Biology subjects

Altman, R. B.

Publications and source records attributed to Altman, R. B..

At least 19 recordsLinked to original sources

bioRxiv2026

Inverse FoldDir: Structure-conditioned Protein Sequence Design by Dirichlet Flow Matching

Protein engineering has important implications in the bioeconomy, enabling applications in materials, medicine, and energy. A key challenge is designing protein sequences that have a specific form and function. Protein inverse folding seeks to address this challenge by identifying amino acid sequences compatible with a desired protein backbone. This task is central to protein redesign and can provide a sequence design capability for de novo backbones produced by structure-generation methods. Ideally, inverse folding can provide diverse sequence alternatives, fixed residues or motifs, soft biochemical preferences at selected positions, and candidates that remain experimentally useful. We developed Inverse FoldDir, a controllable inverse-folding method that performs iterative denoising on the amino acid probability simplex. Given a backbone structure, the model updates all positions jointly through a learned Dirichlet flow, supporting full sequence generation, fixed-residue inpainting, and user-defined soft residue priors. On the held-out CATH 4.2 test set, Inverse FoldDir achieved a mean TM-score of 84.5 (on a 0-100 scale) and a mean C RMSD of 1.76[A], compared with 83.3 and 1.86[A], respectively, for ESM-IF1, the strongest evaluated baseline on both metrics. Denoising trajectory analyses showed that positions commit at different rates and that some residues change identity late in generation, illustrating whole-sequence refinement rather than one-shot prediction or irreversible sequential decoding. We experimentally tested Inverse FoldDir in an anti-GFP nanobody redesign task, where two of 35 redesigned sequences retained reproducible sfGFP-binding signal across independent assay runs with approximately 43% sequence divergence from the native nanobody. Inverse FoldDir is a structure-conditioned protein redesign method that combines structural recovery, user control, experimental validation, and a natural route toward future property-guided sampling.

bioinformatics↗

bioRxiv2026

Homology-Based Variant-Effect Predictors Break Down on Cytochrome P450 Pharmacogenes

Cytochrome P450 (CYP) enzymes metabolize roughly three-quarters of clinically used drugs; genetic variation in these enzymes is a leading source of interindividual differences in drug response. Predicting a variants functional effect is therefore critical, yet the consequences of most CYP variants remain unknown. Many state-of-the-art variant-effect predictors rest on a homology-based paradigm that scores variants by evolutionary conservation -- an assumption that pharmacogenes including CYPs violate. Indeed, focusing on human CYPs, we show that homology-based models fail systematically. AlphaMissense (AM) assigns variants to its "ambiguous" class at nearly twice the proteome-wide rate across six CYPs, and within that class the scores are essentially uncorrelated with CYP2C9 DMS activity (Spearmans{rho} = 0.069); Evolutionary Scale Modeling 2 (ESM-2) shows the same pattern. We further hypothesized that non-homology-based features (sequence position, substitution chemistry, binding-site distance, and secondary structure) might help resolve the ambiguous calls, but their explanatory power is weak. Using CYP2C9 DMS activity as ground truth, we built a k-nearest-neighbors model over ESM-2 embeddings and ensembled it with AM and ESM-2 masked marginal probability, improving the ambiguous-class correlation roughly ten-fold, from{rho} = 0.069 to 0.715 (overall{rho} from 0.638 to 0.825). However, this markedly improved accuracy does not translate into agreement with clinical annotations. Drawing on evidence that a variants effect can depend on the drug, we hypothesize that substrate identity is the key missing feature in current models, and that predicting function for these multi-substrate enzymes may require redefining function as substrate-conditioned.

bioinformatics↗

bioRxiv2026

Interpreting Protein Language Models: high attention sites predict functional regions

