Mohammed AlQuraishi

Cells employ dense networks of transient protein–protein interactions mediated by modular peptide-binding domains and unstructured peptidic motifs for high-fidelity information processing. How these networks physically execute computations through protein interactions governed by complex intra- and intermolecular mechanisms remains indiscernible from current, sparse and non-quantitative, maps of the human interactome. Here, we introduce a quantitative statistical mechanical modeling (QSM) approach for machine learning domain–peptide affinities with experimental-level accuracy. Leveraging a new, principled algorithm for data harmonization and a biophysically informed neural network architecture, QSM learns to predict dissociation constants directly from amino acid sequences with calibrated confidence. We use QSM to construct the first quantitative drafts of human signaling networks and study these networks across three physical scales—recognition mechanisms of modular binding domains, combinatorial logic of multi-dentate proteins, and pathways inferred from de novo inference of protein interaction networks. We find that (i) modular domains, based on their binding preferences, selectivities, and strengths, fall into a limited number of biophysical equivalence groups, (ii) those domains, along with peptidic motifs, are “syntactically” combined within proteins to yield multivalent recognition mechanisms, and (iii) the organization of cellular function can be traced back to algorithmically detectable modules induced by domain-mediated interactions. In aggregate, these analyses instantiate a tractable roadmap towards a comprehensive, mechanistic, and simulatable articulation of the systems biology of signaling.

High-quality experimental datasets that link protein–ligand structures with binding affinity data are essential for developing and evaluating structure-based machine learning methods. To help address this need, we established OpenBind as an open-science initiative to generate large-scale experimental datasets for structure-based AI and molecular discovery. Here, we describe the first public OpenBind release, which, to the best of our knowledge, is the largest public single-target experimental structure–affinity dataset. The dataset focuses on enteroviral 2A protease, comprising 925 crystallographic binding events from 699 compounds and associated affinity measurements for 601 compounds. It combines structures from an initial fragment screen and follow-on molecules, together with affinity data, linking experimentally determined protein–ligand binding modes to biophysical measurements within a coherent antiviral discovery campaign. We used this dataset to evaluate protein–ligand structure prediction, binding-affinity prediction, and virtual screening using representative structure-based methods, including docking and cofolding. This exposed several challenges that are central to practical structure-based modelling: docking performance depends strongly on binding-pocket conformation, poses are difficult to rank, and structure-based affinity prediction remains challenging. Fine-tuning OpenFold3-p2 on the fragment-screen structures substantially improved pose prediction and virtual screening for related follow-on compounds, demonstrating how early-stage experimental structures can support target-specific model adaptation.

The universe of possible protein sequences is astronomically large, yet our understanding of sequence-structure relationships is derived almost entirely from the infinitesimal fraction used by life. Determining which physical properties are already accessible to unevolved polypeptides, and which require subsequent evolutionary refinement, is central both to understanding protein emergence and to expanding protein design beyond biological sequence space. Here, we explore this largely uncharted space using a high-throughput FRET-based selection of a library containing millions of random protein sequences, complemented by fluorescence imaging, biophysical characterization and sequence-based modelling. We find that random sequence space contains multiple physical phenotypes that cannot be captured by a simple distinction between soluble and aggregated proteins. Individual sequences can produce substantial FRET in both the cytosol and visible assemblies, while assemblies differ in packing density and cellular responses. Repeated selection for stronger FRET reduces population solubility indicating that proximity-based selection alone does not necessarily favor biology-like protein states. These experimentally defined phenotypes are predictable from sequence. Our results reveal resolvable and learnable variation within unevolved protein sequence space, providing a route to investigate how far random sequences lie from soluble, stable protein solutions and to learn from regions of the protein universe not previously sampled by evolution.

Despite the breakneck pace of progress in protein design methodology, frontier problems remain challenging, with leading methods struggling to design high-affinity binders, scaffold multiple functional motifs, or stabilize large multi-domain proteins. Recent research efforts have focused on two areas: improving model reasoning when generating active sites or binding interfaces, and improving concordance between the design process and the in silico oracle used to select promising designs. In addressing the first, the field has shifted towards all-atom models that capture sidechain conformations in atomistic detail by eschewing data-efficient SE(3)-equivariance, mirroring the evolution of AlphaFold2 to AlphaFold3. In addressing the second, recent work has focused on replacing generative models employing diffusion or flow-matching with hallucination approaches that directly optimize the oracle in sequence space; this improves success rates but reduces computational efficiency. Here, we close and surpass the generation-hallucination gap by revisiting SE(3)-equivariance using a branched polymer treatment of protein structures. The resulting diffusion model, Genie 3, achieves state-of-the-art performance on binder design, motif scaffolding, and unconditional generation, while being significantly faster than the best existing methods. We use Genie 3 to design a nanomolar binder of Nipah Glycoprotein G, a tetramer with minimal structural or biophysical characterization, as part of the Adaptyv Bio Nipah Competition, achieving a 12.5% success rate. Taken together, our results present a new frontier in protein design capability and a reexamination of the role of SE(3)-equivariance in molecular modeling.

