 2016). With these images, we used CellProfiler (Stirling et al., 2021) to extract single-cell morphological features from individual cells. The signal from each of the five channels was analyzed in the nucleus, cytoplasm, and total cell and characterized based on features (object parameters) such as signal intensity, shape of the object, texture of the staining pattern, etc. yielding a total of ~3500 features. These cellular features were combined and analyzed on a per well basis (well profiles) and then compared across cells and experimental conditions to determine whether morphological features of drug resistance could be reliably detected in the absence of drug treatment.

A subset of morphological features contribute to the signature of bortezomib resistance

To examine whether there were any clear qualitative morphological differences between bortezomib-resistant and -sensitive cells we chose HCT116 parental cells, bortezomib-sensitive clones WT01-WT05, and bortezomib-resistant clones A, E, and BZ01-BZ05 for our initial studies. We treated cells with 0.1% DMSO (to allow for comparison with future experiments using drug-treated cells) and performed Cell Painting, staining fixed cells and imaging as per the published protocol (Bray et al., 2016). We observed cellular heterogeneity within each clone as well as between clones with similar bortezomib sensitivities (Figure 2A and Figure 2—figure supplement 1). This heterogeneity obscured any potential morphological differences between clones and prevented us from qualitatively distinguishing bortezomib-resistant from -sensitive clones by eye, supporting the need for high-content quantitative analysis.

Figure 2 with 7 supplements see all

Download asset Open asset

A subset of morphological features contributes to the signature of bortezomib resistance.

(A) Representative fixed fluorescence microscopy images of two bortezomib-sensitive (WT02 and WT03) and two bortezomib-resistant (BZ02 and BZ03) clones stained and imaged as per the Cell Painting protocol. Channels are labeled as mito (mitochondria; magenta), AGP (actin, golgi, plasma membrane; yellow), RNA (ribonucleic acid; green), ER (endoplasmic reticulum; orange), and DNA (deoxyribonucleic acid; blue). See Figure 2—figure supplement 1 for single-channel images. Scale bars, 50 μm. (B) Volcano plot of the variability of morphological features (β) by resistance status. Y-axis -log₁₀p values are from Tukey’s Honestly Significant Difference test score (see Materials and methods). Red circles are features included in the final signature of resistance and gray circles are features excluded from the final signature. Features above the red dashed line (-log₁₀[0.05/number of unique features]) were considered significantly varying and those that had not been excluded as technical variables (Figure 2—figure supplement 3) were included in the signature of bortezomib resistance. n = 6 independent experiments (biological replicates).

Figure 2—source data 1 A subset of CellProfiler features contribute to the signature of bortezomib resistance. List of 45 CellProfiler features that, after the exclusion of technical variables, were found to contribute to the signature of bortezomib resistance. Feature details can be found in the documentation for CellProfiler (see Materials and methods).: https://cdn.elifesciences.org/articles/91362/elife-91362-fig2-data1-v2.xlsx
Download elife-91362-fig2-data1-v2.xlsx

We then pre-processed profiles to remove low-variance and highly correlated features, and population-averaged single cell measurements at the well level to generate well profiles (see Materials and methods). The morphological profiles of bortezomib-resistant and bortezomib-sensitive cells did not cleanly distinguish clones based on bortezomib resistance (Figure 2—figure supplement 2A). We saw a similar failure to distinguish clones based on bortezomib sensitivity after a short, 4 hr treatment with 7 nM bortezomib (Figure 2—figure supplement 2B), suggesting that if there is a morphological difference between bortezomib-resistant and -sensitive cells, further feature refinement would be needed for its identification.

Each observed morphological measurement results from a combination of both technical and biological variables. It is therefore important to control and test for technical variables as these can confound subtle biologically relevant signatures. Using bortezomib-sensitive clones WT01-05 and bortezomib-resistant clones BZ01-05 to quantify and reduce the impact of technical variables, we fit a linear model to each morphological feature adjusting for technical variables (experimental run/batch, incubation time, cell count/density, clone ID) and biological variables (resistance status) (see Materials and methods). We then discarded morphological features with variances that correlated with experimental run (batch), incubation time (4 or 13 hours with 0.1% DMSO), cell density, or those features that varied between two or more pairs of bortezomib-sensitive clones (clone ID) (Figure 2—figure supplement 3A-E). Of the remaining morphological features, we only considered those that varied based on the resistance status of a clone (Figure 2B). This resulted in 45 morphological features that significantly contributed to a clone’s bortezomib resistance (Figure 2—source data 1). We used these 45 features to compute a rank-based resistance score or ‘Bortezomib Signature’ for each well profile based on the direction-sensitive method called singscore (Foroutan et al., 2018). Singscore ranks these 45 resistance-related features on a per sample basis and calculates a normalized score between –1 and 1, with higher values expected for bortezomib-resistant clones and lower values expected for bortezomib-sensitive clones. With the exception of some texture-based features, the Bortezomib Signature features were largely independent, displaying low pairwise correlation (Figure 2—figure supplement 4). We then examined the grouping of features across compartments and channels and found radial distribution features were higher in resistant cells (Figure 2—figure supplement 5). Anticipating well location as a possible technical artifact, we plated our cells in a repeating serpentine pattern, ensuring that each clone would be imaged in multiple locations on a plate (Figure 2—figure supplement 6A). We found that the pattern of Bortezomib Signatures did not correspond to well position relative to the plate (Figure 2—figure supplement 6B), indicating that the well position for each clone was not strongly contributing to its Bortezomib Signature. In addition, we found that the Bortezomib Signature correlated with the resistance status of clones and not technical variables (Figure 2—figure supplement 7). These data suggest that our analysis pipeline and signature building process minimized technical artifacts.

Cell morphology predicts the bortezomib response of multiple clones across datasets

To evaluate the performance of the Bortezomib Signature, we used machine learning best practices, separating our data into training, validation, test, and holdout datasets (Figure 3—source data 1; see Materials and methods). The data initially used to create the Bortezomib Signature, which included well-based morphological profiles from clones WT01-05 and BZ01-05, was designated as the training dataset. The validation dataset was composed of well profiles from clones WT01-05 and BZ01-05 that were not used to generate the Bortezomib Signature but were collected on the same plates as the well profiles used for the training dataset. The test dataset was composed of well profiles from HCT116 parental cells and bortezomib-resistant clones A and E; these cells were not included in the training dataset, but their well profiles were collected on the same plates as those used for the training dataset. The holdout dataset was a separate plate and contained HCT116 parental cells, bortezomib-sensitive clones WT01-05, and bortezomib-resistant clones A, E, and BZ01-05. These datasets allowed us to test generalizability across clones and plates for the Bortezomib Signature.

We next examined whether the Bortezomib Signature was able to predict the bortezomib resistance of a clone based on morphological profiling data (Figure 3A–E and Figure 3—figure supplement 1A and B). We called the clone bortezomib-resistant if the median Bortezomib Signature of all replicate well profiles was greater than zero and bortezomib-sensitive if the median Bortezomib Signature was less than zero. In the training dataset, the Bortezomib Signature correctly predicted the bortezomib resistance of all ten clones, with median Bortezomib Signatures for eight out of ten clones beyond the 95% confidence interval for the randomly permuted data (Figure 3A). The accuracy of the Bortezomib Signature was 88% while the average precision was 81% for the training dataset (Figure 3—figure supplement 1A and B) (see Materials and methods). The signature performed similarly well in the validation dataset (Figure 3B), with an accuracy of 92% and an average precision of 89% (Figure 3—figure supplement 1A and B). In the test dataset the Bortezomib Signature correctly predicted the bortezomib resistance of all clones, although only HCT116 parental cells had a median Bortezomib Signature outside the 95% confidence interval for the randomly permuted data (Figure 3C). The test dataset had an accuracy of 80% and an average precision of 68% (Figure 3—figure supplement 1A and B). Similarly, in the holdout dataset the Bortezomib Signature had an accuracy of 78% and an average precision of 69% (Figure 3—figure supplement 1A and B), and correctly predicted the bortezomib resistance of twelve out of thirteen clones, with WT01 misclassified as bortezomib-resistant (Figure 3D). In the holdout dataset, four of the twelve correctly characterized clones had median Bortezomib Signatures outside the 95% confidence interval for the randomly permuted data. The Bortezomib Signature performed better than random chance in all testing conditions, as demonstrated by comparison with the mean accuracy and average precision for the randomly shuffled data (Figure 3—figure supplement 1A and B), and as reflected in receiver operating characteristic (ROC) curves, which describe the classification trade-off between true positive and false positive rates in predicting bortezomib resistance (Figure 3E). We then compared our linear-modeling approach to feature selection against other feature spaces and found that the Bortezomib Signature clusters same-type clones (bortezomib-resistant vs. bortezomib-sensitive) with higher enrichment compared to the full feature space, standard feature selection (see Materials and methods), or a random selection of 45 features (Figure 3—figure supplement 2). These data are consistent with the Bortezomib Signature being able to distinguish bortezomib-resistant from -sensitive clones better than random chance across datasets.

Figure 3 with 2 supplements see all

Download asset Open asset

Cell morphology predicts the bortezomib sensitivity of clones across datasets.

Box plots of Bortezomib Signatures for clones in the (A) training, (B) validation, (C) test, and (D) holdout datasets. Plots show values for individual well profiles (points), range (error bars), 25th and 75th percentiles (box boundaries), and median. Dashed vertical black line is Bortezomib Signature = 0, dashed vertical red lines are the 95% confidence interval for Bortezomib Signatures of 1000 random permutations of the data. (E) ROC curves for the performance of the Bortezomib Signature on the indicated dataset (solid line) or its shuffled counterpart (dashed line). Datasets are designated by color: training (magenta), validation (orange), test (teal), and holdout (purple). Colored points are the corresponding false positive and true positive rates at the absolute minimum thresholds for each respective dataset. Black dotted line is the identity line where false positive rate = true positive rate. AUROC values reported for data and shuffled data. See Figure 3—source data 1 for breakdown of profiles and experiments per dataset.

Figure 3—source data 1 Datasets for Bortezomib Signature generation and evaluation. Breakdown of the number of well-based profiles from cells in the training, validation, test, and holdout datasets used to generate and evaluate the Bortezomib Signature. Data from 6 independent experiments contributed to the training, validation, and test datasets. The holdout dataset is one separate experiment. Note: the Bortezomib Signature was evaluated on clones not included in the training dataset (WT10, WT12-15, BZ06-10) in a separate analysis, see Figure 5.: https://cdn.elifesciences.org/articles/91362/elife-91362-fig3-data1-v2.xlsx
Download elife-91362-fig3-data1-v2.xlsx

Bortezomib Signature has moderate specificity for bortezomib over other ubiquitin-proteasome system inhibitors

To examine whether the Bortezomib Signature was specific to the drug bortezomib or was a general signature of UPS-targeting drug resistance we performed Cell Painting on HCT116 clones that were resistant to either ixazomib (another proteasome inhibitor that targets the PSMB5 subunit) or CB-5083 (a p97 inhibitor that acts upstream of the proteasome). If the Bortezomib Signature were a general signature of UPS-targeting drug resistance, we would expect it to perform equally well at characterizing the drug sensitivity of bortezomib-, ixazomib-, and CB-5083-resistant clones. The Bortezomib Signature performed better than chance at identifying ixazomib-resistant and CB-5083-resistant clones (Figure 4A), correctly identifying four of five ixazomib-resistant clones (Figure 4B) and three of five CB-5083-resistant clones (Figure 4C). However, only two of the four correctly identified ixazomib-resistant clones and one of the three CB-5083-resistant clones had median Bortezomib Signatures outside the 95% confidence interval of the randomly permuted data. The area under the ROC (AUROC) curve for ixazomib-resistant and CB-5083-resistant clones (0.63 and 0.60, respectively) was lower than those calculated for the training, validation, test, and holdout datasets for bortezomib-resistant clones. In addition, many of the Bortezomib Signatures for well profiles of ixazomib- and CB-5083-resistant clones, particularly those for CB-5083-resistant clones, landed within the 95% confidence interval of the randomly permuted data. These results suggest that the Bortezomib Signature is not a general signature of UPS-targeting drug resistance, nor of our cloning protocol, and instead has some specificity for bortezomib.

