Biology subjects

Gray, J. J.

Publications and source records attributed to Gray, J. J..

At least 19 recordsLinked to original sources

bioRxiv2026

ProtAff: Protein Binding Affinity Prediction via LoRA-Finetuned ESM-2

Predicting the binding affinity of protein-protein interactions remains a central challenge in computational biology. Structure prediction models such as AlphaFold3 (AF3) and Boltz-2 can produce high-quality docking poses, and their confidence scores indicate structure quality, but these same scores fail to rank binding affinity among confirmed binders. Here we present ProtAff, a sequence-only affinity prediction model built on ESM-2 (650M parameters) with low-rank adaptation (LoRA) fine-tuning and a cross-attention module. ProtAff is trained using a margin ranking loss on 362,567 affinity measurements spanning 20 heterogeneous data sources, and we removed all training samples whose target sequence exceeds 50% similarity to the test target EGFR. On the AdaptyvBio EGFR benchmark (N =55), ProtAff achieves a Spearman correlation coefficient{rho} = 0.413, outperforming the best AF3 metric ({rho} = 0.054), the best Boltz-2 metric ({rho} = -0.046), and ML-based predictors MINT ({rho} = 0.242) and CrossAffinity ({rho} = 0.216). Applied to the AdaptyvBio Nipah virus binder design competition, a pipeline incorporating ProtAff for affinity ranking produced a design with KD = 0.132 nM (2 of 5 designs confirmed binding), a 2.8-fold improvement over the competition winner. On a cross-target discrimination benchmark of 91 VHH-antigen crystal structures, ProtAff underperforms structural methods for distinguishing cognate from non-cognate pairings, indicating that sequence-based affinity models are effective for within-target ranking but not for cross-target specificity.

bioinformatics↗

bioRxiv2026

GermRL: Alleviating The Germline Bias In Autoregressive Antibody Language Models Through Reinforcement Learning

Antibodies are powerful therapeutics whose antigen specificity arises from sequence diversity shaped during development. Recently, language models trained on large antibody repertoire datasets have enabled the generation and screening of novel candidates, but these models retain a strong germline bias. As AI adoption increases in therapeutic workflows, it is crucial to develop models that harness the diversity of antibodies necessary for the discovery of mutations that encode desirable properties. Previous work explored the germline bias in masked antibody language models, yet the bias in generative autoregressive language models has not yet been addressed. Here, we present GermRL, a lightweight and modular reinforcement learning (RL) framework capable of alleviating the germline bias in pre-trained antibody autoregressive language models through group relative policy optimization (GRPO). GermRL achieves consistent one-shot generation of antibodies that satisfy specified mutation thresholds from germline while maintaining structural plausibility. Under the lowest and highest mutation thresholds tested (5 and 35 mutations from germline), GermRL scores 0.992 and 0.950 pass@1, respectively, compared to 0.398 and 0.034 for the pre-trained language model. Within GermRL, we introduce a key pair of modifications to GRPO that increase training efficiency by discouraging reward hacking under our antibody application. Furthermore, comparison of RL generated and natural antibody sequences reveals how RL based optimization can explore alternative evolutionary mutational patterns and residue compositional strategies while preserving key global properties of natural antibodies, including identifiable germline assignments, embedding-level similarity and comparable developability profiles. Thus, RL-trained generative models optimized to promote antibody mutations through diversity from germline provide a promising framework for navigating the antibody sequence landscape, enabling exploration of novel yet biologically plausible candidates for therapeutic design.