Advances in machine learning have transformed structural biology, enabling swift and accurate prediction of protein structure from sequence. However, key challenges persist in modeling side-chain packing, condition-dependent conformational changes and biomolecular interactions, largely because of limited high-quality training data. At the same time, emerging experimental techniques such as cryo-electron microscopy (cryo-EM), cryo-electron tomography (cryo-ET) and high-throughput crystallography are generating vast amounts of structural information but converting these data into mechanistically interpretable atomic models often remains difficult. Here we show that integrating experimental measurements directly into protein structure prediction can overcome these limitations. We introduce ROCKET, an augmentation of AlphaFold2 that refines predicted structures using cryo-EM, cryo-ET and X-ray crystallography data. By optimizing structures in the space of coevolutionary embeddings rather than Cartesian coordinates, ROCKET captures biologically meaningful structural variation that is inaccessible to AlphaFold2 alone and to existing automated modeling approaches, especially when the signal-to-noise ratio is low. ROCKET enables scalable, automated model building without retraining and provides a general framework for integrating experimental observables with biomolecular machine learning.

A bstract: The ever-increasing availability of large-scale single-cell profiles presents an opportunity to develop foundation models to capture cell properties and behavior. However, standard language models such as transformers benefits from sequentially structured data with well defined absolute or relative positional relationships, while single cell RNA data have orderless gene features. Molecular-interaction graphs, such as gene regulatory networks (GRN) or protein-protein interaction (PPI) networks, offer graph structure-based models that effectively encode both non-local gene token dependencies, as well as potential causal relationships. We introduce GREmLN ( G ene R egulatory Em bedding-based L arge N eural model), a foundation model that leverages graph signal processing to embed gene token graph structure directly within its attention mechanism, producing biologically informed single cell specific gene embeddings. Our model faithfully captures transcriptomics landscapes and achieves superior performance relative to state-of-the-art baselines on cell type annotation, graph structure understanding, and fine-tuned reverse perturbation prediction tasks. It offers a unified and interpretable framework for learning high-capacity foundational representations that capture complex, long-range regulatory dependencies from high-dimensional single-cell transcriptomic data. Moreover, the incorporation of graph-structured inductive biases enables more parameter-efficient architectures and accelerates training convergence.

Understanding the mechanism of T-cell activation and T-cell receptor (TCR) discrimination of MHC-presented epitope peptides (pMHCs) remains an open problem. Machine learning (ML)-based prediction of TCR specificity has gained considerable recent attention. However, the capacity of current models to generalize to peptides unseen during training is currently unknown. Here, we use a proprietary cancer-patient data set that profiles TCR binding to novel regions of peptide space to show that peptide generalization remains an unsolved problem. Specifically, we show that while ML methods have demonstrable utility in predicting TCR specificity for known peptides, they fail to generalize to novel peptides. We also show that physics-based methods utilizing classical energy functions outperform ML methods when predicting TCR binding to novel peptides but underperform them on known peptides. In light of these observations, we develop a new ML model that leverages general knowledge acquired by protein foundation models to achieve better or comparable performance than either ML or biophysical methods on both in- and out-of-distribution TCR-pMHC specificity prediction. We furthermore analyze model performance as a function of distance of TCR sequence specificity between the training and test sets to quantitatively characterize the generalization potential of any given TCR-pMHC model. Our analysis sheds light on the status of modeling TCR-pMHC interactions and suggests new paths forward for continued method development and data acquisition.

Advances in machine learning have transformed structural biology, enabling swift and accurate prediction of protein structure from sequence. However, challenges persist in capturing sidechain packing, condition-dependent conformational dynamics, and biomolecular interactions, primarily due to scarcity of high-quality training data. Emerging techniques, including cryo-electron tomography (cryo-ET) and high-throughput crystallography, promise vast new sources of structural data, but translating experimental observations into mechanistically interpretable atomic models remains a key bottleneck. Here, we address these challenges by improving the efficiency of structural analysis through combining experimental measurements with a landmark protein structure prediction method – AlphaFold2. We present an augmentation of AlphaFold2, ROCKET, that refines its predictions using cryo-EM, cryo-ET, and X-ray crystallography data, and demonstrate that this approach captures biologically important structural variation that AlphaFold2 does not. By performing structure optimization in the space of coevolutionary embeddings, rather than Cartesian coordinates, ROCKET automates difficult modeling tasks, such as flips of functional loops and domain rearrangements, that are beyond the scope of current state-of-the-art methods and, in some instances, even manual human modeling. The ability to efficiently sample these barrier-crossing rearrangements unlocks a new horizon for scalable and automated model building. Crucially, ROCKET does not require retraining of AlphaFold2 and is readily adaptable to multimers, ligand-cofolding, and other data modalities. Conversely, our differentiable crystal-lographic and cryo-EM target functions are capable of augmenting other structure prediction methods. ROCKET thus provides an extensible framework for the integration of experimental observables with biomolecular machine learning.