Figure 4

Download asset Open asset

Bortezomib Signature has limited ability to characterize clones resistant to other UPS-targeting drugs.

(A) ROC curves for ixazomib-resistant (magenta) and CB-5083-resistant (blue) experimental data. Colored solid lines are the actual data while colored dashed lines are the shuffled data for each set of clones. Colored points are the corresponding false positive and true positive rates at the absolute minimum thresholds for each respective cell type. Black dotted line is the identity line where false positive rate = true positive rate. AUROC reported for the data and shuffled data. Box plots of Bortezomib Signatures for (B) ixazomib-resistant and bortezomib-sensitive clones (n = 18 profiles, 3 independent experiments) and (C) CB-5083-resistant and bortezomib-sensitive clones (n = 24 profiles, 4 independent experiments). Plots show values for individual well profiles (points), range (error bars), 25th and 75th percentiles (box boundaries), and median. Dashed vertical black line is Bortezomib Signature = 0, dashed vertical red lines are the 95% confidence interval for Bortezomib Signatures of 1000 random permutations of the data.

Bortezomib Signature characterizes bortezomib sensitivity of clones not included in the training dataset

To examine whether the Bortezomib Signature could correctly characterize the bortezomib sensitivity of clones not included in the training, validation, holdout, or test datasets, we imaged bortezomib-sensitive (WT10, WT12-WT15) and bortezomib-resistant clones (BZ06-BZ10) using the Cell Painting protocol. The Bortezomib Signature had an AUROC of 0.75, compared to 0.55 for the shuffled data (Figure 5A) and correctly characterized the bortezomib resistance of HCT116 parental cells and our bortezomib-resistant clones A and E, which we included as controls (Figure 5B). The Bortezomib Signature also correctly characterized the bortezomib resistance of four out of five bortezomib-resistant clones and three out of five bortezomib-sensitive clones not included in the training dataset (Figure 5C). In addition, the majority of Bortezomib Signatures for these well profiles landed outside the 95% confidence interval for the randomly permuted data. While the Bortezomib Signature correctly characterized the bortezomib sensitivity of most clones, it consistently misclassified others (WT10, WT15, and BZ06; Figure 5—figure supplement 1A). Proliferation assays conducted in earlier experiments showed that WT10 and WT15 were sensitive to bortezomib while BZ06 was resistant (Figure 1—figure supplement 2A and B). By comparing these incorrect predictions with high-confidence correct predictions, we observed differences that varied by clone type, suggesting unique morphology may be driving each of these misclassifications (Figure 5—figure supplement 1B and C). These results are consistent with the Bortezomib Signature being generalizable to clones not included in the training dataset and suggest that morphological profiling has the potential to identify bortezomib-resistant clones based on the morphological features of cells in the absence of drug treatment.

Figure 5 with 1 supplement see all

Download asset Open asset

Bortezomib Signature correctly characterizes bortezomib sensitivity of seven out of ten clones not included in the training, validation, test, or holdout datasets.

(A) ROC curve for the Bortezomib Signature of clones in (B) and (C) (solid line) and shuffled data (dashed line). Colored point is the corresponding false positive and true positive rate at the absolute minimum threshold. Black dashed line is the identity line where false positive rate = true positive rate. AUROC reported for the data and shuffled data. (B) Box plots of Bortezomib Signatures for bortezomib-resistant clones A and E (n = 16 profiles each) and HCT116 parental cells (n = 48 profiles). (C) Box plots of Bortezomib Signatures for bortezomib-sensitive clones WT10, WT12-15 (n = 16 profiles each) and bortezomib-resistant clones BZ06-10 (n = 8 profiles each). Plots show values for individual well profiles (points), range (error bars), 25th and 75th percentiles (box boundaries), and median. Dashed vertical black line is Bortezomib Signature = 0, dashed vertical red lines are the 95% confidence interval for Bortezomib Signatures of 1000 random permutations of the data. 4 independent experiments (biological replicates).

Discussion

We used Cell Painting, a high-throughput, high-content image acquisition and analysis assay, as a target-independent method to capture the morphological profiles of cells that were either resistant or sensitive to the UPS-targeting anticancer drug, bortezomib, in the absence of drug treatment. After processing profiles to reduce the impact of technical variables, we generated a signature of bortezomib resistance and characterized the performance of this signature using machine learning best practices. This Bortezomib Signature correctly predicted the bortezomib resistance of seven out of ten clones not included in the training dataset and was more specific to bortezomib-resistance given its limited ability to identify clones that were resistant to other UPS-targeting drugs, ixazomib and CB-5083. All three drugs tested target the UPS, however bortezomib and ixazomib both bind the same subunit of the proteasome, albeit with potentially non-overlapping spectrums of off-targets (Baggish et al., 2010). These data suggest that the Bortezomib Signature may be specific to the drug bortezomib and not a general signature of resistance to UPS-targeting drugs. The Bortezomib Signature is conceptually similar to the on-disease/off-disease score (Heiser et al., 2020). Both require three phenotypic measurements: a target phenotype representing ideal, a disease phenotype, and a new phenotype to classify. However, our approach is technically different (non-parametric compared to linear projection) and our goals are different (phenotypic classification compared to perturbation alignment). Other methods also enable phenotype labeling, but they focus on single-sample annotation without regard to a target phenotype (Nyffeler et al., 2020; Rohban et al., 2017; Simm et al., 2018; Wawer et al., 2014). Our work demonstrates that there are morphological features of drug resistance in cells that can be identified using Cell Painting and provides a reproducible pipeline for generating morphological signatures of drug resistance.

The Bortezomib Signature’s performance was not perfect; it misclassified three clones not included in the training dataset. Interestingly, one of the misclassified clones (BZ06) had reduced sensitivity to mitoxantrone as well as bortezomib. Given the considerable genetic heterogeneity in this mismatch repair-deficient HCT116 cell line (Glaab and Tindall, 1997; Umar et al., 1994), it is possible that some misclassified clones have orthogonal mechanisms of resistance or unrelated mutations contributing to their morphological profiles. Targeted sequencing of the PSMB5 proteasome subunit in bortezomib-resistant clones may provide information regarding origins of these misclassifications, as multiple mutations have been identified in bortezomib-resistant clones (Wacker et al., 2012). Determining the underlying reason for the misclassification of bortezomib-sensitive clones would require further studies.

Together, our work demonstrates the potential for morphological profiling with Cell Painting to be used as an unbiased method to characterize resistance in the absence of drug treatment. Our results indicate that different mechanisms of bortezomib resistance may generate distinct morphological profiles; with larger and broader training datasets, it may be possible to identify signatures for distinct mechanisms of bortezomib resistance as well as signatures of resistance to other drugs. Although it is unclear whether this method can be extended to patient samples, where identifying intrinsic drug resistance in cells prior to treatment has the potential to improve targeted cancer therapy, our results are an encouraging proof of concept. We expect that further refinement may develop Cell Painting as a tool for identifying drug-resistant cells, perhaps even guiding strategies to overcome resistance.

Materials and methods

Key resources table

Reagent type (species) or resource	Designation	Source or reference	Identifiers
Cell line (Homo sapiens)	HCT116	ATCC	CCL-247; RRID:CVCL_0291
Chemical compound, drug	bortezomib	LC Laboratories	cat # 1408
Chemical compound, drug	taxol	Sigma	cat # T7402
Chemical compound, drug	mitoxantrone	TOCRIS	cat # 4250
Chemical compound, drug	ixazomib	ApexBio	cat # A4007 batch 2
Chemical compound, drug	CB-5083	MedChemExpress	cat # HY-12861 batch 19554
Other	MitoTracker Deep Red	Invitrogen	cat # M22426
Other	Phalloidin AF568	Invitrogen	cat # A12380
Other	Concanavalin A AF488	Invitrogen	cat # C11252
Other	Hoechst 33342	ThermoFisher	cat # 62249
Other	Hoechst 33342	Invitrogen	cat # H3570
Other	Wheat-germ agglutinin AF555	Invitrogen	cat # W32464
Other	SYTO14 Green	Invitrogen	cat # S7576

Cell culture

Request a detailed protocol

HCT116 cells (RRID: CVCL_0291) were purchased from ATCC (CCL-247) and were maintained in McCoy’s 5 A Medium (Gibco) supplemented with 10% (v/v) FBS (Sigma) and cultured at 5% CO2 and 37 °C. Cells were determined to be mycoplasma-free using a PCR-based method (Uphoff and Drexler, 2013). Bortezomib-resistant, ixazomib-resistant, and CB-5083-resistant clones were isolated as previously described (Wacker et al., 2012). Briefly, HCT116 cells were plated in 150 mm dishes and grown in the presence of the desired drug at a concentration that resulted in the death of the majority of cells (selection concentrations: bortezomib [LC Laboratories], 12 nM; ixazomib [ApexBio], 150 nM; CB-5083 [MedChemExpress], 600 and 700 nM). The locations of single surviving cells were identified using brightfield microscopy and marked on the plate. Cells were allowed to expand into colonies over 2–4 weeks and colonies were isolated using cloning rings. Bortezomib-sensitive clones were generated by diluting HCT116 cells into 96-well plates and wells containing single cells as identified by brightfield microscopy were marked. Colonies that grew in these marked wells were expanded and used for experiments. Bortezomib-resistant clones A and E were provided by the Kapoor laboratory having been previously published (Wacker et al., 2012).

Proliferation assays

Request a detailed protocol

Cell proliferation was evaluated using an Alamar Blue assay (O’Brien et al., 2000). Briefly, cells were plated in duplicate or triplicate in sterile 96-well Clear Microplates (Falcon) under described culture conditions, with 1000 cells in 100 μL per well and allowed to adhere overnight. After cells attached to the plate, 50 μL of media containing drug (bortezomib, ixazomib, CB-5083, taxol [Sigma], or mitoxantrone [TOCRIS]) was added to each well. The final DMSO concentration was 0.1% for all wells, including three wells with media only as background measurements. Plates were incubated for 72 hr at 5% CO2 and 37 °C before adding Alamar Blue (resazurin sodium salt, final concentration 50 μM). Cells were incubated with Alamar Blue for 3–4 hr and then imaged with a Synergy Neo plate reader using excitation: 550 nm and emission: 590 nm (Agilent). The average plate background (media only with 0.1% DMSO) was subtracted from the average fluorescence for each condition and the resulting value was normalized by dividing by the background-subtracted value for each condition’s control (cells treated with 0.1% DMSO). Using the data from our proliferation assays, we calculated the median lethal dose (LD50) for each of our drugs of interest by fitting data of normalized growth vs. log[drug concentration] to a sigmoidal dose-response curve using GraphPad Prism (v.9.2.0) (Figure 1—source data 1).

Cell Painting

Request a detailed protocol

High-throughput imaging was performed according to the published Cell Painting protocol (Bray et al., 2016). HCT116 cells were plated at concentrations of 2.5 or 5x10³ cells/mL in 96-well glass-bottomed tissue culture dishes (Greiner Bio-One) and allowed to adhere for 48–72 hr prior to fixation. At either 4 or 13 hr prior to fixation, cells were treated with either 0.1% DMSO or 7 nM bortezomib and 30 min prior to fixation cells were treated with MitoTracker Deep Red (500 nM, Invitrogen). 16% paraformaldehyde (EMS) was added to each well for a final concentration of 3.2% and cells were fixed in the dark at room temperature for 20 min. Wells were washed with HBSS (Invitrogen), permeabilized with 0.1% Triton-X for 15 min, and then washed twice with HBSS before incubating with staining solution (5 U/mL phalloidin AF568 [Invitrogen], 100 μg/mL concanavalin A AF488 [Invitrogen], 5 μg/mL Hoechst 33342 [ThermoFisher or Invitrogen], 1.5 μg/mL wheat-germ agglutinin AF555 [Invitrogen], 3 μM SYTO14 Green [Invitrogen], and 1% bovine serum albumin [BioWorld] in HBSS) in the dark for 30 minutes. Wells were then washed twice with HBSS and imaged using an ImageXpress high-content imaging system (Molecular Devices) with a 20x0.45 NA S Plan Fluor ELWD objective (Nikon) and captured with a Zyla 5.5 sCMOS detector (Andor Technology). Each well was imaged at 12–17 non-overlapping sites in five channels using Semrock filters (mito: Cy5-4040B-NTE-ZERO, AGP: TxRed-4040C-NTE-ZERO, RNA: Cy3-4040C-NTE-ZERO, ER: FITC-3540C-NTE-ZERO, and DNA: DAPI-5060C-NTE-ZERO).