bioinformatics↗

bioRxiv2026

Predicting Supramolecular Self-Assembly of Peptide Structures with AlphaFold3

Self-assembled peptide-based nanostructures have diverse applications in the pharmaceutical and materials fields, but accurately predicting their self-assembly behavior without time-intensive organic synthesis and characterization remains a significant challenge. Here, we assess the effectiveness of AlphaFold3 (AF3), a deep learning model for protein structure prediction, in modeling peptide-based nanostructures and the interactions driving supramolecular self-assembly. We designed amphiphilic peptides composed of alternating hydrophobic residues (valine, leucine, isoleucine, phenylalanine) and hydrophilic residues (glutamic acid), varying both sequence length and residue order. Using AF3's multimer mode, we modeled assemblies with copy numbers ranging from 10 to 1000, generating diverse morphologies such as micelles and nanotubes. We qualitatively analyzed hydrophobic regions, secondary structures, and intermolecular interactions, while also calculating radii of gyration, packing scores, and aspect ratios using PyRosetta. Our results indicate that AF3 predicts morphologies consistent with hydrophobic driving forces and steric constraints. Increased hydrophobicity correlates with smaller radii of gyration, while higher copy numbers correspond to smaller aspect ratios (more compact structures). Longer hydrophobic segments lead to disordered structures, whereas longer hydrophilic segments promote organization. While AF3 captures systemic trends consistent with biophysical principles, comparisons to literature reveal discrepancies driven by charge effects and secondary structure bias, including an overemphasis on helical propensity (e.g., alanine-rich sequences) and sensitivity to terminal charge repulsion. Additionally, since AF3 is predisposed to predict a single assembled entity rather than higher-order assemblies such as multiple micelles or fibers, finding the optimal copy number for the best prediction requires system-specific iteration. These limitations highlight the need for complementary approaches with controlled chemical potential and environmental conditions, though qualitative agreement with experimental trends in morphology and compactness supports AF3's utility for initial structure generation. Our findings highlight AF3's potential as a user-friendly first pass screening tool for peptide design, enabling rapid generation and prioritization of candidate assemblies for subsequent evaluation using coarse-grained or all-atom molecular simulations toward the efficient development of functional self-assembled peptide nanomaterials.

bioengineering↗

bioRxiv2025

Fitness Landscape for Antibodies 2: Benchmarking Reveals That Protein AI Models Cannot Yet Consistently Predict Developability Properties

1A prominent application of machine learning in therapeutic antibody design is the development of models that can generate or screen antibody candidates with a high probability of success in manufacturing and clinical trials. These models must accurately represent sequence-structure-function relationships, also known as the fitness landscape. Previous protein function benchmarks examine fitness landscapes across diverse protein families, but they exclude antibody data. Here, we introduce the second iteration of the Fitness Landscape for Antibodies (FLAb2), the largest public therapeutic antibody design benchmark to date. The datasets collected in FLAb2 contain developability assay data for over 4M antibodies across 32 studies, encompassing seven properties of therapeutic antibodies: thermostability, expression, aggregation, binding affinity, pharmacokinetics, polyreactivity, and immunogenicity. Using the curated data, we evaluate the performance of 30 artificial intelligence (AI) and biophysical models in learning these properties. Protein AI models on average do not produce statistically significant correlations for most (80%) of developability datasets. No models correlate with all properties or across multiple datasets of similar properties. Zero-shot predictions from pretrained models are incapable of accurately predicting all developability properties, although several models (IgLM, ProGen2, Chai-1, ESM2, ISM, IgFold) produce statistically significant correlations for multiple datasets for thermostability, expression, binding, or immunogenicity. Fine-tuning with at least 102 points improves performance on thermostability, aggregation, and binding, but polyreactivity and pharmacokinetics lack enough data for significance. Yet it is humbling to observe that given enough developability data (103 points), a fine-tuned one-hot encoding model can match the performance of fine-tuned billion-parameter pretrained models. Training data composition influences performance more than model architecture, and intrinsic biophysical properties (thermostability) are more readily learned than extrinsic properties (immunogenicity, pharmacokinetics). Controlling for germline distance with partial correlation reveals that protein language models draw substantially on evolutionary signal; on average, germline edit distance accounts for 40% of their apparent predictive power. FLAb2 data are accessible at https://github.com/Graylab/FLAb, together with scripts that allow researchers to benchmark, compare, and iteratively improve new AI-based developability prediction models.

bioinformatics↗

bioRxiv2025

Adapting Co-Folding Models for Structure-Based Protein-Protein Docking Through Flow Matching

