The biological function of a protein is often determined by its distinct functional units, such as folded domains and intrinsically disordered regions. Identifying and categorizing these protein segments from sequence has been a major focus in computational biology which has enabled the automatic annotation of folded protein domains. Here we show that embeddings from the unsupervised protein language model ProtT5 can be used to identify and categorize protein segments without relying on conserved patterns in primary amino acid sequence. We present Zero-shot Protein Segmentation (ZPS), where we use embeddings from ProtT5 to predict the boundaries of protein segments without training or fine-tuning any parameters. We find that ZPS boundary predictions for the human proteome are better at reproducing reviewed annotations from UniProt than established bioinformatics tools and ProtT5 embeddings of ZPS segments can categorize over 200 of the most common UniProt annotations in the human proteome, including folded domains, sub-domains, and intrinsically disordered regions. To explore ZPS predictions, we introduce a new way to visualize protein embeddings that closely resembles diagrams of distinct functional units in protein biology. Since ZPS and segment embeddings can be used without training or fine-tuning, the approach is not biased towards known annotations and can be used to identify and categorize unannotated protein segments. We used the segment embeddings to identify unannotated mitochondrion targeting signals and SYGQ-rich prion-like domains, which are functional regions within intrinsically disordered regions. We expect that the analysis of protein segment embedding similarity can lead to valuable information about protein function, including about intrinsically disordered regions and poorly understood protein regions.

MOTIVATION: Supervised deep learning is used to model the complex relationship between genomic sequence and regulatory function. Understanding how these models make predictions can provide biological insight into regulatory functions. Given the complexity of the sequence to regulatory function mapping (the cis-regulatory code), it has been suggested that the genome contains insufficient sequence variation to train models with suitable complexity. Data augmentation is a widely used approach to increase the data variation available for model training, however current data augmentation methods for genomic sequence data are limited. RESULTS: Inspired by the success of comparative genomics, we show that augmenting genomic sequences with evolutionarily related sequences from other species, which we term phylogenetic augmentation, improves the performance of deep learning models trained on regulatory genomic sequences to predict high-throughput functional assay measurements. Additionally, we show that phylogenetic augmentation can rescue model performance when the training set is down-sampled and permits deep learning on a real-world small dataset, demonstrating that this approach improves data efficiency. Overall, this data augmentation method represents a solution for improving model performance that is applicable to many supervised deep-learning problems in genomics. AVAILABILITY AND IMPLEMENTATION: The open-source GitHub repository agduncan94/phylogenetic_augmentation_paper includes the code for rerunning the analyses here and recreating the figures.

A major challenge to the characterization of intrinsically disordered regions (IDRs), which are widespread in the proteome, but relatively poorly understood, is the identification of molecular features that mediate functions of these regions, such as short motifs, amino acid repeats and physicochemical properties. Here, we introduce a proteome-scale feature discovery approach for IDRs. Our approach, which we call "reverse homology", exploits the principle that important functional features are conserved over evolution. We use this as a contrastive learning signal for deep learning: given a set of homologous IDRs, the neural network has to correctly choose a held-out homolog from another set of IDRs sampled randomly from the proteome. We pair reverse homology with a simple architecture and standard interpretation techniques, and show that the network learns conserved features of IDRs that can be interpreted as motifs, repeats, or bulk features like charge or amino acid propensities. We also show that our model can be used to produce visualizations of what residues and regions are most important to IDR function, generating hypotheses for uncharacterized IDRs. Our results suggest that feature discovery using unsupervised neural networks is a promising avenue to gain systematic insight into poorly understood protein sequences.

