Biology subjects

Skolnick, J.

Publications and source records attributed to Skolnick, J..

15 recordsLinked to original sources

bioRxiv2026

Protein Pocket Coengagement Prioritizes Drug-Drug Interactions

Drug-drug interactions (DDIs) are largely catalogued empirically, with limited structural explanation of why two drugs interact. We asked whether a purely structural mechanism - competing or cooperating engagement of the same protein pocket - is sufficient to recover documented DDIs without any training on clinical interaction labels. Using composite protein multiple-ligand (COLIG) assignments, which classify co-bound ligand pairs as attractive or repulsive, we scored 21,945 pairs of 210 drugs by predicted engagement of each proteins top-ranked pocket. Confidence-summed COLIG ranking achieved top-5% precision of 49.3% against DDInter and 25.1% against KEGG, versus background prevalences of 29.6% and 14.7%, respectively; pooled top-pocket engagement reached 50.3% DDInter precision. These results support pocket-based DDI screening and class-dependent hypotheses about direct and noncontact ligand engagement, while leaving causal mechanisms to be established.

bioinformatics↗

bioRxiv2026

The human metabolite - protein interactome reveals a global layer of cellular coordination

Metabolites are substrates, products, cofactors, and regulators, but protein-protein interaction networks do not represent their potential to organize proteins across conventional pathway boundaries. Using the LIGMAP virtual-screening algorithm, we mapped 308 human metabolite codes to pockets in monomers, dimer interfaces, and non-interface sites in dimers and represented attractive or repulsive COLIG states involving pairs of metabolites in the same pocket. On a fixed cohort of 3,938 proteins, the mean coverage of 104 strict non-enzyme pathways was 44.6% for LIGMAP, 70.1% for STRING, and 82.5% for STRING+LIGMAP; the union placed 86.3% of eligible proteins in the largest connected component and 95.1% in the two largest components. STRING+LIGMAP protein coverage was 88.0% for 68 enzyme-only pathways and 88.9% for 1,283 mixed pathways. In pathway-held-out, degree-matched prediction, adding LIGMAP to degree plus STRING increased the mean area under the precision-recall curve from 0.651 to 0.660 (paired P = 0.024); adding BioLiP2 increased it to 0.663 (paired P = 0.005). Ancient-only and non-ancient-only subnetworks were each globally connected; ancient features were denser, whereas non-ancient features covered more proteins and pathways. At the full 5,426-protein scale, retaining only features assigned to 2-100 proteins recovered 698 of 1,691 strict non-enzyme reference edges (41.3%) and exceeded both protein-label and exact bipartite degree-preserving nulls. Uncapped recovery approached saturation and lost identity-selective enrichment. Experimentally established metabolite-dependent complexes validate the local mechanism independently of LIGMAP; LIGMAP fully recovered two of seven stringent direct mechanisms and all three broader serial axes examined. Our findings reveal a global metabolite-mediated architecture with the capacity to coordinate proteins across otherwise distinct cellular systems.

bioinformatics↗

bioRxiv2026

Metabolites form a globally connected chemical network across protein families

Metabolites are generally viewed as substrates, products, cofactors, or regulators of individual proteins, whereas metabolites recurring across many protein families are often regarded as promiscuous binders. Here, we analyzed 989,058 BioLiP2 protein-ligand binding sites and assigned 929,546 sites to ECOD v295 homologous groups to quantify ligand specificity, cross-fold scatter, structural breadth, and metabolite-mediated connectivity across protein-family space. Many ancient metabolites preferentially occupied cognate structural groups, demonstrating that broad evolutionary reuse can coexist with local structural discrimination. After excluding elemental metals, BioLiP potential-artifact/dual-use ligands, and metabolites containing fewer than six heavy atoms, 32 ancient metabolites occupied a mean of 185.38 ECOD F-groups per metabolite, compared with 6.32 F-groups for 2,540 mapped filtered non-ancient metabolites-- a 29.35-fold enrichment (bootstrap 95% CI, 18.66-43.46). The complete 40-ancient-metabolite network connected all 6,798 associated F-groups into a single giant connected component (GCC). Even after stringent filtering, all 3,135 ancient-metabolite-associated F-groups remained in one GCC. Degree-preserving configuration-model randomizations and maximum-degree capping showed that this connectivity follows from the broad, recurrent distribution of metabolite binding rather than dependence on a few extreme hubs or a specialized higher-order topology. Differences between ancient and filtered non-ancient networks were not explained by metabolite size, whereas generic crystallization additives preferentially occupied smaller pockets. These results indicate that a limited ancient chemical repertoire established a globally connected protein- family architecture that subsequent metabolite diversification expanded while preserving its basic organization. SignificanceMetabolites are conventionally viewed as substrates, products, cofactors, or regulators acting on individual proteins. Global examination of experimentally observed metabolite-protein interactions reveals a broader organizing principle. Ancient metabolites combine local binding discrimination with extraordinary reuse across protein families, such that only 40 metabolites generate an almost completely connected network spanning thousands of ECOD (evolutionary classification of domains) protein families. The much larger non-ancient metabolite repertoire expands the protein-family space covered by this network, while preserving near-global connectivity. Thus, metabolite diversification appears to have elaborated, rather than created, a chemically connected protein architecture established early in evolution, suggesting that overlapping metabolite-binding repertoires could coordinate proteins, pathways, and cellular processes.