Co-folding models like AlphaFold have revolutionized protein complex structure prediction, yet their reliance on multiple sequence alignments (MSAs) limits their applicability on challenging targets such as antibody-antigen complexes. An alternative approach, structure-based protein-protein docking, predicts the bound complex structure from the unbound monomer structures without requiring MSAs. In this work, we propose a novel method to adapt co-folding models for structure-based protein-protein docking by replacing their template module with a docking module, followed by training end-to-end with a flow-matching objective. We apply our method to AlphaFold-Multimer (AF-M) using the OpenFold implementation and transform it into a generative docking model, which we name AF2Dock. We evaluate AF2Dock and various baseline methods on the PINDER-AF2 benchmark and an antibody/nanobody test set. When using non-holo inputs, AF2Dock shows competitive performance compared to other structure-based docking methods and, in the case of nanobody complexes, outperforms all other docking methods tested here. Although AF2Dock underperforms co-folding AF-M and AF3 in success rates when using non-holo inputs, it produces orthogonal predictions and successfully identifies correct structures for targets where co-folding models fail. Ablation studies confirm that full-parameter fine-tuning of the AF-M components is critical for performance and reveal that, surprisingly, the inclusion of ESM embeddings can hinder success rates in certain cases such as nanobody complexes. The code is available at https://github.com/Graylab/AF2Dock.

biophysics↗

bioRxiv2025

Can We Extract Physics-like Energies from Generative Protein Diffusion Models?

Diffusion models have emerged as the state-of-the-art method in generative artificial intelligence (AI) and have shown great success in image synthesis, video generation, molecular design, and protein structure prediction. For biophysical problems, such as protein folding and association, a fundamental question in diffusion-based methods is how their learned functions correspond to thermodynamics. In this paper, we study diffusion models through the lens of theoretical biophysics, analyzing their underlying formulation of potentials and exploring their applications in scoring protein interactions. We develop simple theories rooted in statistical physics that relate thermodynamic potentials to the negative log of the probability of observing a system in a particular state. We include dimensional analysis of diffusion model equations and a table mapping AI and physics jargon. We then test a diffusion models ability to capture learned energies as negative log-likelihood values, -log p0(x0), by integrating over the diffusion-generated path or a probability flow path. We test these integrals on a simple 1D Gaussian mixture diffusion model and a protein-docking diffusion model, DFMDock. In the 1D case, we find that integration over both diffusion and flow paths can accurately recover ground truth probabilities. When we extract the learned docking energies for cases where DFMDock succeeds, we observe energy funnels with the minimum energy near the experimental docked structure, like those we observe with Rosetta, an empirically tuned physics-based biomolecular modeling suite. The learned energy performs comparably or outperforms Rosetta interface energy in 9 out of 25 cases at ranking the correctness of docked poses. These data show that we can extract a relevant learned energy function from a diffusion model and compare it to physical energy functions.

biophysics↗

bioRxiv2025

Evaluation of De Novo Deep Learning Models on the Protein-Sugar Interactome

Advances in deep learning have produced a range of models for predicting the protein-sugar interactome; however, structural docking of noncovalent protein-carbohydrate complexes remains largely unexplored. Although all-atom structure prediction models like AlphaFold3 (AF3), Boltz-1, Chai-1, DiffDock, and RosettaFold-All Atom (RFAA) were validated on protein-small molecule complexes, no benchmark or evaluation exists specifically for noncovalent protein-carbohydrate docking. To address this, we developed a high-quality dataset of experimental structures - Benchmark of CArbohydrate Protein Interactions (BCAPIN). Using BCAPIN and a novel evaluation metric, DockQC, we assessed the performance of all-atom structure prediction models on non-covalent protein-carbohydrate docking. We found all methods achieved comparable results, with an 85% success rate for structures of at least acceptable quality. However, we found that the predictive power of all models declined with increasing carbohydrate polymer length. With the capabilities and limitations assessed, we evaluated AF3s ability to predict binding for a set of putative human carbohydrate binding and carbohydrate non-binding proteins. While current models show promise, further development is needed to enable high-confidence, high-throughput prediction of the complete protein-sugar interactome.

biophysics↗

bioRxiv2025

Anti-citrullinated protein antibodies arise during affinity maturation of germline antibodies to carbamylated proteins in rheumatoid arthritis