Computational proteomics has revolutionized biomedical research, guiding targeted experimental exploration to accelerate protein-based mechanistic discovery. Protein Language Models (PLMs) enable scalable, resource-efficient study; through large-scale training on only primary protein sequences, PLMs generate vector representations of protein structure that have been shown to capture biochemical, evolutionary, and structural properties. A core component of PLMs is the attention mechanism, which specifically captures long-range interactions across a protein sequence in attention matrices. Using the Evolutionary Scale Modelling 2 (ESM-2) PLM, we previously developed a novel method to identify "High Attention" (HA) sites. HA sites are specific residues that ESM-2 assigns the most attention to early during encoding. Here, we further characterize these HA sites across structural and functional metrics. Using unsupervised clustering, we find HA sites can be categorized as "structural core", "structural pathogenic", "core pathogenic", or "low-confidence". We further use AlphaMissense pathogenicity predictions and the pan-cancer analysis of whole genomes (PCAWG)-labeled pathogenic variant positions to show that HA sites predict protein regions with high pathogenic risk. Finally, we explore the utility of HA sites for suggesting candidate binding sites, identifying multiple cancer protein examples where HA sites identified regions with previously undiscovered high interaction likelihood and thus potential therapeutic utility. Our work demonstrates the biological interpretability of PLM representations and offers a valuable method to prioritize functionally relevant protein residues for targeted biomedical research.

bioinformatics↗

bioRxiv2026

Drug-Target Interaction Prediction with PIGLET

Drug-target interaction (DTI) prediction is a key task for computed-aided drug development that has been widely approached by deep learning models. Despite extremely high reported performance, these models have yet to find widespread success in accelerating real-world drug discovery. In contrast with the most common approach of creating embeddings from one-dimensional or three-dimensional representations of the input drug and input target, we create a novel graph transformer method for DTI prediction that operates on a proteome-wide knowledge graph of binding pocket similarity, protein-protein interactions, drug similarity, and known binding relationships. We benchmark our method, named PIGLET, against existing DTI prediction models on the Human dataset. We assess performance with two different splitting strategies: the frequently-reported random split, and a novel, more rigorous drug-based split. All models perform similarly well on the random split, and PIGLET outperforms all models on the drug-based split. We highlight the utility of PIGLET through a real-world drug discovery case study.

bioinformatics↗

bioRxiv2026

GATSBI: Improving context-aware protein embeddingsthrough biologically motivated data splits

MotivationUnderstanding protein function requires integrating diverse biological evidence while accounting for strong contextual dependence. Recent protein embedding methods increasingly leverage heterogeneous biological networks, yet their evaluation protocols often fail to reflect the specific biological tasks for which the embeddings are intended. Prediction of missing interactions, annotation of new proteins, and discovery of functional modules require fundamentally different data partitions, such as edge-masked versus node-held-out splits. Moreover, most approaches report performance primarily on well-studied proteins, where computational predictions are least needed, risking substantial overestimation of real-world utility. ResultsWe introduce a graph attention-based framework (Gatsbi) to construct context-aware protein embeddings from integrated protein-protein interactions, co-expression, sequence representations, and tissue-specific associations. Using task-aligned evaluation protocols, we show that models trained with biologically appropriate partitions achieve markedly better generalization. Across interaction, function, and functional set prediction, Gatsbi consistently outperforms existing pretrained embeddings for both well-studied and understudied proteins, with the largest gains observed for the understudied regime and under inductive node-held-out evaluation. To enable broad reuse, we provide the learned embeddings for download for application to other protein prediction tasks. Availability and Implementationhttps://github.com/Helix-Research-Lab/GATSBI-embedding