Image data processing

Request a detailed protocol

We used CellProfiler versions 3.1.8 and 3.1.9 (Caicedo et al., 2018) to perform the standard processing pipeline of illumination correction, single-cell segmentation, and morphology feature extraction. We performed per-plate illumination correction to adjust for uneven background intensity that commonly impacts microscopy images. We also developed per-plate analysis pipelines for single-cell segmentation and feature extraction. We extracted 3,528 total cell morphology features from all 25,331,572 cells we captured in this experiment. The 3528 features represent stain intensities, stain co-localization, textures, areas, and other patterns extracted from all five imaging channels and different segmentation objects (nuclei, cytoplasm, total cells). Feature details are described in the documentation for CellProfiler (https://cellprofiler-manual.s3.amazonaws.com/CellProfiler-3.1.9/help/output_measurements.html). We include all image analysis pipelines at https://github.com/broadinstitute/profiling-resistance-mechanisms (Way et al., 2023).

Following feature extraction, we applied an image-based analysis pipeline to generate our final analytical set of treatment profiles (Caicedo et al., 2017). We first used cytominer-database to ingest all single-cell, per-compartment CellProfiler output files (comma separated) to clean column names, evaluate integrity of CellProfiler output CSVs, and output single-cell SQLite files for downstream processing. Next, we used pycytominer (github hash c1aa34b641b4e07eb5cbd424166f31355abdbd4d) for all image-based profiling pipeline steps. In the first step, we median aggregated all single cells to form well-level profiles (Way et al., 2022). Next, we performed a step called annotation, which merges the consistent platemap metadata with the well-level profiles. Third, we performed standard z-score normalization to ensure all features are measured on the same scale with zero mean and unit variance. Lastly, we performed feature selection, which removed features with low variance, high correlation (>0.9 Pearson correlation), features with missing values, features on our blocklist (Way, 2020), and features with outliers greater than 15 standard deviations, which we suspected were measured in error. For developing our final analytical datasets (see section, Constructing the resistance signature) we performed normalization within each plate but performed a combined feature selection across all plates per analytical dataset using the same procedures described previously, which resulted in 782 features. We applied the same pipeline uniformly across all plates. We did not detect large differences in variance that could be attributed to well position and batch and therefore did not apply batch effect correction. Our full image data processing pipeline is publicly available at https://github.com/broadinstitute/profiling-resistance-mechanisms (Way et al., 2023).

Constructing the resistance signature

Request a detailed protocol

After processing all images and forming normalized and feature selected profiles per well, we performed additional analyses to explore the results and discover a morphology profile of bortezomib resistance. We performed initial comparisons of morphological profiles using Morpheus (https://software.broadinstitute.org/morpheus) to create similarity matrix heatmaps.

We aimed to discover a generalizable signature of bortezomib resistance from the normalized profiles. Our approach was to identify features that were significantly different by resistance status and not significantly impacted by technical covariates. To do so, we carefully constructed datasets for training and evaluating signature performance (Figure 3—source data 1). To generate our training dataset, we selected a set of six plates consisting of five bortezomib-sensitive and five bortezomib-resistant clones that we collected on three different days, which showed high within-replicate reproducibility (technical replicates; data not shown). A seventh plate was held-out from signature generation in order to analyze generalizability between plates (holdout dataset). We evaluated the signature in five scenarios: (1) clones held-out on the same plates used to generate the training dataset (validation dataset, Figure 3B), (2) HCT116 parental cells and clones with PSMB5 mutations known to confer resistance to bortezomib (test dataset, Figure 3C; Wacker et al., 2012), (3) clones held-out on a separate plate (holdout dataset, Figure 3D), (4) clones selected to be resistant to other drugs (ixazomib and CB-5083, Figure 4), and (5) bortezomib-resistant clones not included in the training dataset (Figure 5). All cells on these plates were incubated with 0.1% DMSO for either 4 or 13 hours.

Using data from the ten clones in our training dataset (20–21 technical replicates per clone, see Figure 3—source data 1), we fit two linear models for all 782 CellProfiler features (post normalization and feature selection) to discover features that varied strongly with technical variants (batch, cell count, incubation time, or clone ID) and features that varied strongly with resistance status (bortezomib-sensitive or bortezomib-resistant). In the first linear model, we quantified the per feature variance contribution of resistance status (β_{resistance status}), batch (β_batch), incubation time (β_{incubation time}), and clone (β_{clone ID}) to each CellProfiler feature (Y_j) where $ε$ is the error term:

Y_{j} = β_{i n t e r c e p t} + β_{r e s i s t a n c e s t a t u s} X_{r e s i s t a n c e s t a t u s} + β_{b a t c h} X_{b a t c h} + β_{i n c u b a t i o n t i m e} X_{i n c u b a t i o n t i m e} + β_{c l o n e I D} X_{c l o n e I D} + ε

Fitting this model produced a goodness of fit R² value per feature and individual beta coefficients per covariate. Furthermore, we calculated a Tukey’s Honestly Significant Difference (Tukey’s HSD) post hoc test per model to determine which categorical covariate comparison contributed to a significant finding and to control for within-covariate-group multiple comparisons through a family-wise error rate (FWER) adjustment that accounts for different within-group sizes (e.g. three different batches in the comparison, two different resistance statuses, etc.) (Tukey, 1949).

Separately, we fit another linear model on continuous features to adjust for features that were significantly impacted by well confluence (β_{cell count}) as it is expected that dense wells will impact certain morphology features, which we want to avoid in the resistance signature:

Y_{j} = β_{i n t e r c e p t} + β_{r e s i s t a n c e s t a t u s} X_{r e s i s t a n c e s t a t u s} + β_{c e l l c o u n t} X_{c e l l c o u n t} + ε

By fitting these models, we quantified the variance contribution of four technical covariates (incubation time, batch, clone ID, and cell count) and our biological variable of interest (resistance status), and, based on the first linear model, we have knowledge of which group comparisons were significant in each category (via Tukey’s HSD). We further refined the signature by filtering features that did not pass a Bonferonni adjusted alpha threshold calculated across all 782 features (0.05/782 = 6.4x10^–6).

We next applied an exclusion criterion to isolate features that contributed to resistance status. We excluded features that were significantly different across incubation times, batches, and cell counts. We also excluded features that were different within clone type (features varying between two or more pairs of bortezomib-sensitive clones) to reduce the contribution of features that may mark generic inter-clone differences nonspecific to resistance status. This procedure resulted in a total of 45 features that were significantly different by resistance status and not significantly impacted by any of the technical covariates we considered. Of the 45 features, 14 had higher values in resistant clones and 31 had lower values in resistant clones (Figure 2—source data 1).

We were also interested in comparing the ability of different feature spaces to cluster clones of the same type (resistant vs. sensitive). This analysis would determine if the Bortezomib Signature features, which we derived using linear modeling to isolate biological from technical variables, had a greater ability to cluster. We compared the Bortezomib Signature against three other feature spaces: (1) the full feature space, (2) standard feature selection (see Image data processing methods), and (3) 45 randomly selected features. We performed two analyses using these four feature spaces including Uniform Manifold Approximation and Projection (UMAP) (McInnes et al., 2018) and k-means clustering. For UMAP, we used default umap-learn parameters to identify two UMAP coordinates per feature space. We then visualized the clusters by their resistance status and Bortezomib Signature score. The UMAP analysis represents a qualitative analysis. Next, we applied k-means clustering with 25 initializations across a range of 2–14 clusters (k). Prior to clustering and for each feature space, we applied principal component analysis (PCA) and transformed each feature space into 30 principal components. This step was necessary to compare k-means clustering metrics, which are sensitive to the feature space dimensionality. We applied a Fisher’s exact test to each cluster using a two-by-two contingency matrix that specified cluster membership for each clone classification (resistant vs. sensitive). We visualized the mean odds ratio and max cluster odds ratio for each feature space across k. A high odds ratio tells us that the feature space effectively clusters clones of the same resistance status. Lastly, we calculated Silhouette width (the average proximity between samples in one cluster to the second nearest cluster) for each feature space across k.

Applying the signature

Request a detailed protocol

We used the singscore method (Foroutan et al., 2018) to characterize individual profiles of different clones as either bortezomib-resistant or bortezomib-sensitive. Singscore is a rank-based method that was originally developed to analyze the direction and significance of previously defined molecular signatures on transcriptomic data. The method calculates a two-part signature for each direction list (14 up and 31 down) and calculates an internal rank per profile of how highly ranked and lowly ranked each of the up and down features are, respectively. The method then adds the up and down rank scores to form a total singscore per sample, which ranges between –1 and 1 and represents a rank-based normalized concordance score that can be directly compared across profiles that may have been normalized differently. Therefore, the score is robust to outliers and different normalization procedures. In addition to calculating the singscore per sample, we also calculated singscore with 1,000 random permutations, in which we randomly shuffled feature rankings to derive a range in which a sample may be scored by chance. Some profiles were consistently predicted incorrectly with high confidence but in the opposite direction (see Figure 5—figure supplement 1). For a well-level profile to be categorized as high-confidence (in either the correct or incorrect directions), it needed to score beyond the 95% confidence interval of the randomly permuted data range. For example, a high-confidence incorrect resistant profile would have a Bortezomib Signature below 95% confidence interval of the randomly permuted data. To evaluate the features driving the differences in these samples, we applied two-sample Kolmogorov–Smirnov (KS) tests per Bortezomib Signature feature. We applied these tests to two separate groups: (1) misclassified bortezomib-sensitive vs. high-confidence accurate bortezomib-sensitive and (2) misclassified bortezomib-resistant vs. high-confidence accurate bortezomib-resistant.

Signature evaluation

Request a detailed protocol

We used several metrics to evaluate signature quality across five different evaluation scenarios (validation, test, holdout, other UPS-targeting drugs, and clones not included in the training dataset [biological replicates]). Because we are measuring a binary decision in a balanced dataset (roughly the same amount of positive as negative classes), we used accuracy (total correct / total chances) to quantify performance. We also calculated mean average precision using sci-kit learn, averaging over samples along the precision recall curve (Varoquaux et al., 2011), which is a measure of separation between the two resistance classes (higher being more separation). We also calculated receiver operating characteristic (ROC) curves and area under the ROC curve (AUROC) using sci-kit learn. AUROC compares the ability to distinguish positive samples across signatures.

Data availability

All data generated during this study are provided in the dataset cpg0028-kelley-resistance, available in the Cell Painting Gallery on the Registry of Open Data on AWS (https://registry.opendata.aws/cellpainting-gallery/). Processed data, source data files, and code to reproduce this analysis are available at https://github.com/broadinstitute/profiling-resistance-mechanisms (Way et al., 2023).