Why autoantibodies in rheumatoid arthritis (RA) primarily target physiologically modified proteins, called citrullinated proteins, is unknown. Recognizing the inciting event in the production of anti-citrullinated protein antibodies (ACPAs) may shed light on the origin of RA. Here, we demonstrate that ACPAs originate from germline-encoded antibodies targeting a distinct but structurally similar modification, called carbamylation, which is pathogenic and environmentally driven. The transition from anti-carbamylated protein (anti-CarP) antibodies to ACPAs results from somatic hypermutations, indicating that the change in reactivity is acquired via antigen-driven affinity maturation. During this process, a single germline anti-CarP antibody transitions from anti-CarP to double positive (anti-CarP/ACPA) to ACPA according to the pattern and number of somatic hypermutations, explaining their coexistence and diverse specificity in RA. Artificial intelligence-based structural modeling revealed that an ACPA and its germline precursor exhibit distinct structural and biophysical properties, and pointed to heavy-chain tryptophan 48 (H-W48) as a critical residue in the differential recognition of citrullinated vs. carbamylated proteins. Indeed, a single methionine substitution in H-W48 changes the antibody specificity from ACPA to anti-CarP. These data indicate that the existence of germline-encoded anti-CarP antibodies is most likely the first event in the production of ACPAs during the early stages of RA development.

immunology↗

bioRxiv2025

The Proteome as a Lectome: Predictions from Deep Learning Propose Substantial Protein-Carbohydrate Interplay

It is a grand challenge to identify all the protein - carbohydrate interactions in an organism. Direct experiments would require extensive libraries of glycans to definitively distinguish binding from non-binding proteins. Computational screening of proteins for carbohydrate-binding provides an attractive and ultimately testable alternative. Recent computational techniques have focused primarily on which protein residues interact with carbohydrates or which carbohydrate species a protein binds to. Current estimates label 1.5 to 5% of proteins as carbohydrate-binding proteins; however, 50-70% of proteins are known to be glycosylated, suggesting a potential wealth of proteins that bind to carbohydrates. We therefore developed a novel dataset and neural network architecture, named Protein interaction of Carbohydrates Predictor (PiCAP), to predict whether a protein non-covalently binds to a carbohydrate. We trained PiCAP on a dataset of known carbohydrate binders, and we selected proteins that we identified as likely not to bind carbohydrates, including DNA-binding transcription factors, cytoskeletal components, selected antibodies, and selected small-molecule-binding proteins. PiCAP achieves a 90% balanced accuracy on protein-level predictions of carbohydrate binding/non-binding. Using the same dataset, we developed a model named Carbohydrate Protein Site Identifier 2 (CAPSIF2) to predict protein residues that interact non-covalently with carbohydrates. CAPSIF2 achieves a Dice coefficient of 0.57 on residue-level predictions on our independent test dataset, outcompeting all previous models for this task. To demonstrate the biological applicability of PiCAP and CAPSIF2, we investigated cell surface proteins of human neural cells and further predicted the likelihood of three proteomes, notably E. coli, M. musculus, and H. sapiens, to bind to carbohydrates. PiCAP predicts that approximately 35-40% of proteins in these proteomes bind carbohydrates, indicating a substantial interplay of protein-carbohydrate interactions for cellular functionality. Significance StatementThe totality of carbohydrate-protein interactions remains elusive, in part due to the inability to test proteomes versus glycomes in a high throughput manner. Here we show the first high-throughput methodology to predict protein-carbohydrate interactions at proteomic scales by using structural and sequence information. This information will allow scientists to target predicted protein-carbohydrate interactions to better determine how the elusive carbohydrate biomolecules play roles in all cellular functions.