bioinformatics↗

bioRxiv2025

The Human Omnibus of Targetable Pockets

Hundreds of computational methods for predicting ligand binding pockets exist, but the problem of finding druggable pockets throughout the human proteome persists. Different strategies for pocket-finding excel in different use cases. Ensemble models that leverage multiple different pocket-finding strategies can best capture diverse pockets at scale. Despite this, no publicly available human-proteome-wide datasets of pocket predictions from multiple pocket-finding methods exist. We present the Human Omnibus of Targetable Pockets (HOTPocket), a dataset of over 2.4 million predicted pockets over the entire human proteome that utilizes both experimentally-determined and computationally-predicted protein structures. We assembled this dataset by running seven diverse, established pocket-finding methods over all PDB and AlphaFold2 structures of the canonical human proteome. We created a novel pocket scoring method, hotpocketNN, which we used to filter candidate pockets and assemble the final proteome-wide dataset. Our hotpocketNN method is able to recover known ligand binding pockets, including those which are dissimilar from any pocket seen in its training set. The hotpocketNN method outperforms all constituent methods, including P2Rank and Fpocket, when assessing the DCCcriterion on the Astex Diverse Set and PoseBusters dataset. Additionally, hotpocketNN was able to identify recently-discovered druggable pockets on KRAS and the mu opioid receptor. We make both the HOTPocket dataset and the hotpocketNN method freely available. Scientific ContributionWe introduce HOTPocket, a human proteome-wide dataset of known and predicted binding pockets from a diverse assortment of pocket-finding algorithms. We also introduce hotpocketNN, a machine learning model that scores candidate pockets, which we used to filter and harmonize the HOTPocket dataset. Both the data and the model are made publicly available for scientific use.

bioinformatics↗

bioRxiv2025

Exposing the Molecular Reaction Blind Spots of LLMs with PathwayQA

Proteins mediate a large portion of cellular activity, and understanding protein pathways can yield novel biological insights. Large language models (LLMs) have become increasingly adept at performing inference tasks across different fields of science and engineering. These models could facilitate the analysis of protein networks and help generate hypotheses about protein interactions in a scalable and accessible manner. However, the performance of LLMs in inferring protein-mediated biochemical reactions remains understudied. Here, we evaluate nine LLMs in reasoning over protein pathways included in the curated Reactome database. We find that all nine models struggle to infer products of a reaction when given reactants and enzymes. GPT-4o mini performed the best with a median recovery score of 0.6667, but no model surpassed the baseline strategy of parroting reactants back as predicted products. Most LLMs also performed poorly when inferring whether a protein pathway is associated with a human disease, with an average accuracy of 0.5980. DeepSeek 7B Chat performed the best with an accuracy of 0.9100. This study highlights an area where LLMs still struggle to make correct inferences and provides an opportunity for further work in developing biological LLMs. We also provide a novel question-answer dataset, PathwayQA, which is based on the Reactome database. PathwayQA can be used to benchmark and improve model performance on reasoning over protein-interaction networks. PathwayQA is available at https://github.com/Helix-Research-Lab/PathwayQA.

bioinformatics↗

bioRxiv2025

The E2F4 transcriptional repressor is a key mechanistic regulator of colon cancer resistance to irinotecan (CPT-11).