Constraint-based modeling is a powerful framework for studying cellular metabolism, with applications ranging from predicting growth rates and optimizing production of high value metabolites to identifying enzymes in pathogens that may be targeted for therapeutic interventions. Results from modeling experiments can be affected at least in part by the quality of the metabolic models used. Reconstructing a metabolic network manually can produce a high-quality metabolic model but is a time-consuming task. At the same time, current methods for automating the process typically transfer metabolic function based on sequence similarity, a process known to produce many false positives. We created Architect, a pipeline for automatic metabolic model reconstruction from protein sequences. First, it performs enzyme annotation through an ensemble approach, whereby a likelihood score is computed for an EC prediction based on predictions from existing tools; for this step, our method shows both increased precision and recall compared to individual tools. Next, Architect uses these annotations to construct a high-quality metabolic network which is then gap-filled based on likelihood scores from the ensemble approach. The resulting metabolic model is output in SBML format, suitable for constraints-based analyses. Through comparisons of enzyme annotations and curated metabolic models, we demonstrate improved performance of Architect over other state-of-the-art tools, notably with higher precision and recall on the eukaryote C. elegans and when compared to UniProt annotations in two bacterial species. Code for Architect is available at https://github.com/ParkinsonLab/Architect. For ease-of-use, Architect can be readily set up and utilized using its Docker image, maintained on Docker Hub.

Understanding if classifiers generalize to out-of-sample datasets is a central problem in machine learning. Microscopy images provide a standardized way to measure the generalization capacity of image classifiers, as we can image the same classes of objects under increasingly divergent, but controlled factors of variation. We created a public dataset of 132,209 images of mouse cells, COOS-7 (Cells Out Of Sample 7-Class). COOS-7 provides a classification setting where four test datasets have increasing degrees of covariate shift: some images are random subsets of the training data, while others are from experiments reproduced months later and imaged by different instruments. We benchmarked a range of classification models using different representations, including transferred neural network features, end-to-end classification with a supervised deep CNN, and features from a self-supervised CNN. While most classifiers perform well on test datasets similar to the training dataset, all classifiers failed to generalize their performance to datasets with greater covariate shifts. These baselines highlight the challenges of covariate shifts in image data, and establish metrics for improving the generalization capacity of image classifiers.

SUMMARY: We introduce YeastSpotter, a web application for the segmentation of yeast microscopy images into single cells. YeastSpotter is user-friendly and generalizable, reducing the computational expertise required for this critical preprocessing step in many image analysis pipelines. AVAILABILITY AND IMPLEMENTATION: YeastSpotter is available at http://yeastspotter.csb.utoronto.ca/. Code is available at https://github.com/alexxijielu/yeast_segmentation. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Mammalian genomes can contain thousands of enhancers but only a subset are actively driving gene expression in a given cellular context. Integrated genomic datasets can be harnessed to predict active enhancers. One challenge in integration of large genomic datasets is the increasing heterogeneity: continuous, binary and discrete features may all be relevant. Coupled with the typically small numbers of training examples, semi-supervised approaches for heterogeneous data are needed; however, current enhancer prediction methods are not designed to handle heterogeneous data in the semi-supervised paradigm. RESULTS: We implemented a Dirichlet Process Heterogeneous Mixture model that infers Gaussian, Bernoulli and Poisson distributions over features. We derived a novel variational inference algorithm to handle semi-supervised learning tasks where certain observations are forced to cluster together. We applied this model to enhancer candidates in mouse heart tissues based on heterogeneous features. We constrained a small number of known active enhancers to appear in the same cluster, and 47 additional regions clustered with them. Many of these are located near heart-specific genes. The model also predicted 1176 active promoters, suggesting that it can discover new enhancers and promoters. AVAILABILITY AND IMPLEMENTATION: We created the 'dphmix' Python package: https://pypi.org/project/dphmix/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Cellular microscopy images contain rich insights about biology. To extract this information, researchers use features, or measurements of the patterns of interest in the images. Here, we introduce a convolutional neural network (CNN) to automatically design features for fluorescence microscopy. We use a self-supervised method to learn feature representations of single cells in microscopy images without labelled training data. We train CNNs on a simple task that leverages the inherent structure of microscopy images and controls for variation in cell morphology and imaging: given one cell from an image, the CNN is asked to predict the fluorescence pattern in a second different cell from the same image. We show that our method learns high-quality features that describe protein expression patterns in single cells both yeast and human microscopy datasets. Moreover, we demonstrate that our features are useful for exploratory biological analysis, by capturing high-resolution cellular components in a proteome-wide cluster analysis of human proteins, and by quantifying multi-localized proteins and single-cell variability. We believe paired cell inpainting is a generalizable method to obtain feature representations of single cells in multichannel microscopy images.