bioinformatics↗

bioRxiv2024

Unified Sampling and Ranking for Protein Docking with DFMDock

Recent diffusion-based approaches to protein-protein docking typically decouple structure generation from decoy ranking. We introduce DFMDock (Denoising Force Matching for Docking), a unified diffusion model that integrates generative sampling and energy-based ranking through physically motivated supervision. DFMDock predicts both denoising forces and a scalar energy, trained using force matching and energy contrastive objectives. The predicted forces guide the reverse diffusion process, while the energy enables decoy ranking without relying on a separately trained confidence model. On the Docking Benchmark 5, DFMDock achieves a 32.8% Oracle success rate and 5.3% Top-1 success rate, outperforming DiffDock-PP (16.2% and 4.3%, respectively). Unlike co-folding models, DFM-Dock does not require MSAs and generalizes to unseen targets. In decoy ranking, its learned energy function outperforms Rosetta energy and model-derived confidence scores, producing funnel-shaped energy landscapes enriched for near-native structures. These results suggest DFMDock as an efficient and physically grounded approach to diffusion-based protein docking.

bioinformatics↗

bioRxiv2024

What has AlphaFold3 learned about antibody and nanobody docking, and what remains unsolved?

Antibody therapeutic development is a major focus in healthcare. To accelerate drug development, significant efforts have been directed towards the in silico design and screening of antibodies for which high modeling accuracy is necessary. To probe AlphaFold3s (AF3) capabilities and limitations, we tested AF3s ability to capture the fine details and interplay between antibody structure prediction and antigen docking accuracy. With one seed, AF3 achieves an 11.0% and 11.4% high-accuracy docking success rate for antibodies and nanobodies, respectively, and a median unbound CDR H3 RMSD accuracy of 2.73 [A] and 2.30 [A]. CDR H3 accuracy boosts complex prediction accuracy, with antigen context improving CDR H3 accuracy, particularly for loops longer than 15 residues. Combining I-pLDDT with {Delta}GB improves discriminative power for correctly docked complexes. However, AF3s 60% failure rate for antibody and nanobody docking (with single seed sampling) demonstrates necessary refinement to improve antibody design endeavors.

bioengineering↗

bioRxiv2024

Discovery of a single-subunit oligosaccharyltransferase that enables glycosylation of full-length IgG antibodies in Escherichia coli

Human immunoglobulin G (IgG) antibodies are one of the most important classes of biotherapeutic agents and undergo glycosylation at the conserved N297 site in the CH2 domain, which is critical for IgG Fc effector functions and anti-inflammatory activity. Hence, technologies for producing authentically glycosylated IgGs are in high demand. While attempts to engineer Escherichia coli for this purpose have been described, they have met limited success due in part to the lack of available oligosaccharyltransferase (OST) enzymes that can install N-linked glycans within the QYNST sequon of the IgG CH2 domain. Here, we identified a previously uncharacterized single-subunit OST (ssOST) from the bacterium Desulfovibrio marinus that exhibited greatly relaxed substrate specificity and, as a result, was able to catalyze glycosylation of native CH2 domains in the context of both a hinge-Fc fragment and a full-length IgG. Although the attached glycans were bacterial in origin, conversion to a homogeneous, asialo complex-type G2 N-glycan at the QYNST sequon of the E. coli-derived hinge-Fc was achieved via chemoenzymatic glycan remodeling. Importantly, the resulting G2-hinge-Fc exhibited strong binding to human Fc{gamma}RIIIa (CD16a), one of the most potent receptors for eliciting antibody-dependent cellular cytotoxicity (ADCC). Taken together, the discovery of a unique ssOST from D. marinus provides previously unavailable biocatalytic capabilities to the bacterial glycoprotein engineering toolbox and opens the door to using E. coli for the production and glycoengineering of human IgGs and fragments derived thereof.

synthetic biology↗

bioRxiv2024

Advancing membrane-associated protein docking with improved sampling and scoring in Rosetta

