Mapping variation in the morphological landscape of human cells with optical pooled CRISPRi screening

Reviewer #1 (Public Review):

Summary:
Theriot et al. are proposing here a technically very impressive screening method. Their optimization of single-cell sgRNA barcode sequencing/reading is fundamental progress towards the use of CRISPRi technology in phenotypic screening.

The biology side of the manuscript focuses on cell morphology and cytoskeleton. For this, others are also proposing innovative methods for phenotypic quantification and analysis. The output of the phenotypic analysis shows interesting hit correlations between the methods used and identifies well-known hit genes. Nevertheless, the strength and the validity of the results are yet difficult to assess. The complexity and the amount of features extracted from the cell images do not always seem justified. Indeed, the visual conclusion from the authors at the end mostly refers to basic features (cell size, shape, nuclear localization, actin network polarity), which in my opinion could be quantified in a more straightforward way, which then would facilitate the ultimate goal of such a work, which is the biological interpretation of the screening campaign.

Strengths:
A very impressive technology work on molecular biology, microscopy, image analysis, and data analysis. The investment of such efforts seems fundamental for the development of phenotypic and CRISPR screens.

Weaknesses:
The phenotypic analysis method seems too complex in regard to the actual output. The biological interpretation of the screen is therefore suffering from this complexity. Having said that the quantification of cell morphology and actin network phenotype is a very risky and complex task.

Reviewer #2 (Public Review):

Summary:
In this study, the authors present a robust pipeline that integrates high-content phenotypic imaging of a targeted pool of 366 CRISPRi-screened genes with in situ sequencing of single cells, achieving a resolution for 1.3 million cells. The application of this pipeline on the U2OS cell line effectively screens for nuclear and actin morphology changes. One study's strength lies in the utilization of a barcode system, enabling efficient imaging and genotype determination for 85% of cells. The authors employ two distinct approaches to delineate phenotypic changes. In the first approach, cells are characterized by approximately 1,000 morphological features, with dimensionality reduction via PCA using 25 principal components and a novel image sampling method called VIEWED (Visual Interpretation of Embedding by Constrained Walkthrough Sampling). The second approach employs a deep learning technique, specifically the Beta-variational encoder, to identify morphological differences, offering a generative AI approach for visualizing interpreted distinctions learned through the algorithm. While the Beta-variational encoder is deemed simpler to use and interpret, the classical PCA approach demonstrates superiority due to its heightened sensitivity in identifying more genes with phenotypic changes. Both methods, however, successfully identify shared phenotypic gene hits, showing consistent replication across multiple individual guides for each gene hit. Key phenotypic clusters are identified and replicated similarly by both the conventional PCA feature approach and the Beta-variational encoder approach.

Strengths:
- A novel barcode methodology for efficient genotyping via in situ sequencing, minimizing rounds of imaging and genotyping 85% of cells.
- Use of a beta variational autoencoder, generative AI approach to facilitate detection of morphological change in cells, gene hits, and phenotypic gene clusters.

Weaknesses:
Although the outcome is reproduced with 3 gRNA/gene, no biological replicate is presented and is as such limiting on convincing on reproducibility of the phenotypic detection approach.

The presented work is highly compelling as it employs an optical pooled CRISPRi screen, showcasing the capability to conduct pool screening beyond the typical frequency count of guides with the next-generation sequencing approach, effectively establishing a direct link between cell images and guide RNAs in the pool screen approach. This achievement, typically associated with arrayed screens, sets the study apart. Moreover, the study offers captivating images of individual cells that vividly portray convincing phenotypic changes. Additionally, the work effectively highlights the potency of generative AI in interpreting cell phenotypic changes detected by the algorithm. This aspect of the study is particularly relevant in the present time, as it introduces a potentially highly valuable methodology. Overall, the research provides a robust demonstration of innovative techniques and methodologies, contributing significantly to the field.

Reviewer #3 (Public Review):

Summary:
Pooled optical screening has recently emerged as a powerful approach to associate complex phenotypic information from microscope images with specific genetic perturbations at the single-cell level. This is achieved by amplifying and sequencing DNA barcodes within individual cells through in-situ sequencing. This paper leverages these advances in pooled screening technology to examine the effects of gene knockdowns on high-dimensional cell morphological phenotypes beyond binary readouts.

A key challenge is how to effectively distill meaningful phenotypic dimensions from information-rich image data to connect genotype to phenotype. By screening 366 genes using CRISPRi and analyzing tens of thousands of single-cell images, this paper provides insights into genetic regulators of morphology in osteosarcoma cells. In developing this screen and analyzing its readout, the authors make several notable contributions.

First, the authors tested and optimized molecular inversion probes (MIPs) to improve rolling circle amplification and barcode imaging. Through these optimization experiments, they identified a shortened MIP design that yielded 11-fold more visible amplicons, enabling more robust barcode readout from complex images. Second, the authors address several unresolved questions regarding how to work with single-cell images at this scale. A critical aspect of this is the need to develop analysis strategies using single-cell data rather than commonly used current methodologies that condense down to an agglomerated perturbation level cell morphology information. The authors compare morphological profiling using curated feature extraction and an unsupervised deep learning approach called a β-variational autoencoder on single-cell imaging data, suggesting that the latter can capture salient aspects of variation without requiring much human input. Finally, and perhaps more importantly, the authors develop an approach, Visual Interpretation of Embeddings by constrained Walkthrough Sampling (VIEWS), towards sampling cells at the end of distributions such as a principal component dimension in a reduction of curated features or a latent space dimension extracted from an autoencoder. This allows for a rapid and efficient way of understanding extremes of morphological profiles and allows for quick interpretability of extracted morphological signal which in turn assists with downstream functional understandings of groups of genes that similarly alter a cell's morphology.

Strengths and Weaknesses:
This is an interesting and rigorous paper that provides an important advance in conducting large-scale microscopy-based approaches. The methods development and computational analyses described in this paper are strong and innovative. However, the screening conducted in this paper did not identify a large number of modifiers of general U2OS cell morphology. As the authors rightly point out, several factors could contribute to the modest hit rate, including variable CRISPRi knockdown efficiency and limited phenotypic readout from just two imaging channels. Despite these limitations, the paper makes several key methodological contributions and in the opinion of this reviewer merits revision or benchmarking.