BACKGROUND: Crm1-dependent Nuclear Export Signals (NESs) are clusters of alternating hydrophobic and non-hydrophobic amino acid residues between 10 to 15 amino acids in length. NESs were largely thought to follow simple consensus patterns, based on which they were categorized into 6-10 classes. However, newly discovered NESs often deviate from the established consensus patterns. Thus, identifying NESs within protein sequences remains a bioinformatics challenge. RESULTS: We describe a probabilistic representation of NESs using a new generative model we call NoLogo that can account for a large diversity of NESs. Using this model to predict NESs, we demonstrate improved performance over PSSM and GLAM2 models, but do not achieve the performance of the state-of-the-art NES predictor LocNES. Our findings illustrate that over 30% of NESs are best described by novel NES classes rather than the 6-10 classes proposed by current/existing models. Finally, many NESs have additional hydrophobic residues either upstream or downstream of the canonical four residues, suggesting possible functionality. CONCLUSION: Applying the NoLogo model highlights the observation that NESs are more diverse than previously appreciated. Our work questions the practice of assigning each NES to one of several predefined NES classes. Finally, our analysis suggests a novel and testable biophysical perspective on interaction between Crm1 receptor and Crm1-dependent NESs.

Abstract Motivation: Quantifying variability in protein expression is a major goal of systems biology and cell-to-cell variability in subcellular localization pattern has not been systematically quantified. Results: We define a local measure to quantify cell-to-cell variability in high-throughput microscope images and show that it allows comparable measures of variability for proteins with diverse subcellular localizations. We systematically estimate cell-to-cell variability in the yeast GFP collection and identify examples of proteins that show cell-to-cell variability in their subcellular localization. Conclusions: Automated image analysis methods can be used to quantify cell-to-cell variability in microscope images. Contact: [email protected] Availability and Implementation: Software and data are available at http://www.moseslab.csb.utoronto.ca/louis-f/ Supplementary information: Supplementary data are available at Bioinformatics online.

Gene duplication is an important evolutionary mechanism that can result in functional divergence in paralogs due to neo-functionalization or sub-functionalization. Consistent with functional divergence after gene duplication, recent studies have shown accelerated evolution in retained paralogs. However, little is known in general about the impact of this accelerated evolution on the molecular functions of retained paralogs. For example, do new functions typically involve changes in enzymatic activities, or changes in protein regulation? Here we study the evolution of posttranslational regulation by examining the evolution of important regulatory sequences (short linear motifs) in retained duplicates created by the whole-genome duplication in budding yeast. To do so, we identified short linear motifs whose evolutionary constraint has relaxed after gene duplication with a likelihood-ratio test that can account for heterogeneity in the evolutionary process by using a non-central chi-squared null distribution. We find that short linear motifs are more likely to show changes in evolutionary constraints in retained duplicates compared to single-copy genes. We examine changes in constraints on known regulatory sequences and show that for the Rck1/Rck2, Fkh1/Fkh2, Ace2/Swi5 paralogs, they are associated with previously characterized differences in posttranslational regulation. Finally, we experimentally confirm our prediction that for the Ace2/Swi5 paralogs, Cbk1 regulated localization was lost along the lineage leading to SWI5 after gene duplication. Our analysis suggests that changes in posttranslational regulation mediated by short regulatory motifs systematically contribute to functional divergence after gene duplication.

Protein subcellular localization has been systematically characterized in budding yeast using fluorescently tagged proteins. Based on the fluorescence microscopy images, subcellular localization of many proteins can be classified automatically using supervised machine learning approaches that have been trained to recognize predefined image classes based on statistical features. Here, we present an unsupervised analysis of protein expression patterns in a set of high-resolution, high-throughput microscope images. Our analysis is based on 7 biologically interpretable features which are evaluated on automatically identified cells, and whose cell-stage dependency is captured by a continuous model for cell growth. We show that it is possible to identify most previously identified localization patterns in a cluster analysis based on these features and that similarities between the inferred expression patterns contain more information about protein function than can be explained by a previous manual categorization of subcellular localization. Furthermore, the inferred cell-stage associated to each fluorescence measurement allows us to visualize large groups of proteins entering the bud at specific stages of bud growth. These correspond to proteins localized to organelles, revealing that the organelles must be entering the bud in a stereotypical order. We also identify and organize a smaller group of proteins that show subtle differences in the way they move around the bud during growth. Our results suggest that biologically interpretable features based on explicit models of cell morphology will yield unprecedented power for pattern discovery in high-resolution, high-throughput microscopy images.