The oligomerization of protein macromolecules on cell membranes plays a fundamental role in regulating cellular function. From modulating signal transduction to directing immune response, membrane proteins (MPs) play a crucial role in biological processes and are often the target of many pharmaceutical drugs. Despite their biological relevance, the challenges in experimental determination have hampered the structural availability of membrane proteins and their complexes. Computational docking provides a promising alternative to model membrane protein complex structures. Here, we present Rosetta-MPDock, a flexible transmembrane (TM) protein docking protocol that captures binding-induced conformational changes. Rosetta-MPDock samples large conformational ensembles of flexible monomers and docks them within an implicit membrane environment. We benchmarked this method on 29 TM-protein complexes of variable backbone flexibility. These complexes are classified based on the root-mean-square deviation between the unbound and bound states (RMSDUB) as: rigid (RMSDUB <1.2 [A]), moderately-flexible (RMSDUB [isin] [1.2, 2.2) [A]), and flexible targets (RMSDUB > 2.2 [A]). In a local docking scenario, i.e. with membrane protein partners starting {approx}10 [A] apart embedded in the membrane in their unbound conformations, Rosetta-MPDock successfully predicts the correct interface (success defined as achieving 3 near-native structures in the 5 top-ranked models) for 67% moderately flexible targets and 60% of the highly flexible targets, a substantial improvement from the existing membrane protein docking methods. Further, by integrating AlphaFold2-multimer for structure determination and using Rosetta-MPDock for docking and refinement, we demonstrate improved success rates over the benchmark targets from 64% to 73%. Rosetta-MPDock advances the capabilities for membrane protein complex structure prediction and modeling to tackle key biological questions and elucidate functional mechanisms in the membrane environment. The benchmark set and the code is available for public use at github.com/Graylab/MPDock.

biophysics↗

bioRxiv2024

Molecular Display of the Animal Meta-Venome for Discovery of Novel Therapeutic Peptides

Animal venoms, distinguished by their unique structural features and potent bioactivities, represent a vast and relatively untapped reservoir of therapeutic molecules. However, limitations associated with extracting or expressing large numbers of individual venoms and venom-like molecules have precluded their therapeutic evaluation via high throughput screening. Here, we developed an innovative computational approach to design a highly diverse library of animal venoms and "metavenoms". We employed programmable M13 hyperphage display to preserve critical disulfide-bonded structures for highly parallelized single-round biopanning with quantitation via high-throughput DNA sequencing. Our approach led to the discovery of Kunitz type domain containing proteins that target the human itch receptor Mas-related G protein-coupled receptor X4 (MRGPRX4), which plays a crucial role in itch perception. Deep learning-based structural homology mining identified two endogenous human homologs, tissue factor pathway inhibitor (TFPI) and serine peptidase inhibitor, Kunitz type 2 (SPINT2), which exhibit agonist-dependent potentiation of MRGPRX4. Highly multiplexed screening of animal venoms and metavenoms is therefore a promising approach to uncover new drug candidates.

molecular biology↗

bioRxiv2024

FLAb: Benchmarking deep learning methods for antibody fitness prediction

The successful application of machine learning in therapeutic antibody design relies heavily on the ability of models to accurately represent the sequence-structure-function landscape, also known as the fitness landscape. Previous protein bench-marks (including The Critical Assessment of Function Annotation [33], Tasks Assessing Protein Embeddings [23], and FLIP [6]) examine fitness and mutational landscapes across many protein families, but they either exclude antibody data or use very little of it. In light of this, we present the Fitness Landscape for Antibodies (FLAb), the largest therapeutic antibody design benchmark to date. FLAb currently encompasses six properties of therapeutic antibodies: (1) expression, (2) thermosta-bility, (3) immunogenicity, (4) aggregation, (5) polyreactivity, and (6) binding affinity. We use FLAb to assess the performance of various widely adopted, pretrained, deep learning models for proteins (IgLM [28], AntiBERTy [26], ProtGPT2 [11], ProGen2 [21], ProteinMPNN [7], and ESM-IF [13]); and compare them to physics-based Rosetta [1]. Overall, no models are able to correlate with all properties or across multiple datasets of similar properties, indicating that more work is needed in prediction of antibody fitness. Additionally, we elucidate how wild type origin, deep learning architecture, training data composition, parameter size, and evolutionary signal affect performance, and we identify which fitness landscapes are more readily captured by each protein model. To promote an expansion on therapeutic antibody design benchmarking, all FLAb data are freely accessible and open for additional contribution at https://github.com/Graylab/FLAb.

bioinformatics↗

bioRxiv2023

Enhancement of antibody thermostability and affinity by computational design in the absence of antigen