The following data sets were generated

1. Kelley ME
2. Berman AY
3. Stirling DR
4. Cimini BA
5. Han Y
6. Singh S
7. Carpenter AE
8. Kapoor TM
9. Way GP
(2023) Registry of Open Data on AWS
ID cpg0028-kelley-resistance/. Cell Painting images to produce a high-content phenotypic signature of Bortezomib resistance.

https://registry.opendata.aws/cellpainting-gallery/

References

1. Baggish AL
2. Hutter AM
3. Wang F
4. Yared K
5. Weiner RB
6. Kupperman E
7. Picard MH
8. Wood MJ
(2010) Cardiovascular screening in college athletes with and without electrocardiography: A cross-sectional study
Annals of Internal Medicine 152:269–275.
https://doi.org/10.7326/0003-4819-152-5-201003020-00004
- PubMed
- Google Scholar
1. Barrio S
2. Stühmer T
3. Da-Viá M
4. Barrio-Garcia C
5. Lehners N
6. Besse A
7. Cuenca I
8. Garitano-Trojaola A
9. Fink S
10. Leich E
11. Chatterjee M
12. Driessen C
13. Martinez-Lopez J
14. Rosenwald A
15. Beckmann R
16. Bargou RC
17. Braggio E
18. Stewart AK
19. Raab MS
20. Einsele H
21. Kortüm KM
(2019) Spectrum and functional validation of PSMB5 mutations in multiple myeloma
Leukemia 33:447–456.
https://doi.org/10.1038/s41375-018-0216-8
- PubMed
- Google Scholar
1. Bray MA
2. Singh S
3. Han H
4. Davis CT
5. Borgeson B
6. Hartland C
7. Kost-Alimova M
8. Gustafsdottir SM
9. Gibson CC
10. Carpenter AE
(2016) Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes
Nature Protocols 11:1757–1774.
https://doi.org/10.1038/nprot.2016.105
- PubMed
- Google Scholar
1. Caicedo JC
2. Cooper S
3. Heigwer F
4. Warchal S
5. Qiu P
6. Molnar C
7. Vasilevich AS
8. Barry JD
9. Bansal HS
10. Kraus O
11. Wawer M
12. Paavolainen L
13. Herrmann MD
14. Rohban M
15. Hung J
16. Hennig H
17. Concannon J
18. Smith I
19. Clemons PA
20. Singh S
21. Rees P
22. Horvath P
23. Linington RG
24. Carpenter AE
(2017) Data-analysis strategies for image-based cell profiling
Nature Methods 14:849–863.
https://doi.org/10.1038/nmeth.4397
- PubMed
- Google Scholar
(2018) Weakly supervised learning of single-cell feature embeddings
Proceedings. IEEE Computer Society Conference on Computer Vision and Pattern Recognition 2018:9309–9318.
https://doi.org/10.1109/CVPR.2018.00970
- PubMed
- Google Scholar
1. Caicedo JC
2. Arevalo J
3. Piccioni F
4. Bray MA
5. Hartland CL
6. Wu X
7. Brooks AN
8. Berger AH
9. Boehm JS
10. Carpenter AE
11. Singh S
(2022) Cell Painting predicts impact of lung cancer variants
Molecular Biology of the Cell 33:ar49.
https://doi.org/10.1091/mbc.E21-11-0538
- PubMed
- Google Scholar
1. Chen D
2. Frezza M
3. Schmitt S
4. Kanwar J
5. Dou QP
(2011) Bortezomib as the first proteasome inhibitor anticancer drug: current status and future perspectives
Current Cancer Drug Targets 11:239–253.
https://doi.org/10.2174/156800911794519752
- PubMed
- Google Scholar
(2022) Artificial intelligence to identify genetic alterations in conventional histopathology
The Journal of Pathology 257:430–444.
https://doi.org/10.1002/path.5898
- PubMed
- Google Scholar
1. Foroutan M
2. Bhuva DD
3. Lyu R
4. Horan K
5. Cursons J
6. Davis MJ
(2018) Single sample scoring of molecular phenotypes
BMC Bioinformatics 19:404.
https://doi.org/10.1186/s12859-018-2435-4
- PubMed
- Google Scholar
1. Franke NE
2. Niewerth D
3. Assaraf YG
4. van Meerloo J
5. Vojtekova K
6. van Zantwijk CH
7. Zweegman S
8. Chan ET
9. Kirk CJ
10. Geerke DP
11. Schimmer AD
12. Kaspers GJL
13. Jansen G
14. Cloos J
(2012) Impaired bortezomib binding to mutant β5 subunit of the proteasome is the underlying basis for bortezomib resistance in leukemia cells
Leukemia 26:757–768.
https://doi.org/10.1038/leu.2011.256
- PubMed
- Google Scholar
1. Garraway LA
2. Jänne PA
(2012) Circumventing cancer drug resistance in the era of personalized medicine
Cancer Discovery 2:214–226.
https://doi.org/10.1158/2159-8290.CD-12-0012
- PubMed
- Google Scholar
1. Glaab WE
2. Tindall KR
(1997) Mutation rate at the hprt locus in human cancer cell lines with specific mismatch repair-gene defects
Carcinogenesis 18:1–8.
https://doi.org/10.1093/carcin/18.1.1
- PubMed
- Google Scholar
(2020) Resistance to the proteasome inhibitors: Lessons from multiple myeloma and mantle cell lymphoma
Advances in Experimental Medicine and Biology 1233:153–174.
https://doi.org/10.1007/978-3-030-38266-7_6
- PubMed
- Google Scholar
1. Gottesman MM
2. Lavi O
3. Hall MD
4. Gillet JP
(2016) Toward a better understanding of the complexity of cancer drug resistance
Annual Review of Pharmacology and Toxicology 56:85–102.
https://doi.org/10.1146/annurev-pharmtox-010715-103111
- PubMed
- Google Scholar
Preprint
1. Heiser K
2. McLean PF
3. Davis CT
4. Fogelson B
5. Gordon HB
6. Jacobson P
7. Hurst B
8. Miller B
9. Alfa RW
10. Earnshaw BA
11. Victors ML
12. Chong YT
13. Haque IS
14. Low AS
15. Gibson CC
(2020) Identification of Potential Treatments for COVID-19 through Artificial Intelligence-Enabled Phenomic Analysis of Human Cells Infected with SARS-CoV-2
bioRxiv.
https://doi.org/10.1101/2020.04.21.054387
- Google Scholar
(2001)
The proteasome inhibitor PS-341 inhibits growth, induces apoptosis, and overcomes drug resistance in human multiple myeloma cells

Cancer Research 61:3071–3076.
- PubMed
- Google Scholar
(2007) Understanding multiple myeloma pathogenesis in the bone marrow to identify new therapeutic targets
Nature Reviews. Cancer 7:585–598.
https://doi.org/10.1038/nrc2189
- PubMed
- Google Scholar
1. Kapoor TM
2. Miller RM
(2017) Leveraging chemotype-specific resistance for drug target identification and chemical biology
Trends in Pharmacological Sciences 38:1100–1109.
https://doi.org/10.1016/j.tips.2017.09.003
- PubMed
- Google Scholar
(2014) DrugTargetSeqR: a genomics- and CRISPR-Cas9-based method to analyze drug targets
Nature Chemical Biology 10:626–628.
https://doi.org/10.1038/nchembio.1551
- PubMed
- Google Scholar
1. Lee SH
2. Jang HJ
(2022) Deep learning-based prediction of molecular cancer biomarkers from tissue slides: A new tool for precision oncology
Clinical and Molecular Hepatology 28:754–772.
https://doi.org/10.3350/cmh.2021.0394
- PubMed
- Google Scholar
1. Liu LF
(1989) DNA topoisomerase poisons as antitumor drugs
Annual Review of Biochemistry 58:351–375.
https://doi.org/10.1146/annurev.bi.58.070189.002031
- PubMed
- Google Scholar
1. Ljosa V
2. Caie PD
3. Ter Horst R
4. Sokolnicki KL
5. Jenkins EL
6. Daya S
7. Roberts ME
8. Jones TR
9. Singh S
10. Genovesio A
11. Clemons PA
12. Carragher NO
13. Carpenter AE
(2013) Comparison of methods for image-based profiling of cellular morphological responses to small-molecule treatment
Journal of Biomolecular Screening 18:1321–1329.
https://doi.org/10.1177/1087057113503553
- PubMed
- Google Scholar
(2018) UMAP: Uniform Manifold Approximation and Projection
Journal of Open Source Software 3:861.
https://doi.org/10.21105/joss.00861
- Google Scholar
(2004) Focus on multiple myeloma
Cancer Cell 6:439–444.
https://doi.org/10.1016/j.ccr.2004.10.020
- PubMed
- Google Scholar
(2020) Bioactivity screening of environmental chemicals using imaging-based high-throughput phenotypic profiling
Toxicology and Applied Pharmacology 389:114876.
https://doi.org/10.1016/j.taap.2019.114876
- PubMed
- Google Scholar
1. O’Brien J
2. Wilson I
3. Orton T
4. Pognan F
(2000) Investigation of the Alamar Blue (resazurin) fluorescent dye for the assessment of mammalian cell cytotoxicity
European Journal of Biochemistry 267:5421–5426.
https://doi.org/10.1046/j.1432-1327.2000.01606.x
- PubMed
- Google Scholar
1. Papadopoulos N
2. Nicolaides NC
3. Wei YF
4. Ruben SM
5. Carter KC
6. Rosen CA
7. Haseltine WA
8. Fleischmann RD
9. Fraser CM
10. Adams MD
(1994) Mutation of a mutL homolog in hereditary colon cancer
Science 263:1625–1629.
https://doi.org/10.1126/science.8128251
- PubMed
- Google Scholar
1. Perlman ZE
2. Slack MD
3. Feng Y
4. Mitchison TJ
5. Wu LF
6. Altschuler SJ
(2004) Multidimensional drug profiling by automated microscopy
Science 306:1194–1198.
https://doi.org/10.1126/science.1100709
- PubMed
- Google Scholar
1. Pisa R
2. Kapoor TM
(2020) Chemical strategies to overcome resistance against targeted anticancer therapeutics
Nature Chemical Biology 16:817–825.
https://doi.org/10.1038/s41589-020-0596-8
- PubMed
- Google Scholar
1. Rohban MH
2. Singh S
3. Wu X
4. Berthet JB
5. Bray MA
6. Shrestha Y
7. Varelas X
8. Boehm JS
9. Carpenter AE
(2017) Systematic morphological profiling of human gene and allele function via Cell Painting
eLife 6:e24060.
https://doi.org/10.7554/eLife.24060
- PubMed
- Google Scholar
1. Simm J
2. Klambauer G
3. Arany A
4. Steijaert M
5. Wegner JK
6. Gustin E
7. Chupakhin V
8. Chong YT
9. Vialard J
10. Buijnsters P
11. Velter I
12. Vapirev A
13. Singh S
14. Carpenter AE
15. Wuyts R
16. Hochreiter S
17. Moreau Y
18. Ceulemans H
(2018) Repurposing high-throughput image assays enables biological activity prediction for drug discovery
Cell Chemical Biology 25:611–618.
https://doi.org/10.1016/j.chembiol.2018.01.015
- PubMed
- Google Scholar
(2021) CellProfiler 4: improvements in speed, utility and usability
BMC Bioinformatics 22:433.
https://doi.org/10.1186/s12859-021-04344-9
- PubMed
- Google Scholar
1. Teraishi F
2. Wu S
3. Zhang L
4. Guo W
5. Davis JJ
6. Dong F
7. Fang B
(2005) Identification of a novel synthetic thiazolidin compound capable of inducing c-Jun NH2-terminal kinase-dependent apoptosis in human colon cancer cells
Cancer Research 65:6380–6387.
https://doi.org/10.1158/0008-5472.CAN-05-0575
- PubMed
- Google Scholar
1. Tukey JW
(1949)
Comparing individual means in the analysis of variance

Biometrics 5:99–114.
- PubMed
- Google Scholar
1. Umar A
2. Boyer JC
3. Thomas DC
4. Nguyen DC
5. Risinger JI
6. Boyd J
7. Ionov Y
8. Perucho M
9. Kunkel TA
(1994)
Defective mismatch repair in extracts of colorectal and endometrial cancer cell lines exhibiting microsatellite instability

The Journal of Biological Chemistry 269:14367–14370.
- PubMed
- Google Scholar
1. Uphoff CC
2. Drexler HG
(2013) Eradication of mycoplasma contaminations
Methods in Molecular Biology 946:15–26.
https://doi.org/10.1007/978-1-62703-128-8_2
- PubMed
- Google Scholar
1. van de Ven R
2. Scheffer GL
3. Reurs AW
4. Lindenberg JJ
5. Oerlemans R
6. Jansen G
7. Gillet J-P
8. Glasgow JN
9. Pereboev A
10. Curiel DT
11. Scheper RJ
12. de Gruijl TD
(2008) A role for multidrug resistance protein 4 (MRP4; ABCC4) in human dendritic cell migration
Blood 112:2353–2359.
https://doi.org/10.1182/blood-2008-03-147850
- PubMed
- Google Scholar
(2011) Multi-subject dictionary learning to segment an atlas of brain spontaneous activity
Information Processing in Medical Imaging 22:562–573.
https://doi.org/10.1007/978-3-642-22092-0_46
- PubMed
- Google Scholar
(2019) A view on drug resistance in cancer
Nature 575:299–309.
https://doi.org/10.1038/s41586-019-1730-1
- PubMed
- Google Scholar
(2013) Endoplasmic reticulum stress and the unfolded protein response: targeting the Achilles heel of multiple myeloma
Molecular Cancer Therapeutics 12:831–843.
https://doi.org/10.1158/1535-7163.MCT-12-0782
- PubMed
- Google Scholar
(2012) Using transcriptome sequencing to identify mechanisms of drug action and resistance
Nature Chemical Biology 8:235–237.
https://doi.org/10.1038/nchembio.779
- PubMed
- Google Scholar
1. Wawer MJ
2. Li K
3. Gustafsdottir SM
4. Ljosa V
5. Bodycombe NE
6. Marton MA
7. Sokolnicki KL
8. Bray M-A
9. Kemp MM
10. Winchester E
11. Taylor B
12. Grant GB
13. Hon CS-Y
14. Duvall JR
15. Wilson JA
16. Bittker JA
17. Dančík V
18. Narayan R
19. Subramanian A
20. Winckler W
21. Golub TR
22. Carpenter AE
23. Shamji AF
24. Schreiber SL
25. Clemons PA
(2014) Toward performance-diverse small-molecule libraries for cell-based phenotypic screening using multiplexed high-dimensional profiling
PNAS 111:10911–10916.
https://doi.org/10.1073/pnas.1410933111
- Google Scholar
Software
1. Way G
(2020) Blocklist features - cell profiler
Figshare.