bioinformatics↗

bioRxiv2026

Understanding Substance Dependence: What Differentiates Addictive from Non-Addictive Drugs?

Addiction is a global health challenge, yet the molecular features that distinguish addictive from non-addictive drugs remain incompletely understood at the pathway and circuit levels. Here, we present a systematic computational framework that integrates drug-target binding predictions (provided by FINDSITEcomb2.0) with brain-region-specific protein expression to compare addictive and non-addictive compounds. We analyzed 457 addictive and 1,774 non-addictive blood-brain barrier permeable drugs and mapped their predicted targets and associated pathways onto proteins expressed across 120 addiction-relevant brain regions. This analysis reveals widespread convergence between the two classes (addictive and non-addictive drugs) on shared molecular pathways, accompanied by distinct patterns of target and pathway engagement. Functional annotation of differentially engaged targets highlights biases toward plasticity-associated components for addictive drugs. In contrast, non-addictive drugs interact with both plasticity-associated proteins and proteins within the same molecular complex that have addiction suppression, regulatory, and homeostatic functions. Notably, both target classes co-localize within the same addiction-relevant circuits and form an integrated protein-protein interaction network. Together, these results define a differential engagement landscape that links chemical interactions to pathway-level utilization in the brain, revealing molecular features associated with differences in addiction propensity. Significance StatementAddiction is a global health crisis, yet the molecular features that distinguish addictive from non-addictive drugs remain poorly defined. By systematically comparing drug-target and pathway engagement across shared addiction-relevant brain circuits, this study identifies distinct signatures in which addictive compounds preferentially engage plasticity- and stress-associated targets, while non-addictive compounds engage these as well as addiction suppression, regulatory, transport, and metabolic pathways. These results provide a circuit- and pathway-level framework for interpreting addiction liability and guiding the design of safer therapeutics.

biophysics↗

bioRxiv2026

AlphaInterp: Probing AlphaFold 3's Internal Representations Reveals Evolutionary Determinants of Predicted Structure and Confidence

AlphaFold 3 predicts biomolecular structures with unprecedented accuracy, yet the computations transforming sequence and evolutionary data into structural coordinates remain poorly understood. Here, we present a systematic mechanistic interpretability analysis of AlphaFold 3, tracking its internal representations across the forward pass. Probing four critical network checkpoints reveals that the Pairformer compresses diffuse co-evolutionary inputs into a compact latent geometry where complex biophysical features become linearly decodable. Using causal activation patching, we demonstrate that predicted confidence is directly manipulable within this latent space, allowing geometric certainty to be transferred across entirely unrelated proteins. Furthermore, across adversarial-mutation, fold-switching, and generalization benchmarks, we show that AlphaFold 3s representational coherence strictly requires comparative evolutionary context. The latent space collapses when multiple sequence alignments are removed, regardless of sequence familiarity or training-set membership. This stability requires phylogenetic diversity rather than alignment depth, and a minimal set of highly divergent homologs is sufficient to anchor the latent space and activate the models structural priors. These findings indicate that AlphaFold 3s representational coherence is deeply tied to evolutionary scaffolding, suggesting it functions similarly to an advanced fold-recognition system and highlighting that protein structure prediction from sequence alone is not yet fully solved.

bioinformatics↗

bioRxiv2026

PHENOCAUZ: Linking Human Symptoms, Drug Side Effects and Efficacy to Their Molecular Causes Using Mendelian Disease Biology