MOTIVATION: Protein kinases represent critical links in cell signaling. A central problem in computational biology is to systematically identify their substrates. RESULTS: This study introduces a new method to predict kinase substrates by extracting evolutionary information from multiple sequence alignments in a manner that is tolerant to degenerate motif positioning. Given a known consensus, the new method (ConDens) compares the observed density of matches to a null model of evolution and does not require labeled training data. We confirmed that ConDens has improved performance compared with several existing methods in the field. Further, we show that it is generalizable and can predict interesting substrates for several important eukaryotic kinases where training data is not available. AVAILABILITY AND IMPLEMENTATION: ConDens can be found at http://www.moseslab.csb.utoronto.ca/andyl/. CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: Detection of false-positive motifs is one of the main causes of low performance in de novo DNA motif-finding methods. Despite the substantial algorithm development effort in this area, recent comprehensive benchmark studies revealed that the performance of DNA motif-finders leaves room for improvement in realistic scenarios. RESULTS: Using large-deviations theory, we derive a remarkably simple relationship that describes the dependence of false positives on dataset size for the one-occurrence per sequence motif-finding problem. As expected, we predict that false-positives can be reduced by decreasing the sequence length or by adding more sequences to the dataset. Interestingly, we find that the false-positive strength depends more strongly on the number of sequences in the dataset than it does on the sequence length, but that the dependence on the number of sequences diminishes, after which adding more sequences does not reduce the false-positive rate significantly. We compare our theoretical predictions by applying four popular motif-finding algorithms that solve the one-occurrence-per-sequence problem (MEME, the Gibbs Sampler, Weeder, and GIMSAN) to simulated data that contain no motifs. We find that the dependence of false positives detected by these softwares on the motif-finding parameters is similar to that predicted by our formula. CONCLUSIONS: We quantify the relationship between the sequence search space and motif-finding false-positives. Based on the simple formula we derive, we provide a number of intuitive rules of thumb that may be used to enhance motif-finding results in practice. Our results provide a theoretical advance in an important problem in computational biology.

BACKGROUND: Genetic variations contribute to normal phenotypic differences as well as diseases, and new sequencing technologies are greatly increasing the capacity to identify these variations. Given the large number of variations now being discovered, computational methods to prioritize the functional importance of genetic variations are of growing interest. Thus far, the focus of computational tools has been mainly on the prediction of the effects of amino acid changing single nucleotide polymorphisms (SNPs) and little attention has been paid to indels or nonsense SNPs that result in premature stop codons. RESULTS: We propose computational methods to rank insertion-deletion mutations in the coding as well as non-coding regions and nonsense mutations. We rank these variations by measuring the extent of their effect on biological function, based on the assumption that evolutionary conservation reflects function. Using sequence data from budding yeast and human, we show that variations which that we predict to have larger effects segregate at significantly lower allele frequencies, and occur less frequently than expected by chance, indicating stronger purifying selection. Furthermore, we find that insertions, deletions and premature stop codons associated with disease in the human have significantly larger predicted effects than those not associated with disease. Interestingly, the large-effect mutations associated with disease show a similar distribution of predicted effects to that expected for completely random mutations. CONCLUSIONS: This demonstrates that the evolutionary conservation context of the sequences that harbour insertions, deletions and nonsense mutations can be used to predict and rank the effects of the mutations.

BACKGROUND: Nuclear localization signals (NLSs) are stretches of residues within a protein that are important for the regulated nuclear import of the protein. Of the many import pathways that exist in yeast, the best characterized is termed the 'classical' NLS pathway. The classical NLS contains specific patterns of basic residues and computational methods have been designed to predict the location of these motifs on proteins. The consensus sequences, or patterns, for the other import pathways are less well-understood. RESULTS: In this paper, we present an analysis of characterized NLSs in yeast, and find, despite the large number of nuclear import pathways, that NLSs seem to show similar patterns of amino acid residues. We test current prediction methods and observe a low true positive rate. We therefore suggest an approach using hidden Markov models (HMMs) to predict novel NLSs in proteins. We show that our method is able to consistently find 37% of the NLSs with a low false positive rate and that our method retains its true positive rate outside of the yeast data set used for the training parameters. CONCLUSION: Our implementation of this model, NLStradamus, is made available at: (http://www.moseslab.csb.utoronto.ca/NLStradamus/).

