Proteins in complex with small-molecule ligands represent the core of structure-based drug discovery. However, three-dimensional representations are absent from most deep-learning-based generative models. Here, we present a graph-based generative modeling technology that encodes explicit 3D protein-ligand contacts within a relational graph architecture and evaluate its behavior using the dopamine D2 receptor (DD2R) as a model system. The models combine a conditional variational autoencoder that allows for activity-specific molecule generation with putative contact generation that provides predictions of molecular interactions within the target-binding pocket. We show that molecules generated with our 3D procedure are more compatible with the DD2R-binding pocket than those produced by a comparable ligand-based 2D generative method, as measured by docking scores, expected stereochemistry, and recoverability in commercial chemical databases. Predicted protein-ligand contacts were found to be among the highest-ranked docking poses with a high recovery rate. Overall, this work shows how the structural context of a protein target can enhance the generation of small molecules within a realistic binding environment.

The rapid growth of quantum computing is driven by promises of performing complex calculations with unprecedented speed; however, current use cases have been limited by quantum hardware and the difficulty of identifying problems that classical computers cannot easily address. Within these constraints, biological problems including drug discovery, protein folding, and precision medicine present an opportunity to understand how current quantum hardware can make advances. In immunology, accurate prediction of cancer neoantigens remains a major challenge, limited by small, noisy datasets and the inability of classical models to generalize. In approaching the problem, we explore multiple noise mitigation techniques, including Pauli twirling and dynamical decoupling, in conjunction with controlled shot-based sampling to stabilize training on real hardware and in a warm start hybrid approach. With these approaches, we demonstrate the use of Quantum Convolutional Neural Networks (QCNNs) for both MHC binding and immunogenicity prediction, including a quantum hardware experiment involving 46 qubits that achieved a 6% increase in classification accuracy with fewer training samples compared to classical approaches. Building on these models, we introduce Quantum Convolutional HLA Immunogenic Peptide Prediction (Q-CHIPP), a combinatorial framework integrating MHC binding and T-cell recognition. It targets HLA-A*02:01-restricted 9-mer peptides and, more accurately, identifies those peptides known to be immunogenic, improving the prognostic impact of predicted neoantigen load. Together, these represent a large-scale application of QCNNs in biomedical modeling, highlighting both the feasibility and promise of quantum machine learning for data-limited biological systems and establishing a scalable foundation for quantum-enhanced biomedical research.

T cell receptors (TCRs) are critical for immune surveillance and successful adaptive immune response against foreign antigens. TCRs drive this key arm of the immune system through recognition of peptide epitopes presented on MHC complexes. However, they are limited due to their stochastic nature and generation via genetic recombination. In silico design of functional TCRs that target defined peptide epitopes would be of considerable utility but has up until now been unsuccessful. Here, we develop an artificial intelligence (AI)-powered approach using a hybrid physics-based simulation and generative AI that successfully engineers TCRs against defined epitopes presented by MHC-I. We use this approach to design TCRs against two cancer antigens, a HERC1 neoantigen and an immunogenic neoepitope in mutant EGFR. We engineer multiple TCRs against the HERC1 neoantigen which activate T cells in response to exposure to peptide-MHC I and kill cancer cells more effectively than a patient-derived TCR. In addition, we used generative AI to design functional TCRs that target the EGFR T790M neoantigen, engineering greater specificity against the mutant sequence. We present an AI-based approach to TCR design with broad utility for efforts to engineer TCRs and for the development of new cell therapies. One sentence summary:Artificial intelligence-based approach enables the directed engineering of functional TCRs with enhanced features that target cancer neoantigens.

Modern AI (Artificial Intelligence) methods offer new opportunities in pharmacology by enabling improved modeling of disease mechanisms and drug action learned from large and heterogeneous biological datasets. A central challenge is developing models that can jointly integrate disparate biomedical modalities. We introduce MAMMAL (Molecular Aligned Multi Modal Architecture and Language), a foundation model for cross-modal learning, designed to address the challenges associated with drug discovery tasks. MAMMAL was pre-trained on 2 billion samples across protein and antibody sequences, small molecules, and gene expression profiles, and supports classification, regression, and generative tasks on cross-modal inputs. Across eleven benchmarks covering multiple stages of the drug discovery pipeline, MAMMAL achieves state-of-the-art performance on nine tasks and competitive results on two. In an antibody-antigen binding benchmark, fine-tuned MAMMAL prediction scores significantly outperform AlphaFold3 confidence scores, used here as a reference proxy for binding likelihood, in five of seven antigen targets. The MAMMAL framework and pretrained models are publicly available to support open and collaborative research.