Human diseases and adverse drug reactions are ultimately recognized through clinical symptoms, yet the molecular determinants of most symptoms remain unknown. To address this key issue, we present PHENOCAUZ, a computational framework that links symptoms to their causative proteins by integrating Mendelian phenotype-gene relationships with molecular features of proteins. Starting from symptom annotations derived from Mendelian phenotypes and their causal genes, PHENOCAUZ identifies biological pathways and processes associated with individual symptoms and trains a machine learning model to predict symptom-causing proteins beyond those currently implicated in Mendelian diseases. The approach is motivated by the hypothesis that if dysfunction of a protein produces a symptom in a Mendelian disorder, the same protein may contribute to the same symptom in a complex disease or cause drug toxicity. Benchmarking across 2,344 symptoms and 4,828 Mendelian proteins using leave-one-out cross-validation yielded an estimated precision of approximately 0.70 among the top predictions. Predicted symptom-protein relationships show strong pathway-level agreement with literature-curated symptom-protein associations, efficacious drug targets and disease mode-of-action proteins. PHENOCAUZ also enables practical applications including prediction of severe drug side effects and identification of candidate therapeutics for ovarian, prostate, and breast cancers as well as noncancer diseases such as dementia and Crohns disease. These results demonstrate that Mendelian disease biology provides a powerful route to connect clinical symptoms with their molecular determinants and translate those insights into drug discovery and safety prediction. Significance StatementClinical symptoms define how diseases and drug side effects are recognized, yet the molecular determinants of most symptoms remain unknown. We introduce PHENOCAUZ, a framework that uses Mendelian disease gene-phenotype relationships to infer the proteins and pathways responsible for human symptoms, by linking symptoms to molecular mechanisms and enabling the prediction of adverse drug reactions and discovery of potential therapeutic targets for both Mendelian and complex diseases.

bioinformatics↗

bioRxiv2026

Adversarial Sequence Mutations in AlphaFold andESMFold Reveal Nonphysical StructuralInvariance, Confidence Failures, and Concerns forProtein Design

AlphaFold has transformed structural biology and spawned an ecosystem of derivative tools for protein design, binding prediction, and drug discovery. However, whether AlphaFold has learned generalizable biophysical principles versus template-based pattern matching remains unclear--a distinction critical for applications beyond its training context. Here, we perform a systematic adversarial evaluation of AlphaFold 3 using point and deletion mutations across 200 proteins. Remarkably, predicted structures remain invariant to mutations of up to 40% of residues--including deliberately destabilizing substitutions--and to deletions of 10%. Notably, this invariance holds even for experimentally validated fold-switching proteins that are known to adopt alternative conformations in response to such mutations, despite the fact that these proteins are small and monomeric--precisely the category where AlphaFold is expected to perform best. Confidence metrics prove unreliable, as they select the most accurate structure at most 35% of the time and correlate with the structural quality of the best available training set template. This suggests that AlphaFolds uncertainty estimates reflect template availability more than biophysical reasoning. ESMFold exhibits greater, though still imperfect, mutational sensitivity, suggesting superior sequence-structure coupling. These findings indicate that AlphaFold may rely heavily on memorized templates rather than biophysical reasoning, with profound implications for the reliability of AlphaFold-based protein design, drug discovery, and modeling workflows.

bioinformatics↗

bioRxiv2025

GlueFinder: A Data-Driven Framework for the Rational Discovery of Molecular Glues

Molecular glues can drive targeted protein degradation by stabilizing ternary complexes between proteins of interest and E3 ubiquitin ligases, but rational design has lagged due to limited rules for interface recognition and an overreliance on a few ligases (e.g., VHL or Cereblon). We introduce GlueFinder, a systematic, unbiased platform that leverages structural bioinformatics to mine the Protein Data Bank for ligand binding pockets adjacent to the protein interface which are ligandable sites near protein-protein interfaces that can nucleate glue-mediated complex formation. After validating its performance on a benchmark of experimentally solved dimeric structures with known and predicted glues, we applied GlueFinder to three therapeutically important targets, EGFR, HER2, and KRAS, and predicted candidate glues that recruit 24, 111, and 148 distinct E3 ligases to these targets, respectively. We further demonstrate that GlueFinder can promote the formation of non-native EGFR complexes, possibly enabling ternary assemblies that would not form on their own. Together, these results establish a general, computation-guided strategy for molecular glue discovery that decouples design from legacy degrader scaffolds and specific ligase dependencies, expands the usable E3 ligase repertoire, and enables rational targeting of interfacial binding pockets. GlueFinder thus broadens both the scope and precision of targeted protein degradation and moves the field toward mechanism-driven, systematic glue development across diverse therapeutic contexts. SignificanceMolecular glues are small molecules that destroy disease-causing proteins by helping them attach to cellular disposal machinery. However, current glue discovery depends on a few known ligases and lacks clear design rules. GlueFinder provides a general, computation-guided method that scans protein structures to find pockets near interfaces where glues can act. Applied to key targets such as EGFR, HER2, and KRAS, it predicts glues that connect many different ligases and even create new protein complexes. By revealing where and how glues can work, GlueFinder expands the range of degradable proteins and accelerates the rational design of future therapeutics.