https://doi.org/10.6084/m9.figshare.10255811.v3
1. Way GP
2. Natoli T
3. Adeboye A
4. Litichevskiy L
5. Yang A
6. Lu X
7. Caicedo JC
8. Cimini BA
9. Karhohs K
10. Logan DJ
11. Rohban MH
12. Kost-Alimova M
13. Hartland K
14. Bornholdt M
15. Chandrasekaran SN
16. Haghighi M
17. Weisbart E
18. Singh S
19. Subramanian A
20. Carpenter AE
(2022) Morphology and gene expression profiling provide complementary information for mapping cell state
Cell Systems 13:911–923.
https://doi.org/10.1016/j.cels.2022.10.001
- PubMed
- Google Scholar
Software
1. Way G
2. Han Y
3. Stirling D
4. Singh S
(2023) broadinstitute/profiling-resistance-mechanisms: Response to reviewers
Zenodo.

https://doi.org/10.5281/zenodo.8170152
(2014) Unique molecular landscapes in cancer: implications for individualized, curated drug combinations
Cancer Research 74:7181–7184.
https://doi.org/10.1158/0008-5472.CAN-14-2329
- PubMed
- Google Scholar

Decision letter

Caigang Liu

Senior and Reviewing Editor; Shengjing Hospital of China Medical University, China

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

[Editors' note: this paper was reviewed by Review Commons.]

https://doi.org/10.7554/eLife.91362.sa1

Author response

General statements

We would like to thank the editorial staff and the reviewers for their handling of our manuscript. We were very pleased with the timely communications from Review Commons, and we are grateful to have been assigned this insightful and constructive group of reviewers.

The reviewers were well-suited to evaluate our work based on their stated areas of expertise (cancer biology, image analysis, machine learning, cell-based screening, etc.). As such, we received thoughtful and constructive feedback, which we have already incorporated into our attached revision. We are confident that these reviews have improved our manuscript.

Our goal with this manuscript is to present a proof-of-concept study where high-content imaging and morphological profiling are used to characterize drug resistance in clonal cell lines. The main criticism from reviewers was that our original manuscript may have overstated our method’s ability to discriminate the signal of bortezomib resistance and that any extension beyond cultured cells (to patient samples for example) would require significant follow-up studies. The reviewers suggested that such work would be beyond the scope of our study, and recommended toning down our language to better reflect the limitations of this proof-of-concept work. We have embraced this suggestion, extensively revising our text, and we now believe our language and tone more accurately reflects our results. The reviewers also suggested follow-up computational analyses to more robustly characterize the bortezomib resistance signature. We have performed these analyses and added their description to our revised manuscript. We feel that these analyses have improved understanding of the signature, and will help a reader to gain a deeper understanding of our results and methodology.

The reviewers also suggested several minor changes; many of which we embraced fully, but others that we chose not to incorporate. We felt that a lack of clarity in our text contributed to these reviewer suggestions. In these cases, we improved clarity in the text and responded to each comment point-by-point in the “prefer not to carry out” section. Further, we address all reviewer comments in the following document point-by-point, grouped by common themes across reviewers (e.g., tone, clarity, analyses, etc.).

Lastly, a common theme among reviewer comments was their appreciation for our strong methodology and data transparency (examples pasted below). We are extremely gratified by this observation as we feel this is a particular strength of our manuscript. In addition, we were pleased to see reviewers engaged by our work, acknowledging the interest this manuscript is likely to generate among a broad range of scientific disciplines.

Examples of reviewer appreciation of our strong methodology and data transparency:

Reviewer 1: “However, this does not imply that the same approach can not achieve the goal, perhaps by using other cell painting markers for bortezomib-sensitivity, or with the same markers to assess sensitivity of different drugs. The cell painting + analysis approaches are not new and the clinical impact is questionable, but the technical aspects (data, analysis) are exceptional and the concept may hold as I described above.”

Reviewer 2: “The paper is well written, and the text is clear, as is the presentation of data and transparency of methods being utilized. The methods were applied appropriately and followed established standards in the field. The paper's premise is timely and interesting, addressing a pressing issue in cancer therapy: making informed treatment decisions fast, based on markers found in tumors early in tumor development, and using image-based screening for characterizing drug resistance before treatment could be an option. A fascinating bit of the manuscript is the description of the feature selection from the screen is done systematically, considering the technical and biological variability and technical artifacts and modeling covariates using linear models seems a very appropriate way of doing so and could serve as another proof of concept that this is indeed the most robust way of modeling and removing signal of technical covariates from the data.”

Reviewer 3: “The strengths of this study are the machine learning best practice and detailed methodology. The experiments could be reproduced and statistical analysis is more than adequate. The analysis takes into account batch effects, well position, differences in cell numbers, and other sources of technical variation that complicate high-content image analysis. It is a good exemplar of how unsupervised morphological profiling can be applied to imaging data. The major limitation is the generalizability of this particular method for patient samples. This could be addressed in the Discussion.”

Description of the revisions

We have incorporated all planned revisions.

Description of the revisions that have already been incorporated in the transferred manuscript

Text revisions already carried out

[Text revision] We have materially toned down our claims in the manuscript in two distinct areas: (A) model performance and (B) potential clinical application.

Model performance. We specifically balanced our discussion of the discriminative signal of the Bortezomib Signature. While the signature adequately separated never-before-seen wildtype and resistant clones with metrics well above randomly permuted baselines (accuracy near 80%, average precision about 70%, area under the ROC curve (AUROC) about 84%), there were many limitations that we should have more explicitly highlighted. For example, many individual profiles were incorrectly classified, some clones were predicted entirely incorrectly, and many profiles did not receive Bortezomib Signature scores above the randomly permuted baseline. We have more clearly discussed these limitations and used more balanced language (see key examples of text-based changes below). Additionally, we modified a figure (now Figure 3) to include boxplots of clones that explicitly show the Bortezomib Signature scores of each well profile and permit examination of the strength of the signature for each clone (previously found in Figure 2-Supplement 9). Lastly, we add a new supplementary figure (now Figure 5-Supplement 1) that describes a feature space analysis of misclassified samples. Please note that this figure rearrangement and new analysis helped to balance our claims, but were also performed in response to other tangential reviewer comments.
Clinical application. In the abstract, introduction, and discussion, we further emphasized that this work is a proof of concept, and that more advances must be made prior to clinical application.

We made these changes in direct response to the following reviewer comments:

Reviewer 1 - Major Comment 1 (relevant excerpts)

While I am convinced that the signature captures morphological phenotypes associated with drug resistance, at the cumulative scale, the discriminative signal of a single cell type seems weak… With Fig. 4, the data fully supports the argument that the bortezomib-signature encodes bortezomib-resistance, but the signal is weak. Thus statements such as "We found the Bortezomib Signature could predict whether a cell line was bortezomib-resistant or bortezomib-sensitive" (line #172) and the specificity statements in the abstract" (line #28) are not supported by the data in my opinion. I would recommend the authors to tune down these and other related statements throughout the manuscript.

Reviewer cross-commenting - Reviewer 1

My main critic is regarding "over selling" a weak discriminative signal. Specifically, I am not convinced that the major claims regarding predicting sensitivity and specificity at the single cell types scales are supported by the data. Since reviewer #2 and #3 did not raise this concern I think it is worth discussion here.

Once these statements are tuned down - I think no significant additional work is needed to make the point that they can measure a discriminative signal. If they want to make these claims, perhaps they'd like to collect more data to gain statistical power (but I am not optimistic this will work at the single cell level).

Personally, I was happy with the authors' choice of cell lines not included in the training dataset. I am not convinced that additional cell lines + validations are necessary for making the point of a proof of principle.

Reviewer cross-commenting - Reviewer 2

I agree that, perhaps, my major criticism of the paper was the manuscript's 'overselling' of claims that were only weakly supported by the data. Yes, if the authors tune down their claims and clearly state that this is an interesting starting point and proof of concept study, it might be ok to publish with only minor revisions. If the claims should be more generalized, then this study needs more data supporting the conclusions and the method's predictive power.

Reviewer 2 - Major Comment 8

Lastly, I find some misfits between the question, the model used, and the conclusions drawn. The authors start by exploring the problem of bortezomib resistance in cancer treatment, which they say is a devastating issue for patients with, e.g., multiple myeloma. Yet, the authors use HCT116 as their model cell line, a microsatellite instable, colorectal cell line with several intrinsic mutations that make it a difficult model to address physiologically relevant medical problems after all. The authors then go on to suppose that their method might be suitable to diagnose resistance in patient samples, but I am not convinced this conclusion can be speculated based on data from HCT cells. I suggest the authors test their approach on at least two other cell lines (maybe from different tissues) and benchmark their results against a dataset of digital pathology where such predictions are made from stained and analyzed tissue slices. This way, after a thorough benchmark against related third-party data sets, the method would significantly gain relevance, the paper would appeal to a broader audience, and the advance gains more merit.

Reviewer 3 - Major Comment 5

It is not clear from the Discussion whether this type of analysis is more broadly applicable to cell lines derived from patients, rather than selected from a parental cell line, or if this approach would be more efficient than genotyping or next-gen sequencing. How many replicates and ground truth cell lines would be necessary for predictive confidence?

We edited the last two sentences of the abstract to tone down specificity claims (“provide evidence”) and clarify that we are establishing a “proof-of-concept framework”.

“This signature predicted bortezomib resistance better than resistance to other drugs targeting the ubiquitin-proteasome system. Our results establish a proof-of-concept framework for the unbiased analysis of drug resistance using high-content microscopy of cancer cells, in the absence of drug treatment.”

We revised the last paragraph of the introduction to contrast bortezomib predictions with ixazomib/CB-5083 predictions, and to remove claims about “using microscopy to guide therapy”.

“This morphological signature correctly predicted the bortezomib resistance of seven out of ten clones not included in the signature training dataset. Overall, our results establish a proof-of-concept framework for identifying unbiased signatures of drug resistance using high-content microscopy. The ability to identify drug-resistant cells based on morphological features provides a valuable orthogonal method for characterizing resistance in the absence of drug treatment.”

To tone down claims in the figures, we added boxplots to Figure 3 (previous Figure 2) showing specific distribution of signature scores per well profile and updated Figure 4 legend (previous Figure 3).

“Figure 4. Bortezomib Signature has limited ability to characterize clones resistant to other ubiquitin-proteasome system inhibitors.”

We modify the following text in the discussion to tone down claims of specificity and clinical utility:

“This Bortezomib Signature correctly predicted the bortezomib resistance of seven out of ten clones not included in the training dataset and was more specific to bortezomib-resistance given its limited ability to identify clones that were resistant to other UPS-targeting drugs.

Though it is unclear whether this method can be extended to patient samples, where identifying intrinsic drug resistance in cells prior to treatment has the potential to improve targeted cancer therapy, our results are an encouraging proof of concept. We expect that further refinement may develop Cell Painting as a tool for identifying drug-resistant cells, perhaps even guiding strategies to overcome intrinsic resistance.”

[Text revision] We defined LD50 in text (originally line #97), changed description of resistant clone selection to remove main text references to LD90 (originally line #87), and stated drug concentrations used for selection in Methods. We also defined LD90 in the Methods and described its role in determining the drug concentrations to use for clone selection. This change was in response to the following comments:

Reviewer 1 - Minor Comment 2

What is LD90 (line #87)? LD50 (line #97)?