Background. Colorectal carcinomas (CRCs) are seldom eradicated by cytotoxic chemotherapy. Cancer cells with stem-like functional properties, often referred to as "cancer stem cells" (CSCs), display preferential resistance to several anti-tumor agents used in cancer chemotherapy, but the molecular mechanisms underpinning their selective survival remain only partially understood. Methods. In this study, we used Transcription Factor Target Genes (TFTG) enrichment analysis to identify transcriptional regulators (activators or repressors) that undergo preferential activation by chemotherapy in CRC cells with a "bottom-of-the-crypt" phenotype (EPCAM+/CD44+/CD166+; CSC-enriched) as compared to CRC cells with a "top-of-the-crypt" phenotype (EPCAM+/CD44neg/CD166neg; CSC-depleted). The two cell populations were purified in parallel by fluorescence-activated cell sorting (FACS) from a patient-derived xenograft (PDX) line representative of a moderately differentiated human CRC, following in vivo chemotherapy with irinotecan (CPT-11). The transcriptional regulators identified as differentially activated were tested for differential expression in normal vs. cancer tissues, and in cell populations enriched in stem/progenitor cell-types as compared to differentiated lineages (goblet cells, enterocytes) in the mouse colon epithelium. Finally, the top candidate was tested for mechanistic contribution to drug-resistance by selective down-regulation using short-hairpin RNAs (shRNAs). Results. Our analysis identified E2F4 and TFDP1, two core components of the DREAM transcriptional repression complex, as transcriptional modulators preferentially activated by irinotecan in EPCAM+/CD44+/CD166+ as compared to EPCAM+/CD44neg/CD166neg cancer cells. The expression levels of both genes (E2F4, TFDP1) were found up-regulated in CRCs as compared to human normal colon tissues, and in a sub-population of mouse colon epithelial cells enriched in stem/progenitor elements (Epcam+/Cd44+/Cd66alow/Kitneg) as compared to other sub-populations enriched in either goblet cells (Epcam+/Cd44+/Cd66alow/Kit+) or enterocytes (Epcam+/Cd44neg/Cd66ahigh). Most importantly, E2F4 down-regulation using shRNAs dramatically enhanced the sensitivity of human CRCs to in vivo treatment with irinotecan, across three independent PDX models. Conclusions. Our data identified E2F4 and the DREAM repressor complex as critical regulators of human CRC resistance to irinotecan, and as candidate targets for the development of chemo-sensitizing agents.

cancer biology↗

bioRxiv2024

Paying Attention to Attention: High Attention Sites as Indicators of Protein Family and Function in Language Models

Protein Language Models (PLMs) use transformer architectures to capture patterns within protein sequences, providing a powerful computational representation of the protein sequence [1]. Through large-scale training on protein sequence data, PLMs generate vector representations that encapsulate the biochemical and structural properties of proteins [2]. At the core of PLMs is the attention mechanism, which facilitates the capture of long-range dependencies by computing pairwise importance scores across residues, thereby highlighting regions of biological interaction within the sequence [3]. The attention matrices offer an untapped opportunity to uncover specific biological properties of proteins, particularly their functions. In this work, we introduce a novel approach, using the Evolutionary Scale Model (ESM) [4], for identifying High Attention (HA) sites within protein sequences, corresponding to key residues that define protein families. By examining attention patterns across multiple layers, we pinpoint residues that contribute most to family classification and function prediction. Our contributions are as follows: (1) we propose a method for identifying HA sites at critical residues from the middle layers of the PLM; (2) we demonstrate that these HA sites provide interpretable links to biological functions; and (3) we show that HA sites improve active site predictions for functions of unannotated proteins. We make available the HA sites for the human proteome. This work offers a broadly applicable approach to protein classification and functional annotation and provides a biological interpretation of the PLMs representation. 1 Author SummaryUnderstanding how proteins work is critical to advancements in biology and medicine, and protein language models (PLMs) facilitate studying protein sequences at scale. These models identify patterns within protein sequences by focusing on key regions of the sequence that are important to distinguish the protein. Our work focuses on the Evolutionary Scale Model (ESM), a state-of-the-art PLM, and we analyze the models internal attention mechanism to identify the significant residues. We developed a new method to identify "High Attention (HA)" sites--specific parts of a protein sequence that are essential for classifying proteins into families and predicting their functions. By analyzing how the model prioritizes certain regions of protein sequences, we discovered that these HA sites often correspond to residues critical for biological activity, such as active sites where chemical reactions occur. Our approach helps interpret how PLMs understand protein data and enhances predictions for proteins whose functions are still unknown. As part of this work, we provide HA-site information for the entire human proteome, offering researchers a resource to further study the potential functional relevance of these residues.

bioinformatics↗

bioRxiv2024

Pool PaRTI: A PageRank-based Pooling Method for Robust Protein Sequence Representation in Deep Learning