bioinformatics↗

bioRxiv2025

Combining structural modeling and deep learning to calculate the E. coli protein interactome and functional networks

We report on the integration of three methods that are computationally efficient enough to predict, on a proteome-wide scale, whether two proteins are likely to form a binary complex. The methods include PrePPI, which uses three-dimensional structure information as a basis for predictions, Topsy-Turvy which analyzes sequences using a protein language model, and ZEPPI which uses evolutionary information to evaluate protein-protein interfaces. We demonstrate how these methods can be integrated and validate the performance of the integrated method and its separate components at predicting E. coli protein-protein interactions through testing on the HINT high-quality literature-curated database of binary interactions. The integrated method identifies more high-confidence (FPR [≤] 0.001) interactions ([~]20K) than any of the component methods. The AF3Complex algorithm was used to predict the structures of 400 protein-protein interactions, and 78% of the integrated method predictions resulted in models deemed accurate by the AF3Complex evaluation score. Notably, essentially all AF3Complex models have at least partially overlapping interfaces with PrePPI models of the complexes. Finally, we clustered the high-confidence E. coli interactome and obtained 385 subnetworks which have high functional coherence defined by enrichment of Gene Ontology Biological Process terms, thus, illustrating that our methods which contain no explicit functional information provide biologically meaningful protein interactions. Biological insights derived from the subnetworks, including the annotation of proteins of unknown function, are discussed in detail. Overall, independent validations support the accuracy of the comprehensive E. coli interactome we have presented.

systems biology↗

bioRxiv2025

MENDELSEEK: An algorithm that predicts Mendelian Genes and elucidates what makes them special

Although individual Mendelian diseases--those caused by a single gene--are rare, their collective disease burden is substantial. Identifying the causal gene for each condition is essential for accurate diagnosis and effective treatment. Yet, despite decades of research, the genetic basis of more than half of all known Mendelian diseases remains unresolved. To address this gap, we introduce MENDELSEEK, a machine learning framework that predicts Mendelian genes by integrating residue variation scores with pathway participation, Gene Ontology processes, and protein language model features. In benchmarking across 16,946 human genes with 10-fold cross-validation, MENDELSEEK achieved an AUC of 0.869 and an AUPR of 0.737--substantially outperforming the next best methods, ENTPRISE+ENTPRISE-X (AUC 0.781; AUPR 0.626), and REVEL (AUC 0.585; AUPR 0.401). When applied to the full set of 17,858 human genes, MENDELSEEK predicted 1,277 novel Mendelian gene candidates with precision greater than 0.7. Analysis further revealed that Mendelian genes engage in significantly more protein-protein interactions than non-Mendelian genes and are evolutionarily ancient. Together, these results highlight MENDELSEEK as a major advance over existing methods, offering new insights into the biochemical features that distinguish Mendelian from non-Mendelian genes.

genomics↗

bioRxiv2025

AF3Complex Yields Improved Structural Predictions of Protein Complexes

MotivationAccurate structures of protein complexes are essential for understanding biological pathway function. A previous study showed how downstream modifications to AlphaFold 2 could yield AF2Complex, a model better suited for protein complexes. Here, we introduce AF3Complex, a model equipped with both similar and novel improvements, built on AlphaFold 3. ResultsBenchmarking AF3Complex and AlphaFold 3 on a large dataset of protein complexes, it was shown that AF3Complex outperforms AlphaFold 3. Moreover, by evaluating the structures generated by AF3Complex on datasets of protein-peptide complexes and antibody-antigen complexes, it was established that AF3Complex could create high-fidelity structures for these challenging complex types. Additionally, when deployed to generate structural predictions for protein complexes used in the recent CASP16 competition, AF3Complex yielded structures that would have placed it among the top models in the competition. AvailabilityThe AF3Complex code is freely available at https://github.com/Jfeldman34/AF3Complex.git. ContactPlease contact jonathanfeldman@gatech.edu.

bioinformatics↗

bioRxiv2022

Deep learning-driven insights into super protein complexes for outer membrane protein biogenesis in bacteria