BACKGROUND: Molecular evolutionary studies of noncoding sequences rely on multiple alignments. Yet how multiple alignment accuracy varies across sequence types, tree topologies, divergences and tools, and further how this variation impacts specific inferences, remains unclear. RESULTS: Here we develop a molecular evolution simulation platform, CisEvolver, with models of background noncoding and transcription factor binding site evolution, and use simulated alignments to systematically examine multiple alignment accuracy and its impact on two key molecular evolutionary inferences: transcription factor binding site conservation and divergence estimation. We find that the accuracy of multiple alignments is determined almost exclusively by the pairwise divergence distance of the two most diverged species and that additional species have a negligible influence on alignment accuracy. Conserved transcription factor binding sites align better than surrounding noncoding DNA yet are often found to be misaligned at relatively short divergence distances, such that studies of binding site gain and loss could easily be confounded by alignment error. Divergence estimates from multiple alignments tend to be overestimated at short divergence distances but reach a tool specific divergence at which they cease to increase, leading to underestimation at long divergences. Our most striking finding was that overall alignment accuracy, binding site alignment accuracy and divergence estimation accuracy vary greatly across branches in a tree and are most accurate for terminal branches connecting sister taxa and least accurate for internal branches connecting sub-alignments. CONCLUSION: Our results suggest that variation in alignment accuracy can lead to errors in molecular evolutionary inferences that could be construed as biological variation. These findings have implications for which species to choose for analyses, what kind of errors would be expected for a given set of species and how multiple alignment tools and phylogenetic inference methods might be improved to minimize or control for alignment errors.

The gain and loss of functional transcription factor binding sites has been proposed as a major source of evolutionary change in cis-regulatory DNA and gene expression. We have developed an evolutionary model to study binding-site turnover that uses multiple sequence alignments to assess the evolutionary constraint on individual binding sites, and to map gain and loss events along a phylogenetic tree. We apply this model to study the evolutionary dynamics of binding sites of the Drosophila melanogaster transcription factor Zeste, using genome-wide in vivo (ChIP-chip) binding data to identify functional Zeste binding sites, and the genome sequences of D. melanogaster, D. simulans, D. erecta, and D. yakuba to study their evolution. We estimate that more than 5% of functional Zeste binding sites in D. melanogaster were gained along the D. melanogaster lineage or lost along one of the other lineages. We find that Zeste-bound regions have a reduced rate of binding-site loss and an increased rate of binding-site gain relative to flanking sequences. Finally, we show that binding-site gains and losses are asymmetrically distributed with respect to D. melanogaster, consistent with lineage-specific acquisition and loss of Zeste-responsive regulatory elements.

Background. Transcriptional regulation in eukaryotes is often multifactorial, involving multiple transcription factors binding to the same transcription control region (e.g., upstream activating sequences and enhancers), and to understand the regulatory content of eukaryotic genomes it is necessary to consider the co-occurrence and spatial relationships of individual binding sites. The identification of sequences conserved among related species (often known as phylogenetic footprinting) has been successfully used to identify individual transcription factor binding sites. Here, we extend this concept of functional conservation to higher-order features of transcription control regions involved in the multifactorial control of gene expression.Results. We used the genome sequences of four yeast species of the genus Saccharomyces to identify sequences potentially involved in multifactorial control of gene expression. We found 1,117 potential regulatory "templates": pairs of hexameric sequences that are jointly conserved in transcription regulatory regions and also exhibit non-random relative spacing. Many of the individual sequences in these templates correspond to known transcription factor binding sites, and the sets of genes containing a particular template in their transcription control regions tend to be differentially expressed in conditions where the corresponding transcription factors are known to be active.Conclusions. The incorporation of both joint conservation and spacing constraints of sequence pairs predicts groups of target genes that were specific for common patterns of gene expression. Our work suggests that positional information, especially the relative spacing between transcription factor binding sites, may represent a common organizing principle of transcription control regions.