MotivationProtein language models produce token-level embeddings for each residue, resulting in an output matrix with dimensions that vary based on sequence length. However, downstream machine learning models typically require fixed-length input vectors, necessitating a pooling method to compress the output matrix into a single vector representation of the entire protein. Traditional pooling methods often result in substantial information loss, impacting downstream task performance. We aim to develop a pooling method that produces more expressive general-purpose protein embedding vectors while offering biological interpretability. ResultsWe introduce Pool PaRTI, a novel pooling method that leverages internal transformer attention matrices and PageRank to assign token importance weights. Our unsupervised and parameter-free approach consistently prioritizes residues experimentally annotated as critical for function, assigning them higher importance scores. Across four diverse protein machine learning tasks, Pool PaRTI enables significant performance gains in predictive performance. Additionally, it enhances interpretability by identifying biologically relevant regions without relying on explicit structural data or annotated training. To assess generalizability, we evaluated Pool PaRTI with two encoder-only protein language models, confirming its robustness across different models. Availability and ImplementationPool PaRTI is implemented in Python with PyTorch and is available at https://github.com/Helix-Research-Lab/Pool_PaRTI.git. The Pool PaRTI sequence embeddings and residue importance values for all human proteins on UniProt are available at https://zenodo.org/records/15036725 for ESM2 and protBERT.

bioinformatics↗

bioRxiv2024

Network-based representation learning enables the identification of risk genes and latent functional pathways in primary open-angle glaucoma

Despite the identification of hundreds of risk genes for primary open-angle glaucoma (POAG), a significant portion of the POAG genetic risk landscape remains unexplored. We hypothesized that unsupervised learning on large protein-protein interaction (PPI) networks could enable comprehensive characterization of the genetic pathways that underlie POAG risk. We used graph representation learning on a proteome-scale PPI network to generate embeddings capturing complex features of each proteins interactions. Using these embeddings, we trained a model with POAG-associated genes from the DisGeNET database to output an inferred POAG risk score for over 12,000 gene products, which identified known POAG risk genes with an area under the receiver operating characteristic curve of 0.739 (95% CI 0.686-0.792). These included well-known POAG risk genes such as RHOA and MMP3, as well as genes with significant contributions to other ocular diseases. Pathway analysis on the proteome-wide risk scores implicated 20 biological processes in POAG pathogenesis. Furthermore, cluster analysis of embeddings for POAG risk genes revealed 5 distinct functional neighborhoods, including cytokine signaling, coagulation response, collagen biosynthesis, extracellular matrix development, and fatty acid metabolism. Our results suggest that representation learning can recognize important patterns of protein interaction that allow in silico prioritization of POAG risk genes and pathways.

bioinformatics↗

bioRxiv2024

CRISPR-GPT: An LLM Agent for Automated Design of Gene-Editing Experiments

Performing effective gene-editing experiments requires a deep understanding of both the CRISPR technology and the biological system involved. Meanwhile, despite their versatility and promise, Large Language Models (LLMs) often lack domain-specific knowledge and struggle to accurately solve biological design problems. We present CRISPR-GPT, an LLM agent system to automate and enhance CRISPR-based gene-editing design and data analysis. CRISPR-GPT leverages the reasoning capabilities of LLMs for complex task decomposition, decision-making, and interactive human-artificial intelligence (AI) collaboration. This system incorporates domain expertise, retrieval techniques, external tools, and a specialized LLM fine-tuned with open-forum discussions among scientists. CRISPR-GPT assists users in selecting CRISPR systems, experiment planning, designing gRNAs, choosing delivery methods, drafting protocols, designing assays, and analyzing data. We showcase the potential of CRISPR-GPT by knocking-out four genes with CRISPR-Cas12a in a human lung adenocarcinoma cell line and epigenetically activating two genes using CRISPR-dCas9 in human melanoma cell line. CRISPR-GPT enables fully AI-guided gene-editing experiment design and analysis across different modalities, validating its effectiveness as an AI co-pilot in genome engineering.

bioinformatics↗

bioRxiv2024

β-amino acids reduce ternary complex stability and alter the translation elongation mechanism