To reach their final destinations, outer membrane proteins (OMPs) of gram-negative bacteria undertake an eventful journey beginning in the cytosol. Multiple molecular machines, chaperones, proteases, and other enzymes facilitate the translocation and assembly of OMPs. These helpers usually associate, often transiently, forming large protein assemblies. They are not well-understood due to experimental challenges in capturing and characterizing protein-protein interactions (PPIs), especially transient ones. Using AF2Complex, we introduce a high-throughput, deep learning pipeline to identify PPIs within the E. coli cell envelope and apply it to several proteins from an OMP biogenesis pathway. Among the top confident hits obtained from screening ~1,500 envelope proteins, we find not only expected interactions but also unexpected ones with profound implications. Subsequently, we predict atomic structures for these protein complexes. These structures, typically of high confidence, explain experimental observations, and lead to mechanistic hypotheses for how a chaperone assists a nascent, precursor OMP emerging from a translocon, how another chaperone prevents it from aggregating and docks to an {beta}-barrel assembly port, and how a protease performs quality control. This work presents a general strategy for investigating biological pathways by using structural insights gained from deep learning-based predictions.

molecular biology↗

bioRxiv2022

Prediction of inter-chain distance maps of protein complexes with 2D attention-based deep neural networks

Residue-residue distance information is useful for predicting the tertiary structures of protein monomers or the quaternary structures of protein complexes. Many deep learning methods have been developed to predict intra-chain residue-residue distances of monomers accurately, but very few methods can accurately predict inter-chain residue-residue distances of protein complexes. We develop a new deep learning method CDPred (i.e., Complex Distance Prediction) based on the 2D attention-powered residual network architecture to address the gap. CDPred predicts the inter-chain distance maps of dimers (homodimers or heterodimers) from the features extracted from multiple sequence alignments (MSAs) and the intra-chain distance maps of predicted tertiary structures of monomers. Tested on two homodimer test datasets, CDPred achieves the precision of 61.56% and 43.26% for top L/5 inter-chain contact predictions (L: length of the monomer in homodimer), respectively, which is substantially higher than DeepHomos 37.40% and 23.08% and GLINTERs 48.09% and 36.74%. And tested on the two heterodimer test datasets, the top L/5 inter-chain contact prediction precision (L: length of the shorter monomer in heterodimer) of CDPred is 47.59% and 22.87% respectively, which surpasses GLINTERs 23.24% and 13.49%. Moreover, we demonstrate that the residue-residue co-evolutionary features calculated from multiple sequence alignments by a deep learning language model are more informative for the inter-chain contact prediction than the traditional statistical optimization approach of maximizing direct co-evolutionary signals, and large intra-chain distances in the intra-chain distance maps of monomers are more useful for the inter-chain distance prediction than small intra-chain distances.

bioinformatics↗

bioRxiv2021

Predicting direct physical interactions in multimeric proteins with deep learning

Accurate descriptions of protein-protein interactions are essential for understanding biological systems. Remarkably accurate atomic structures have been recently computed for individual proteins by AlphaFold2 (AF2). Here, we demonstrate that the same neural network models from AF2 developed for single protein sequences can be adapted to predict the structures of multimeric protein complexes without retraining. In contrast to common approaches, our method, AF2Complex, does not require paired multiple sequence alignments. It achieves higher accuracy than some complex protein-protein docking strategies and provides a significant improvement over AF-Multimer, a new development of AlphaFold for multimeric proteins. Moreover, we introduce metrics for predicting direct protein-protein interactions between arbitrary protein pairs and validate AF2Complex on some challenging benchmark sets and the E. coli proteome. Lastly, using the cytochrome c biogenesis system I as an example, we present high-confidence models of three sought-after assemblies formed by eight members of this system.

bioinformatics↗

bioRxiv2021

A general framework to learn tertiary structure for protein sequence annotation

During the past five years, deep-learning algorithms have enabled ground-breaking progress towards the prediction of tertiary structure from a protein sequence. Very recently, we developed SAdLSA, a new computational algorithm for protein sequence comparison via deep-learning of protein structural alignments. SAdLSA shows significant improvement over established sequence alignment methods. In this contribution, we show that SAdLSA provides a general machine-learning framework for structurally characterizing protein sequences. By aligning a protein sequence against itself, SAdLSA generates a fold distogram for the input sequence, including challenging cases whose structural folds were not present in the training set. About 70% of the predicted distograms are statistically significant. Although at present the accuracy of the intra-sequence distogram predicted by SAdLSA self-alignment is not as good as deep-learning algorithms specifically trained for distogram prediction, it is remarkable that the prediction of single protein structures is encoded by an algorithm that learns ensembles of pairwise structural comparisons, without being explicitly trained to recognize individual structural folds. As such, SAdLSA can not only predict protein folds for individual sequences, but also detects subtle, yet significant, structural relationships between multiple protein sequences using the same deep-learning neural network. The former reduces to a special case in this general framework for protein sequence annotation.

bioinformatics↗