Reviewer 2 - Minor Comment 5

What was the LD 90 per drug on HCT cells? Rather than LD90 foldchanges, absolute concentrations should be used in the results and discussion to allow the reader to vet the conclusions.

“To determine the appropriate drug concentrations to use in order to isolate drug-resistant clones, we performed proliferation assays on HCT116 parental cells with our drugs of interest: bortezomib (proteasome inhibitor), ixazomib (proteasome inhibitor), or CB-5083 (p97 inhibitor) (Fig. 1-Supplement 1 A-D).”

“We characterized the bortezomib-resistant clones and found that the median lethal doses (LD50s) were ~2.8- to ~9-fold that of HCT116 parental cells (Fig. 1-Supplement 2 B).”

“Briefly, HCT116 cells were plated in 150 mm dishes and grown in the presence of the desired drug at a concentration that resulted in the death of the majority of cells (selection concentrations: bortezomib, 12 nM; ixazomib, 150 nM; CB-5083, 600 and 700 nM).”

“Using the data from our proliferation assays, we calculated the median lethal dose (LD50) for each of our drugs of interest by fitting data of normalized growth vs. log[drug concentration] to a sigmoidal dose-response curve using GraphPad Prism (v.9.2.0) (Fig. 1-Supplement 1 D).”

[Text revision] We thank the reviewer for allowing us an opportunity to improve clarity on the clones we used. We now describe the total number of clones generated and removed unnecessary references to specific clones for ease of reading (originally lines #96-98) (We maintain all references to specific clones in the figures, legends, supplement, and methods)

Reviewer 1 - Minor Comment 3

It was not clear to me in the text which and how many cell lines were evaluated and the reader is forced to go to the SI. For example, "(BZ01-10 and BZ clones A and E)" (line #96-97) and "wild-type clones (WT01-05, 10, and 12-15)" (line #98) appeared when presenting the results without a clear explanation and made it harder for me to follow. Summary of the data (for example, based on Figure 2-Supplement 8) can be briefly mentioned in the text to make it more clear for the reader.

We added the following to the second paragraph of the results:

“Together these methods provided a total of twelve bortezomib-resistant, five ixazomib-resistant, five CB-5083-resistant, and twelve bortezomib-sensitive clones as well as HCT116 parental cells for our experiments.”

[Text revision] We removed duplicate text (originally lines #115-125).

Reviewer 1 - Minor Comment 5

Lines #104-111 were duplicated in lines #114-122.

Reviewer 3 - Minor Comment 4

Ten lines of text are duplicated on page 5.

Reviewer 2 - Minor Comment 4 on page 5, paragraph 4, there is a sizeable copy-and-paste error of text being identically replicated.

[Text revision] We provided more intuition of the Bortezomib Signature in the results section (originally lines #150-151).

Reviewer 1 - Minor Comment 6

The "Bortezomib Signature" is a critical measurement but is only briefly mentioned in lines 150-151 ("...based on the direction-sensitive ranking method for phenotype analysis, singscore (Foroutan et al., 2018)"). Please provide more information/intuition.

“We used these 45 features to compute a rank-based resistance score or “Bortezomib Signature” for each well profile based on the direction-sensitive method called singscore (Foroutan et al. 2018). Singscore ranks these 45 resistance-related features on a per sample basis and calculates a normalized score between -1 and 1, with higher values expected for bortezomib-resistant clones and lower values expected for bortezomib-sensitive clones.”

[Text revision] We clarified that DNA sequencing had been performed solely on clones A and E in a previous study (originally lines #88-90). Furthermore, one of the strengths of our approach is that it can identify resistant clones in an unbiased fashion prior to molecular characterization. It is beyond scope to perform these sequencing studies in the present paper.

Reviewer 2 - Minor Comment 3

The authors talk about validating the mutation - PSMB5 by RNA-seq. However, the data for the genotyping/sequencing/characterization of these newly generated BZ-resistant lines are missing.

In the results, we clarify DNA sequencing that was previously performed on clones A and E

“We also isolated bortezomib-sensitive (wild-type; WT) clones by dilution of the HCT116 parental cell line and acquired two bortezomib-resistant clones (BZ clones A and E) both with mutations in PSMB5 identified by RNA sequencing performed in previous work (Fig. 1-Supplement 1 E) (Wacker et al. 2012).”

In the last paragraph of the discussion, we highlight the strength of our unbiased approach.

“Together, our work has demonstrated the potential for morphological profiling with Cell Painting to be used as an unbiased method to characterize resistance in the absence of drug treatment. Our results indicate that different mechanisms of bortezomib resistance may generate distinct morphological profiles; with larger and broader training datasets, it may be possible to identify signatures for distinct mechanisms of bortezomib resistance as well as signatures of resistance to other drugs. Though it is unclear whether this method can be extended to patient samples, where identifying intrinsic drug resistance in cells prior to treatment has the potential to improve targeted cancer therapy, our results are an encouraging proof of concept. We expect that further refinement may develop Cell Painting as a tool for identifying drug-resistant cells, perhaps even guiding strategies to overcome intrinsic resistance.”

[Text revision] We thank the reviewers for their suggestions. We agree that the description of the experimental design was somewhat unclear and have provided greater detail and clarity, particularly regarding the generation of clones. We used the HCT116 parental cell line to generate drug-resistant clones by identifying single surviving cells after drug treatment and allowing these cells to expand prior to isolating colonies for experimentation. We did not perform experiments to confirm whether these “clones” were isogenic and can not exclude cell migration during expansion or genetic drift as convoluting factors. However, we have provided greater detail in the descriptions of our method for clone isolation in order to address this concern.

Reviewer 1 - Minor Comment 1

More information in Fig. 1's legend would be helpful to follow the experimental design. I found it hard to follow in its current form and had to go back to carefully reading the main text to fully understand.

Reviewer 2 - Minor Comment 6

The description of the resistant clonal populations is confusing. As I understand, no single-cell clones were isolated during the selection procedure. Thus, the training lines are not yet isogenic clones but oligoclonal sub-populations of the parental cell line. The authors could provide more details here and discuss the different characteristics of their sub-populations, e.g., their growth kinetics or molecular alterations.

We bolstered the description in the results.

“We first isolated and characterized drug-resistant cells (Fig. 1 A). To isolate drug-resistant clones, we used an approach we have described previously (Wacker et al. 2012; Kasap, Elemento, and Kapoor 2014) and the HCT116 cell line. These cancer cells express multidrug resistance pumps at low levels and are mismatch repair deficient, providing a genetically heterogeneous polyclonal population of cells (Umar et al. 1994; Papadopoulos et al. 1994; Teraishi et al. 2005) allowing for isolation of drug-resistant clones in 2-3 weeks. We hypothesize that a rapid selection of resistance could favor the isolation of clones with intrinsic resistance. To determine the appropriate drug concentrations to use in order to isolate drug-resistant clones, we performed proliferation assays on HCT116 parental cells with our drugs of interest: bortezomib, ixazomib, or CB-5083 (Fig. 1-Supplement 1 A-D). We also isolated bortezomib-sensitive (wild-type; WT) clones by dilution of the HCT116 parental cell line and acquired two published bortezomib-resistant clones (BZ clones A and E) both with mutations in PSMB5 identified by RNA sequencing performed in previous work (Fig. 1-Supplement 1 E) (Wacker et al. 2012). We characterized the bortezomib-resistant clones and found that the median lethal doses (LD50s) for bortezomib were ~2.8- to ~9-fold that of HCT116 parental cells (Fig. 1-Supplement 2 B). In contrast, bortezomib-sensitive clones had LD50s for bortezomib that ranged from ~0.7- to ~1.2-fold that of HCT116 parental cells (Fig. 1-Supplement 2 A). Together these methods provided a total of twelve bortezomib-resistant, five ixazomib-resistant, five CB-5083-resistant, and twelve bortezomib-sensitive clones as well as HCT116 parental cells for our experiments.”

We also updated the legend for Figure 1A.

“Figure 1. Experimental design for using Cell Painting to examine morphological profiles of drug-resistant cells. (A) Graphic of the experimental workflow: we isolated drug-resistant clones by treating parental HCT116 cells with a high dose of the desired drug and then expanded them for experiments. We isolated drug-sensitive clones by diluting HCT116 cells and then expanded them for experiments. We then performed proliferation assays on select clones to screen for multidrug resistance. Next, we performed Cell Painting on both drug-resistant and -sensitive clones, using multiplexed high-throughput fluorescence microscopy of fixed cells followed by feature extraction and morphological profiling to search for features that contribute to a signature of drug resistance.”

[Text revision] We clarified that the Bortezomib Signature did not correspond to well position (originally lines #155-157).

Reviewer 1 - Minor Comment 9

Line #155-156: "We found that the pattern of Bortezomib Signatures corresponded to the cell identity plate layout", the word "not" is missing before "corresponded".

“We found that the pattern of Bortezomib Signatures did not correspond to well position relative to the plate (Fig. 2-Supplement 7 B), indicating that the well position for each clone was not strongly contributing to its Bortezomib Signature.”

[Text revision] We explicitly described the result that some misclassified clones (WT10, WT15, and BZ06) did not have unexpected bortezomib sensitivity as determined by proliferation assays. We also moved the supplementary figure to an updated Figure 3 to better highlight this result (described below in “Figure revisions already carried out”). Lastly, we add a new figure (Figure 5-Supplement 1) to more explicitly analyze the misclassified lines (described below in “New analyses already carried out”).

Reviewer 3 - Minor Comment 3

The bortezomib sensitivity of the WT lines used in the last experiments was determined and did not seem to be greater than parental. This could be mentioned in the text; the figure raises the question and the answer is provided, but it's in the supplemental material.

“While the Bortezomib Signature correctly characterized the bortezomib sensitivity of most clones, it consistently misclassified others (WT10, WT15, and BZ06) (Fig 5-Supplement 1 A). Proliferation assays conducted in earlier experiments showed that WT10 and WT15 were sensitive to bortezomib while BZ06 was resistant (Fig. 1-Supplement 2 A and B). By comparing these incorrect predictions with high-confidence correct predictions, we observed differences that varied by clone type, suggesting unique morphology may be driving each of these misclassifications (Fig. 5-Supplement 1 B and C). These results are consistent with the Bortezomib Signature being generalizable to clones not included in the training dataset and suggest that morphological profiling has the potential to identify bortezomib-resistant clones based on the morphological features of cells in the absence of drug treatment.”

[Text revision] We clarified that the metrics (accuracy and average precision) were based on median Bortezomib Signature scores of all replicate well-level profiles per clone. We can compare samples based on rank, and difference from 95% confidence interval of permuted data. There is no current way for our method to assign a likelihood. Also note that we have updated the discussion to discuss alternative metrics (see Reviewer 1 - Minor Comment 7) These are very important distinctions, and we are grateful to the reviewer for bringing them up.

Reviewer 3 - Major Comment 3

The study classifies cells as binary sensitive or resistant, but would results be improved by scoring based on likelihood of being resistant/sensitive?

Reviewer 3 - Minor Comment 2

It is not clear whether the accuracy was based on a percentage of replicates per cell line that were classified correctly or whether that was referring to classification of the cell line overall as sensitive/resistant.

“We next examined whether the Bortezomib Signature was able to predict the bortezomib resistance of a clone based on morphological profiling data (Fig. 3 A-E and Fig. 3-Supplement 2 A and B). We called the clone bortezomib-resistant if the median Bortezomib Signature of all replicate well profiles was greater than zero and bortezomib-sensitive if the median Bortezomib Signature less than zero. In the training dataset, the Bortezomib Signature correctly predicted the bortezomib resistance of all ten clones, with median Bortezomib Signatures for eight out of ten clones beyond the 95% confidence interval for the randomly permuted data (Fig. 3 A). The accuracy of the Bortezomib Signature was 88% while the average precision was 81% for the training dataset (Fig. 3-Supplement 2 A and B) (see Methods). The signature performed similarly well in the validation dataset (Fig. 3 B), with an accuracy of 92% and an average precision of 89% (Fig. 3-Supplement 2 A and B). In the test dataset the Bortezomib Signature correctly predicted the bortezomib resistance of all clones, though only HCT116 parental cells had a median Bortezomib Signature outside the 95% confidence interval for the randomly permuted data (Fig. 3 C). The test dataset had an accuracy of 80% and an average precision of 68% (Fig. 3-Supplement 2 A and B). Similarly, in the holdout dataset the Bortezomib Signature had an accuracy of 78% and an average precision of 69% (Fig.3 -Supplement 2 A and B), and correctly predicted the bortezomib resistance of twelve out of thirteen clones, with WT01 misclassified as bortezomib-resistant (Fig. 3 D). In the holdout dataset, four of the twelve correctly characterized clones had median Bortezomib Signatures outside the 95% confidence interval for the randomly permuted data.”