Templated synthesis of proteins containing non-natural amino acids (nnAAs) promises to vastly expand the chemical space available to biological therapeutics and materials. Existing technologies limit the identity and number of nnAAs than can be incorporated into a given protein. Addressing these bottlenecks requires deeper understanding of the mechanism of messenger RNA (mRNA) templated protein synthesis and how this mechanism is perturbed by nnAAs. Here we examine the impact of both monomer backbone and side chain on formation and ribosome-utilization of the central protein synthesis substate: the ternary complex of native, aminoacylated transfer RNA (aa-tRNA), thermally unstable elongation factor (EF-Tu), and GTP. By performing ensemble and single-molecule fluorescence resonance energy transfer (FRET) measurements, we reveal the dramatic effect of monomer backbone on ternary complex formation and protein synthesis. Both the (R) and (S)-{beta}2 isomers of Phe disrupt ternary complex formation to levels below in vitro detection limits, while (R)- and (S)-{beta}3-Phe reduce ternary complex stability by approximately one order of magnitude. Consistent with these findings, (R)- and (S)-{beta}2-Phe-charged tRNAs were not utilized by the ribosome, while (R)- and (S)-{beta}3-Phe stereoisomers were utilized inefficiently. The reduced affinities of both species for EF-Tu ostensibly bypassed the proofreading stage of mRNA decoding. (R)-{beta}3-Phe but not (S)-{beta}3-Phe also exhibited order of magnitude defects in the rate of substrate translocation after mRNA decoding, in line with defects in peptide bond formation that have been observed for D--Phe. We conclude from these findings that non-natural amino acids can negatively impact the translation mechanism on multiple fronts and that the bottlenecks for improvement must include consideration of the efficiency and stability of ternary complex formation.

biochemistry↗

bioRxiv2023

Unsupervised learning reveals landscape of local structural motifs across protein classes

Proteins are known to share similarities in local regions of 3D structure even across disparate global folds. Such correspondences can help to shed light on functional relationships between proteins and identify conserved local structural features that lead to function. Self-supervised deep learning on large protein structure datasets has produced high-fidelity representations of local structural microenvironments, enabling comparison of local structure and function at scale. In this work, we leverage these representations to cluster over 15 million environments in the Protein Data Bank, resulting in the creation of a "lexicon" of local 3D motifs which form the building blocks of all known protein structures. We characterize these motifs and demonstrate that they provide valuable information for modeling structure and function at all scales of protein analysis, from full protein chains to binding pockets to individual amino acids. We devise a new protein representation based solely on its constituent local motifs and show that this representation enables state-of-the-art performance on protein structure search and model quality assessment. We then show that this approach enables accurate prediction of drug off-target interactions by modeling the similarity between local binding pockets. Finally, we identify structural motifs associated with pathogenic variants in the human proteome by leveraging the predicted structures in the AlphaFold structure database.

bioinformatics↗

bioRxiv2023

Explainable protein function annotation using local structure embeddings

The rapid expansion of protein sequence and structure databases has resulted in a significant number of proteins with ambiguous or unknown function. While advances in machine learning techniques hold great potential to fill this annotation gap, current methods for function prediction are unable to associate global function reliably to the specific residues responsible for that function. We address this issue by introducing PARSE (Protein Annotation by Residue-Specific Enrichment), a knowledge-based method which combines pre-trained embeddings of local structural environments with traditional statistical techniques to simultaneously predict function and provide residue-level annotations. For the task of predicting the catalytic function of enzymes, PARSE achieves comparable or superior global performance to state-of-the-art machine learning methods (F1 score > 85%) while simultaneously annotating the specific residues involved in each function with much greater precision. Since it does not require supervised training, our method can make one-shot predictions for very rare functions and is not limited to a particular type of functional label (e.g. Enzyme Commission numbers or Gene Ontology codes). Finally, we leverage the AlphaFold Structure Database to perform functional annotation at a proteome scale. By applying PARSE to the dark proteome--predicted structures which cannot be classified into known structural families--we predict several novel bacterial metalloproteases. Each of these proteins shares a strongly conserved catalytic site despite highly divergent sequences and global folds, illustrating the value of local structure representations for new function discovery.