Over the last two decades, therapeutic antibodies have emerged as a rapidly expanding domain within the field biologics. In silico tools that can streamline the process of antibody discovery and optimization are critical to support a pipeline that is growing more numerous and complex every year. In this study, DeepAb, a deep learning model for predicting antibody Fv structure directly from sequence, was used to design 200 potentially stabilized variants of an anti-hen egg lysozyme (HEL) antibody. We sought to determine whether DeepAb can enhance the stability of these antibody variants without relying on or predicting the antibody-antigen interface, and whether this stabilization could increase antibody affinity without impacting their developability profile. The 200 variants were produced through a robust highthroughput method and tested for thermal and colloidal stability (Tonset, Tm, Tagg), affinity (KD) relative to the parental antibody, and for developability parameters (non-specific binding, aggregation propensity, self-association). In the designed clones, 91% and 94% exhibited increased thermal and colloidal stability and affinity, respectively. Of these, 10% showed a significantly increased affinity for HEL (5-to 21-fold increase), with most clones retaining the favorable developability profile of the parental antibody. These data open the possibility of in silico antibody stabilization and affinity maturation without the need to predict the antibody-antigen interface, which is notoriously difficult in the absence of crystal structures.

molecular biology↗

bioRxiv2023

Reliable protein-protein docking with AlphaFold, Rosetta and replica-exchange

Despite the recent breakthrough of AlphaFold (AF) in the field of protein sequence-to-structure prediction, modeling protein interfaces and predicting protein complex structures remains challenging, especially when there is a significant conformational change in one or both binding partners. Prior studies have demonstrated that AF-multimer (AFm) can predict accurate protein complexes in only up to 43% of cases.1 In this work, we combine AlphaFold as a structural template generator with a physics-based replica exchange docking algorithm to better sample conformational changes. Using a curated collection of 254 available protein targets with both unbound and bound structures, we first demonstrate that AlphaFold confidence measures (pLDDT) can be repurposed for estimating protein flexibility and docking accuracy for multimers. We incorporate these metrics within our ReplicaDock 2.0 protocol2 to complete a robust in-silico pipeline for accurate protein complex structure prediction. AlphaRED (AlphaFold-initiated Replica Exchange Docking) successfully docks failed AF predictions including 97 failure cases in Docking Benchmark Set 5.5. AlphaRED generates CAPRI acceptable-quality or better predictions for 63% of benchmark targets. Further, on a subset of antigen-antibody targets, which is challenging for AFm (20% success rate), AlphaRED demonstrates a success rate of 43%. This new strategy demonstrates the success possible by integrating deep-learning based architectures trained on evolutionary information with physics-based enhanced sampling. The pipeline is available at github.com/Graylab/AlphaRED.

biophysics↗

bioRxiv2023

Contextual protein encodings from equivariant graph transformers

The optimal residue identity at each position in a protein is determined by its structural, evolutionary, and functional context. We seek to learn the representation space of the optimal amino-acid residue in different structural contexts in proteins. Inspired by masked language modeling (MLM), our training aims to transduce learning of amino-acid labels from non-masked residues to masked residues in their structural environments and from general (e.g., a residue in a protein) to specific contexts (e.g., a residue at the interface of a protein or antibody complex). Our results on native sequence recovery and forward folding with AlphaFold2 suggest that the amino acid label for a protein residue may be determined from its structural context alone (i.e., without knowledge of the sequence labels of surrounding residues). We further find that the sequence space sampled from our masked models recapitulate the evolutionary sequence neighborhood of the wildtype sequence. Remarkably, the sequences conditioned on highly plastic structures recapitulate the conformational flexibility encoded in the structures. Furthermore, maximum-likelihood interfaces designed with masked models recapitulate wildtype binding energies for a wide range of protein interfaces and binding strengths. We also propose and compare fine-tuning strategies to train models for designing CDR loops of antibodies in the structural context of the antibody-antigen interface by leveraging structural databases for proteins, antibodies (synthetic and experimental) and protein-protein complexes. We show that pretraining on more general contexts improves native sequence recovery for antibody CDR loops, especially for the hypervariable CDR H3, while fine-tuning helps to preserve patterns observed in special contexts.

bioinformatics↗