Machine Learning: Finding the right type of cell
Since the advent of microscopy in the 17th century, it has become well established that organisms are divided into tissues made up of different types of cells, with cells of the same type typically performing the same role. This simplifies the task of understanding a biological system immensely, as there are many fewer cell types than individual cells (Masland, 2001).
Categorizing tissues and cell types has always been done manually, usually by grouping cells that look the same based on their shape, internal structures and various other features. This is also true for images collected using modern day techniques, such as electron microscopy, which can provide three-dimensional reconstructions of tissue samples, or even entire small organisms less than a millimeter cube in size.
While electron microscopy images can be automatically subdivided or ‘segmented’ into individual cells, assigning each one to a cell type by hand is both difficult and time consuming; for example, in a recent project, it took several experts many months to categorize one half of the fruit fly brain (Scheffer et al., 2020). The whole task becomes even more challenging if the object being studied is not a well-known model organism. Now, in eLife, Valentyna Zinchenko, Johannes Hugger, Virginie Uhlmann Detlev Arendt and Anna Kreshuk of the European Molecular Biology Laboratory report a new method that could simplify this process (Zinchenko et al., 2023).
Zinchenko et al. based their program on a machine learning method called unsupervised, contrastive learning (van den Oord et al., 2018). The program works by grouping cell types without having received prior examples of ‘human-classified’ cell types or features (i.e., it is unsupervised) and by finding features that maximize the difference (or contrast) between examples that should be grouped together and those that should not. The method requires many examples, both of cells that should be grouped together, and those that should not. For the positive examples (those that should be grouped together), Zinchenko et al. created synthetic copies of each existing cell with minor modifications, such as different rotations, reflections, and texture or shape variations. In this case, the original cell and the modified cell should be grouped together. For negative examples (those that are of different types), they picked pairs of cells at random from their sample. This will be wrong occasionally but it is sufficiently accurate to train their model while allowing unsupervised operation.
Machine learning was then applied to find features shared by the positive examples only. The system combined the learned features of each original cell into a vector that summarizes the cell’s shape and texture. The team found that cells belonging to the same type were close together within the space of the vector, which can be visualized and interpreted by existing dimension reduction techniques (McInnes et al., 2018).
Zinchenko et al. then tested their model on a three-dimensional reconstruction of the annelid worm (Platynereis dumerilii) obtained through electron microscopy. Their computer model was able to match the different cell types and could identify subgroups of cells that could not be distinguished using human-specified features. Moreover, when compared to a gene expression map of the whole animal, the cells that had been classified as similar based on their features also shared similar genetic signatures, more so than cells that had previously been clustered using ‘human-designed’ features.
Next, they extended their classification method to consider both the shape and texture of each cell, and a combination of these features of all physically adjacent cells. Grouping these enhanced features revealed different tissues and organs within the animal. The classification system of the model strongly agreed with human results, but also found subtle tissue distinctions and rare features that had been overlooked by humans examining the same data set. For example, the analyses revealed a specific type of neuron in the midgut region of the worm, which had previously not been confirmed to be located in this area of the body.
The ‘unsupervised’ aspect of the method created by Zinchenko et al. is critical because it means the program does not require a full library of the relevant cell types (or a full list of the features that can distinguish between the cell types). Instead, the program learns these characteristics from the data itself (Figure 1). This is particularly useful for systems where the cell types are not known. Moreover, it is not restricted to using cell features that humans deem important, such as the roundness of a cell or the presence of dark vesicles. This means that the model can often outperform humans and work without bias, as it is not told what to expect and is thus less likely to overlook rare or unexpected cell types.
Electron microscopy and related techniques provide an incredible level of detail, including the shape, location and structure of every cell. But analyzing this flood of data by hand is nearly impossible and automated techniques are desperately needed to unlock the potential of these findings (Eberle and Zeidler, 2018). Significant progress has already been made in turning some tasks, such as cell segmentation and identifying synapses, into automatic processes, leaving cell and tissue identification as some of the most time-consuming manual steps (Januszewski et al., 2018; Huang et al., 2018). By helping to automate this step, Zinchenko et al. make a critical step in the journey of understanding these invaluable but intimidating data sets.
Fully-automatic synapse prediction and validation on a large data setFrontiers in Neural Circuits 12:87.https://doi.org/10.3389/fncir.2018.00087
The fundamental plan of the retinaNature Neuroscience 4:877–886.https://doi.org/10.1038/nn0901-877
Article and author information
- Version of Record published: February 16, 2023 (version 1)
© 2023, Scheffer
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.
- Page views
Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
- Computational and Systems Biology
Biological age, distinct from an individual's chronological age, has been studied extensively through predictive aging clocks. However, these clocks have limited accuracy in short time-scales. Here we trained deep learning models on fundus images from the EyePACS dataset to predict individuals' chronological age. Our retinal aging clocking, 'eyeAge', predicted chronological age more accurately than other aging clocks (mean absolute error of 2.86 and 3.30 years on quality-filtered data from EyePACS and UK Biobank, respectively). Additionally, eyeAge was independent of blood marker-based measures of biological age, maintaining an all-cause mortality hazard ratio of 1.026 even when adjusted for phenotypic age. The individual-specific nature of eyeAge was reinforced via multiple GWAS hits in the UK Biobank cohort. The top GWAS locus was further validated via knockdown of the fly homolog, Alk, which slowed age-related decline in vision in flies. This study demonstrates the potential utility of a retinal aging clock for studying aging and age-related diseases and quantitatively measuring aging on very short time-scales, opening avenues for quick and actionable evaluation of gero-protective therapeutics.
- Computational and Systems Biology
Computational models starting from large ensembles of evolutionarily related protein sequences capture a representation of protein families and learn constraints associated to protein structure and function. They thus open the possibility for generating novel sequences belonging to protein families. Protein language models trained on multiple sequence alignments, such as MSA Transformer, are highly attractive candidates to this end. We propose and test an iterative method that directly employs the masked language modeling objective to generate sequences using MSA Transformer. We demonstrate that the resulting sequences score as well as natural sequences, for homology, coevolution, and structure-based measures. For large protein families, our synthetic sequences have similar or better properties compared to sequences generated by Potts models, including experimentally validated ones. Moreover, for small protein families, our generation method based on MSA Transformer outperforms Potts models. Our method also more accurately reproduces the higher-order statistics and the distribution of sequences in sequence space of natural data than Potts models. MSA Transformer is thus a strong candidate for protein sequence generation and protein design.