bioinformatics↗

bioRxiv2022

Single-cell analysis of non-alcoholic fatty livers identifies a role for the constitutive androstane receptor

Non-alcoholic fatty liver disease is a heterogeneous disease with unclear underlying molecular mechanisms. While several genetic risk factors have been identified, the cellular and molecular heterogeneity associated with the development of hepatic steatosis are still not fully understood. Here, we perform single-cell RNA sequencing of hepatocytes and hepatic nonparenchymal cells to map the lipid signatures in mice with non-alcoholic fatty liver disease (NAFLD). We uncover previously unidentified clusters of hepatocytes characterized by either high or low srebp1 expression with unique molecular signatures of lipid synthesis. We find that NAFLD livers have elevated expression of the constitutive androstane receptor (CAR), a gene previously associated with lipid synthesis. Furthermore, nuclear expression of CAR positively correlates with steatohepatitis in humans. These findings demonstrate significant cellular differences in lipid signatures and identify a gene with a high likelihood of being linked to hepatic steatosis in humans.

cell biology↗

bioRxiv2022

COLLAPSE: A representation learning framework for identification and characterization of protein structural sites

The identification and characterization of the structural sites which contribute to protein function are crucial for understanding biological mechanisms, evaluating disease risk, and developing targeted therapies. However, the quantity of known protein structures is rapidly outpacing our ability to functionally annotate them. Existing methods for function prediction either do not operate on local sites, suffer from high false positive or false negative rates, or require large site-specific training datasets, necessitating the development of new computational methods for annotating functional sites at scale. We present COLLAPSE (Compressed Latents Learned from Aligned Protein Structural Environments), a framework for learning deep representations of protein sites. COLLAPSE operates directly on the 3D positions of atoms surrounding a site and uses evolutionary relationships between homologous proteins as a self-supervision signal, enabling learned embeddings to implicitly capture structure-function relationships within each site. Our representations generalize across disparate tasks in a transfer learning context, achieving state-of-the-art performance on standardized benchmarks (protein-protein interactions and mutation stability) and on the prediction of functional sites from the PO_SCPLOWROSITEC_SCPLOW database. We use COLLAPSE to search for similar sites across large protein datasets and to annotate proteins based on a database of known functional sites. These methods demonstrate that COLLAPSE is computationally efficient, tunable, and interpretable, providing a general-purpose platform for computational protein analysis.

bioinformatics↗

bioRxiv2021

Training data composition affects performance of protein structure analysis algorithms

The three-dimensional structures of proteins are crucial for understanding their molecular mechanisms and interactions. Machine learning algorithms that are able to learn accurate representations of protein structures are therefore poised to play a key role in protein engineering and drug development. The accuracy of such models in deployment is directly influenced by training data quality. The use of different experimental methods for protein structure determination may introduce bias into the training data. In this work, we evaluate the magnitude of this effect across three distinct tasks: estimation of model accuracy, protein sequence design, and catalytic residue prediction. Most protein structures are derived from X-ray crystallography, nuclear magnetic resonance (NMR), or cryo-electron microscopy (cryo-EM); we trained each model on datasets consisting of either all three structure types or of only X-ray data. We find that across these tasks, models consistently perform worse on test sets derived from NMR and cryo-EM than they do on test sets of structures derived from X-ray crystallography, but that the difference can be mitigated when NMR and cryo-EM structures are included in the training set. Importantly, we show that including all three types of structures in the training set does not degrade test performance on X-ray structures, and in some cases even increases it. Finally, we examine the relationship between model performance and the biophysical properties of each method, and recommend that the biochemistry of the task of interest should be considered when composing training sets.

bioinformatics↗