We also mirrored language when discussing the ixazomib and CB-5083 results.

“However, only two of the four correctly identified ixazomib-resistant clones and one of the three CB-5083-resistant clones had median Bortezomib Signatures outside the 95% confidence interval of the randomly permuted data. The area under the ROC (AUROC) curve for ixazomib-resistant and CB-5083-resistant clones (0.63 and 0.60, respectively) was lower than those calculated for the training, validation, test, and holdout datasets. In addition, many of the Bortezomib Signatures for well profiles of ixazomib- and CB-5083-resistant clones, particularly those for CB-5083-resistant clones, landed within the 95% confidence interval of the randomly permuted data. These results suggest that the Bortezomib Signature is not a general signature of UPS-targeting drug resistance and instead has some specificity for bortezomib.”

[Text revision] We added an explicit note that our image analysis pipelines are also publicly available. Our reporting of our data processing pipelines are documented fully and well above standards in our field. Linking the publicly-available resources with these methods maximizes reproducibility.

Reviewer 1 - Minor Comment 10

Additional details on the processing steps in the analysis pipeline in the Methods will be highly appreciated.

“We include all image analysis pipelines at https://github.com/broadinstitute/profiling-resistance-mechanisms (G. Way et al. 2023).”

[Text revision] We have compared our approach to the on-disease/off-disease scores as introduced in (Heiser et al. 2020). We agree with the reviewer that a discussion of these two methods would help clarify our phenotypic signature concept. The on/off score is about the degree to which a perturbation pushes disease towards a healthy state. In this case we have 3 sets of data: healthy samples (used for training), disease samples (used for training), and the sample we want to score, which should be of the form "disease + perturbation". With our approach, based on singscore, we also have 3 sets of data: sensitive samples (used for training), resistance samples (used for training), and the sample we want to score. Here, our sample we want to score could be anything, not necessarily of the form "resistance + perturbation". Furthermore, singscore does not have the concept of orthogonality to resistance/sensitivity. This would become relevant if we were exploring perturbations or conditions that would induce a resistant cell line to become sensitive, but we are not doing that here. There are other statistical differences (projection vs. rank based etc.) but the key difference is the applicability of the method to the specific problem at hand.

Reviewer 1 - Minor Comment 7

How is the Bortezomib Signature related to the "on-disease"/"off-disease" scores described in https://www.biorxiv.org/content/10.1101/2020.04.21.054387v1.full? Are there other alternatives used for similar binary phenotypic signatures? What is the justification for using these measurements? I would love to see this generalized concept explicitly discussed in the Discussion.

We added the following to the discussion.

“The Bortezomib Signature is conceptually similar to the on-disease/off-disease score (Heiser et al. 2020). Both require three phenotypic measurements: a target phenotype representing ideal, a disease phenotype, and a new phenotype to classify. However, our approach is technically different (non-parametric compared to linear projection) and our goals are different (phenotypic classification compared to perturbation alignment). Other methods also enable phenotype labeling, but they focus on single-sample annotation without regard to a target phenotype (Wawer et al. 2014; Rohban et al. 2017; Simm et al. 2018; Nyffeler et al. 2020).”

Figure revisions already carried out

[Figure revision] We moved all boxplots from the original Fig. 2-Supplement 9 to the main text (also splitting Fig. 2 into Fig. 2 and 3). From the original Figure 2, we moved the accuracy and average precision bar graphs to the supplement. We also note that this change increases transparency of the discriminative signal of our signature.

Reviewer 1 - Minor Comment 8

I would highly recommend showing the Bortezomib Signatures from Figure 2-Supplement 9. in Fig. 2. This was the main measurement used throughout the manuscript and in my opinion, it is very important to consistently visualize the data along the manuscript, for clarity and easier reader interpretation.

[Figure revision] We adjusted the position of the legend in the accuracy and average precision bar graphs (originally Fig. 2 C and D, now Fig. 3-Supplement 2) for clarity. We also note that keeping the bar chart here is standard best practice (compared to a dot plot).

Reviewer 1 - Minor Comment 4

I found the visualization in Fig. 2C-D not intuitive (it is properly explained in the legend). I suggest replacing the accuracy colorbar with a color marker to make it more distinct from the random permutation (|--*--|) The location of the text "mean +- SD of 100 random permutation" made me first think that it is linked to the holdout.

[Figure revision] We changed the point distribution in the boxplots (from expanded to standard) to minimize overlap with the boxplot lines. We also updated the legend text to indicate that individual points in boxplots represent the Bortezomib Signature for well profiles. Note, we paste a representative example of this change above (new Figure 3).

Reviewer 3 - Minor Comment 1

I found the box plots somewhat difficult to interpret (especially where the WT lines had a lot of overlap with the red shaded area). Do the points in these charts correspond to replicate wells?

We also update the figure legend.

“Plots show values for individual well profiles (points), range (error bars), 25th and 75th percentiles (box boundaries), and median.”

[Figure revision] [Response to Reviewer 2 - Major Comment 7] We thank the reviewer for allowing us an opportunity to clarify the mechanism. We feel that it is beyond scope of this manuscript to disentangle the molecular alterations that cause bortezomib resistance based on our Cell Painting insights. This wet lab experimental process is arduous and cost prohibitive, and we argue that one of the benefits of taking a morphology approach to resistance status is that we can detect resistant cells (and therefore cells that won’t die when presented with a treatment) without knowing the molecular mechanism.

Nevertheless, the reviewer has encouraged us to enhance the ability for a reader to view and interpret the signature to perhaps more easily facilitate future work. Previously, we presented our signature in text form in Figure 2-Supplement 4 and in heatmap form in Figure 2-Supplement 5. Here, we add a new figure (Figure 2-Supplement 6; pasted below) which will improve interpretability.

Reviewer 2 - Major Comment 7:

Next to feature importance, the authors do not discuss (or I missed) what biology the features represent. Such the reader is left wondering what the actual mechanism of bortezomib resistance could be and if cell painting could shed light on the molecular alterations that cause the treatment resistance. While reviewing, I thus wondered which audience the authors targeted with their manuscript. A more focused analysis of their data that highlights aspects of the study either for the machine learning community, the cell biology community, or the precision oncology community would greatly benefit the manuscript's impact. In its current form, the study's findings seem diluted and spread across a wide range of research questions.

Additionally, we add the following to the results section:

“We then examined the grouping of features across compartments and channels and found radial distribution features were higher in resistant cells (Fig 2-Supplement 6).”

The code change to generate the signature visualization summary is available at: https://github.com/broadinstitute/profiling-resistance-mechanisms/pull/131

New analyses already carried out

[New analysis] [Response to Reviewer 2 - Major Comment 5] We agree that a systematic analysis of feature selection methods will provide additional insights not already in the manuscript. Therefore, we have performed two new computational experiments to compare our linear modeling feature selection approach against other standard approaches. We demonstrate that our linear modeling approach is effective at isolating the core differences between resistant and sensitive classes.

Specifically, we performed two analyses: (A) UMAP and (B) k-means cluster analysis. We analyzed profiles defined by four different feature selection approaches: (1) Using all traditional CellProfiler features; (2) Using the traditional CellProfiler feature selection approach (removing low variance features, high correlating features, etc.); (3) Using 45 random features (same size as Bortezomib Signature); and (4) Using only the bortezomib signature features. We performed Fisher’s exact tests to derive odds ratios of cluster membership by resistance status and calculated Silhouette widths to quantify relative proximity of clusters.

This analysis generates a new supplementary figure (see below), and demonstrates that the linear-modeling-based feature selection isolated the features driving the differences between the clone types (resistance vs. wildtype) while the standard approaches do not as effectively separate.

Reviewer 2 - Major Comment 5:

A fascinating bit of the manuscript is the description of the feature selection from the screen is done systematically, considering the technical and biological variability and technical artifacts and modeling covariates using linear models seems a very appropriate way of doing so and could serve as another proof of concept that this is indeed the most robust way of modeling and removing signal of technical covariates from the data. Yet, I wondered why the authors do not discuss other means of feature selection or dimensionality reduction; further, they need to show how the features cluster the cell lines or why impact (information content) different features deliver. For an audience interested in the technical aspects of cell painting analysis and machine learning based on the data, that would, IMHO, be the most exciting questions.

Additionally, we add the following to the results section:

“We then compared our linear-modeling approach to feature selection against other feature spaces and found that the Bortezomib Signature clusters same-type clones (bortezomib-resistant vs. bortezomib-sensitive) with higher enrichment compared to the full feature space, standard feature selection (see Methods), or a random selection of 45 features (Fig 3-Supplement 3).”

And methods section, describing this analysis:

“We were also interested in comparing the ability of different feature spaces to cluster clones of the same type (resistant vs. sensitive). This analysis would determine if the Bortezomib Signature features, which we derived using linear modeling to isolate biological from technical variables, had a greater ability to cluster. We compared the Bortezomib Signature against three other feature spaces: (1) the full feature space, (2) standard feature selection (see Image data processing methods), and (3) 45 randomly selected features. We performed two analyses using these four feature spaces including Uniform Manifold Approximation and Projection (UMAP) (McInnes et al. 2018) and k-means clustering. For UMAP, we used default umap-learn parameters to identify two UMAP coordinates per feature space. We then visualized the clusters by their resistance status and Bortezomib Signature score. The UMAP analysis represents a qualitative analysis. Next, we applied k-means clustering with 25 initializations across a range of 2-14 clusters (k). Prior to clustering and for each feature space, we applied principal component analysis (PCA) and transformed each feature space into 30 principal components. This step was necessary to compare k-means clustering metrics, which are sensitive to the feature space dimensionality. We applied a Fisher’s exact test to each cluster using a two-by-two contingency matrix that specified cluster membership for each clone classification (resistant vs. sensitive). We visualized the mean odds ratio and max cluster odds ratio for each feature space across k. A high odds ratio tells us that the feature space effectively clusters clones of the same resistance status. Lastly, we calculated Silhouette width (the average proximity between samples in one cluster to the second nearest cluster) for each feature space across k.”

The code change to derive the UMAP coordinates, perform clustering, and generate the figure is available at https://github.com/broadinstitute/profiling-resistance-mechanisms/pull/132

[New analysis] [Response to Reviewer 3 - Major Comment 1] We thank the reviewer for this suggestion, which allowed us to explore the misclassified samples in more depth. We added a new supplementary figure in which we summarized all bortezomib clones (wildtype and resistant) in their accuracy based on the bortezomib signature (panel A). We did not include training set samples in this analysis. Using samples that were consistently incorrectly classified with high confidence (three samples: WT15, BZ06, WT10) we performed two separate two-sample Kolmogorov–Smirnov (KS) tests. Specifically, we compared high incorrect wildtype to high correct wildtype and high incorrect resistant to high correct resistant. Our results indicate that most bortezomib signatures were significantly different between correct and incorrect assignments (panel B), and that the signature features varied between resistant and wildtype misclassification tests (panel C).

Reviewer 3 - Major Comment 1:

While the claims are largely substantiated, there are a few points where further consideration would improve the manuscript. Several cell lines were mis-classified with what appears to be a high degree of certainty. Can the authors tell what was driving those predictions? Was there something in the morphological signature that weighed more heavily in those cases?

Additionally, we add the following to the results section:

And methods section, describing this analysis:

“Some profiles were consistently predicted incorrectly with high confidence but in the opposite direction (see Figure 5-Supplement 1). For a well-level profile to be categorized as high-confidence (in either the correct or incorrect directions), it needed to score beyond the 95% confidence interval of the randomly permuted data range. For example, a high-confidence incorrect resistant profile would have a Bortezomib Signature below 95% confidence interval of the randomly permuted data. To evaluate the features driving the differences in these samples, we applied two-sample Kolmogorov–Smirnov (KS) tests per Bortezomib Signature feature. We applied these tests to two separate groups: (1) misclassified bortezomib-sensitive vs. high-confidence accurate bortezomib-sensitive and (2) misclassified bortezomib-resistant vs. high-confidence accurate bortezomib-resistant.”

The code change to generate the UMAP coordinates and figure is available at https://github.com/broadinstitute/profiling-resistance-mechanisms/pull/130

[Response to Reviewer 2 - Minor Comments 1 and 2]: These are interesting suggestions! Still, we prefer not to speculate on the biological mechanism of the Bortezomib signature.

Connecting morphological features identified as contributing to the Bortezomib Signature by Cell Painting to specific biological pathways would demand considerable cell-based assays to validate. In addition, our analyses suggest that the features contributing to the Bortezomib Signature are spread across a range of cellular compartments and channels, making it difficult to pin down specific mechanisms or pathways as likely contributors to bortezomib resistance. However, we are adding a figure to increase interpretability of the signature, which will aid in developing future hypotheses. Note that the signature was not possible to detect by eye (Fig. 2 A).

Reviewer 2 - Minor Comment 1:

There could be some speculation on the mechanism of Bortezomib resistance concerning the literature with the existing image data. For example, Bortezomib resistance is connected to serine synthesis and how a particular feature could contribute to the known mechanism.

Reviewer 2 - Minor Comment 2:

Along the same lines, the authors could show that larger cells lead to resistance with microscopic images.

[Response to Reviewer 2 - Major Comment 8]: We appreciate the reviewer’s concern that our work using HCT116 clonal cells lines may not directly reflect results from patient samples. Our choice was based on previously published work demonstrating the efficiency with which HCT116 cells generate resistant clones due to diminished DNA mismatch repair and decreased expression of drug efflux pumps. Since our work is a proof of concept rather than a comprehensive demonstration of translating morphological profiling into clinical practice, we believe that experiments using multiple patient cell lines from different tissues as well as digital pathology records to be beyond the scope of this work. We instead chose to tone down the language of our manuscript to more clearly acknowledge the limitations of our work and clarify this as a proof of concept.

Reviewer 2 - Major Comment 8 (relevant excerpt):

I suggest the authors test their approach on at least two other cell lines (maybe from different tissues) and benchmark their results against a dataset of digital pathology where such predictions are made from stained and analyzed tissue slices. This way, after a thorough benchmark against related third-party data sets, the method would significantly gain relevance, the paper would appeal to a broader audience, and the advance gains more merit.

[Response to Reviewer 3 - Major Comment 2]: The bortezomib sensitivity of ixazomib- and CB-5083-resistant clones was not determined, and hence can not be ruled out as a possible explanation for their high Bortezomib Signature scores. However, we prefer not to conduct additional proliferation assays for the misclassified clones (IX02, WT06, CB14, CB16) in the presence of bortezomib to determine whether coincidental bortezomib resistance might explain the signature performance. Our rationale is that three other misclassified clones (WT10, WT15, and BZ06) had the expected bortezomib sensitivity in proliferation assays (Fig. 1-Supplement 2), meaning that additional proliferation assays may not reveal any insights regarding the signature performance.

Reviewer 3 - Major Comment 2:

Was the bortezomib sensitivity of the IX (or CB) resistant cell lines determined? If there were differences, this could explain some of the variation in the morphological signatures. This could be easily done in one or two growth experiments.

[Response to Reviewer 2 - Major Comment 7]: Thank you for pointing this out. Our goal is to keep the study multi-disciplinary. We are adding a figure to increase interpretability of the signature, and adding text-based clarifications.

Reviewer 2 - Major Comment 7 (relevant excerpt):

While reviewing, I thus wondered which audience the authors targeted with their manuscript. A more focused analysis of their data that highlights aspects of the study either for the machine learning community, the cell biology community, or the precision oncology community would greatly benefit the manuscript's impact. In its current form, the study's findings seem diluted and spread across a wide range of research questions.

[Response to Reviewer 2 and 3 - Major Comments 6 and 4]: We prefer not to expand the scope of the model to predict other drug signatures. This would require a substantial amount of work to generate the appropriate drug-resistant clones, collect the imaging data, and analyze it, and we think it important to convey the purpose of our paper is proof of concept. We do not feel that the time invested in performing this analysis would result in adequate returns beyond what we already demonstrate.

Reviewer 2 - Major Comment 6.

Interestingly, the Bortezomib signature is specific to the drug and not a broad range of proteasomal inhibitors. However, seeing the common features between all the proteasomal inhibitors would be interesting.

Reviewer 3 - Major Comment 4

There was some predictive ability of the Bortezomib Signature for ixazomib resistance. Were there some features that were correlated with IX-resistance, i.e. UPS pathway, versus specific to bortezomib? Do the features suggest anything about resistance mechanisms or is the feature set too abstruse to interpret?

References

Foroutan, Momeneh, Dharmesh D. Bhuva, Ruqian Lyu, Kristy Horan, Joseph Cursons, and Melissa J. Davis. 2018. “Single Sample Scoring of Molecular Phenotypes.” BMC Bioinformatics 19 (1): 404.

Heiser, Katie, Peter F. McLean, Chadwick T. Davis, Ben Fogelson, Hannah B. Gordon, Pamela Jacobson, Brett Hurst, et al. 2020. “Identification of Potential Treatments for COVID-19 through Artificial Intelligence-Enabled Phenomic Analysis of Human Cells Infected with SARS-CoV-2.” bioRxiv. https://doi.org/10.1101/2020.04.21.054387.

McInnes, Leland, John Healy, Nathaniel Saul, and Lukas Großberger. 2018. “UMAP: Uniform Manifold Approximation and Projection.” Journal of Open Source Software 3 (29): 861.

Nyffeler, Johanna, Clinton Willis, Ryan Lougee, Ann Richard, Katie Paul-Friedman, and Joshua A. Harrill. 2020. “Bioactivity Screening of Environmental Chemicals Using Imaging-Based High-Throughput Phenotypic Profiling.” Toxicology and Applied Pharmacology 389 (January): 114876.

Rohban, Mohammad Hossein, Shantanu Singh, Xiaoyun Wu, Julia B. Berthet, Mark-Anthony Bray, Yashaswi Shrestha, Xaralabos Varelas, Jesse S. Boehm, and Anne E. Carpenter. 2017. “Systematic Morphological Profiling of Human Gene and Allele Function via Cell Painting.” eLife 6 (March). https://doi.org/10.7554/eLife.24060.

Simm, Jaak, Günter Klambauer, Adam Arany, Marvin Steijaert, Jörg Kurt Wegner, Emmanuel Gustin, Vladimir Chupakhin, et al. 2018. “Repurposing High-Throughput Image Assays Enables Biological Activity Prediction for Drug Discovery.” Cell Chemical Biology 25 (5): 611–18.e3.

Wacker, Sarah A., Benjamin R. Houghtaling, Olivier Elemento, and Tarun M. Kapoor. 2012. “Using Transcriptome Sequencing to Identify Mechanisms of Drug Action and Resistance.” Nature Chemical Biology 8 (3): 235–37.

Wawer, Mathias J., Kejie Li, Sigrun M. Gustafsdottir, Vebjorn Ljosa, Nicole E. Bodycombe, Melissa A. Marton, Katherine L. Sokolnicki, et al. 2014. “Toward Performance-Diverse Small-Molecule Libraries for Cell-Based Phenotypic Screening Using Multiplexed High-Dimensional Profiling.” Proceedings of the National Academy of Sciences of the United States of America 111 (30): 10911–16.

Way, Gregory, Yu Han, David Stirling, and Shantanu Singh. 2023. Broadinstitute/profiling-Resistance-Mechanisms: Analysis for Preprint. Zenodo. https://doi.org/10.5281/ZENODO.7803787.

Way, Gregory P., Maria Kost-Alimova, Tsukasa Shibue, William F. Harrington, Stanley Gill, Federica Piccioni, Tim Becker, et al. 2021. “Predicting Cell Health Phenotypes Using Image-Based Morphology Profiling.” Molecular Biology of the Cell 32 (9): 995–1005.

https://doi.org/10.7554/eLife.91362.sa2

Article and author information

Author details

Megan E Kelley

Laboratory of Chemistry and Cell Biology, The Rockefeller University, New York City, United States

Contribution
Conceptualization, Data curation, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing – original draft, Writing – review and editing

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0002-0251-5054
Adi Y Berman

Laboratory of Chemistry and Cell Biology, The Rockefeller University, New York City, United States

Contribution
Conceptualization, Data curation, Formal analysis, Validation, Investigation, Methodology, Writing – review and editing

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0002-1460-7562
David R Stirling

Imaging Platform, Broad Institute, Cambridge, United States

Contribution
Data curation, Formal analysis, Validation, Investigation, Methodology, Writing – review and editing

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0001-6802-4103
Beth A Cimini

Imaging Platform, Broad Institute, Cambridge, United States

Contribution
Data curation, Software, Investigation, Writing – review and editing

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0001-9640-9318
Yu Han

Imaging Platform, Broad Institute, Cambridge, United States

Contribution
Software, Formal analysis, Writing – review and editing

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0001-5507-9228
Shantanu Singh

Imaging Platform, Broad Institute, Cambridge, United States

Contribution
Conceptualization, Resources, Supervision, Investigation, Project administration, Writing – review and editing

Contributed equally with
Anne E Carpenter, Tarun M Kapoor and Gregory P Way

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0003-3150-3025
Anne E Carpenter

Imaging Platform, Broad Institute, Cambridge, United States

Contribution
Conceptualization, Resources, Supervision, Funding acquisition, Investigation, Project administration, Writing – review and editing

Contributed equally with
Shantanu Singh, Tarun M Kapoor and Gregory P Way

For correspondence
anne@broadinstitute.org

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0003-1555-8261
Tarun M Kapoor

Laboratory of Chemistry and Cell Biology, The Rockefeller University, New York City, United States

Contribution
Conceptualization, Resources, Supervision, Funding acquisition, Investigation, Project administration, Writing – review and editing

Contributed equally with
Shantanu Singh, Anne E Carpenter and Gregory P Way

For correspondence
kapoor@rockefeller.edu

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0003-0628-211X
Gregory P Way
1. Imaging Platform, Broad Institute, Cambridge, United States
2. Department of Biomedical Informatics, University of Colorado Anschutz Medical Campus, Aurora, United States
Contribution
Conceptualization, Data curation, Software, Formal analysis, Supervision, Validation, Investigation, Visualization, Methodology, Writing – original draft, Project administration, Writing – review and editing

Contributed equally with
Shantanu Singh, Anne E Carpenter and Tarun M Kapoor

For correspondence
Gregory.way@cuanschutz.edu

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0002-0503-9348

Funding

Starr Cancer Consortium (112-0039)

Anne E Carpenter

National Institutes of Health (R35 GM122547)

Anne E Carpenter

National Institutes of Health (R35 GM130234)

Tarun M Kapoor

National Institutes of Health (T32 GM066699)

Megan E Kelley

National Institutes of Health (T32 GM115327)

Adi Y Berman

National Science Foundation (NSF GRFP 2019272977)

Adi Y Berman

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

The authors gratefully acknowledge funding from the Starr Cancer Consortium (112-0039 to TMK and AEC), the National Institutes of Health (NIH MIRA R35 GM122547 to AEC, NIH MIRA R35 GM130234 to TMK, NIH NRSA T32 GM066699 to MEK, NIH T32 GM115327 Chemistry-Biology Interface Training Grant to the Tri-Institutional PhD Program in Chemical Biology to AYB), and the National Science Foundation (NSF GRFP 2019272977 to AYB). We thank Erin Weisbart for uploading image data to the public database. We followed the International Committee of Medical Journal Editors (ICMJE) recommendations for authors and contributors.

Senior and Reviewing Editor

Caigang Liu, Shengjing Hospital of China Medical University, China

Version history

Preprint posted: May 2, 2023 (view preprint)
Received: July 27, 2023
Accepted: September 13, 2023
Accepted Manuscript published: September 27, 2023 (version 1)
Version of Record published: October 18, 2023 (version 2)

Copyright

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.