Monoclonal antibodies (mAbs) represent one of the most prevalent FDA-approved treatments for autoimmune, infectious, and cancer diseases. However, their discovery and development remains a time-consuming and costly process. Recent advancements in machine learning (ML) and artificial intelligence (AI) have shown significant promise in revolutionizing antibody discovery field. Models that predict antibody biological activity enable in silico evaluation of binding and functional properties; such models can prioritize antibodies with the highest likelihood of success in laboratory testing procedures. We explore an AI model for predicting the binding and receptor blocking activity of antibodies against influenza A hemagglutinin (HA) antigens. Our model is developed with the Molecular Aligned Multi-Modal Architecture and Language (MAMMAL) framework for biologics discovery to predict antibody-antigen interactions using only sequence information. To evaluate the model's performance, we tested it under various data split conditions to mimic real-world scenarios. Our model achieved an area under the receiver operating characteristic (AUROC) score of ≥ 0.91 for predicting the activity of existing antibodies against seen HAs and an AUROC score of 0.9 for unseen HAs. For novel antibody activity prediction, the AUROC was 0.73, which further declined to 0.63-0.66 under stringent constraints on similarity to existing antibodies. These results demonstrate the potential of AI foundation models to transform antibody design by reducing dependence on extensive laboratory testing and enabling more efficient prioritization of antibody candidates. Moreover, our findings emphasize the critical importance of diverse and comprehensive antibody datasets to improve the generalization of prediction models, particularly for novel antibody development.

The immune system is an intricately evolved series of cellular and protein-protein interactions, which defend the body against pathogens and abnormal cells such as cancer. A key player in the immunologic recognition of non-self is the immune synapse, where T cell receptors (TCRs) scan peptides presented on major histocompatibility complex (MHC) molecules to detect and eliminate cells displaying non-self antigens. While this interaction is vital for vaccine and immunotherapy success, the underlying rules of TCR recognition remain poorly understood. This is only further challenged as the application of predictive models is very limited due to small training datasets. While traditional machine learning models excel at predicting neoantigen binding to MHC, they often struggle to accurately predict immunogenicity. To address these challenges, we developed a quantum computing approach using Quantum Convolutional Neural Networks (QCNNs). Here we present Quantum Convolutional Human Leukocyte Antigen (HLA) Immunogenic Peptide Prediction (Q-CHIPP), the first application of quantum hardware based on training/predicting both MHC binding and immunogenicity in a combinatorial approach. Additionally, we present a large scale use of quantum hardware at scale with 46 qubits. This study underscores how quantum technology can be used for biological modeling and presents a scalable QCNN design with the potential to overcome current computational bottlenecks as quantum hardware advances. ### Competing Interest Statement T.A.C. is a co-founder of Gritstone Oncology and holds equity. T.A.C. holds equity in An2H. T.A.C. acknowledges grant funding from Bristol-Myers Squibb, AstraZeneca, Illumina, Pfizer, An2H, and Eisai. T.A.C. has served as an advisor for Bristol-Myers, MedImmune, Squibb, Illumina, Eisai, AstraZeneca, and An2H. T.A.C. are inventors on intellectual property held by MSKCC on using tumor mutation burden to predict immunotherapy response, with pending patent, which has been licensed to PGDx. National Science Foundation, No. DBI-1548297

[This corrects the article DOI: 10.1016/j.csbj.2025.03.038.].

Large language models applied to vast biological datasets have the potential to transform biology by uncovering disease mechanisms and accelerating drug development. However, current models are often siloed, trained separately on small-molecules, proteins, or transcriptomic data, limiting their ability to capture complex, multi-modal interactions. Effective drug discovery requires computational tools that integrate multiple biological entities while supporting prediction and generation, a challenge existing models struggle to address. For this purpose, we present MAMMAL - Molecular Aligned Multi-Modal Architecture and Language - a versatile method applied to create a multi-task foundation model that learns from large-scale biological datasets across diverse modalities, including proteins, small-molecules, and omics. MAMMAL's structured prompt syntax supports classification, regression, and generation tasks while handling token and scalar inputs and outputs. Evaluated on eleven diverse downstream tasks, it reaches a new state of the art (SOTA) in nine tasks and is comparable to SOTA in two tasks, all within a unified architecture, unlike prior task-specific models. Additionally, we explored Alphafold 3 binding prediction capabilities on antibody-antigen and nanobody-antigen complexes showing significantly better classification performance of MAMMAL in 3 out of 4 targets. The model code and pretrained weights are publicly available at https://github.com/BiomedSciAI/biomed-multi-alignment and https://huggingface.co/ibm/biomed.omics.bl.sm.ma-ted-458m

Cytotoxic-T-lymphocyte (CTL) mediated control of HIV-1 is enhanced by targeting highly networked epitopes in complex with human-leukocyte-antigen-class-I (HLA-I). However, the extent to which the presenting HLA allele contributes to this process is unknown. Here we examine the CTL response to QW9, a highly networked epitope presented by the disease-protective HLA-B57 and disease-neutral HLA-B53. Despite robust targeting of QW9 in persons expressing either allele, T cell receptor (TCR) cross-recognition of the naturally occurring variant QW9_S3T is consistently reduced when presented by HLA-B53 but not by HLA-B57. Crystal structures show substantial conformational changes from QW9-HLA to QW9_S3T-HLA by both alleles. The TCR-QW9-B53 ternary complex structure manifests how the QW9-B53 can elicit effective CTLs and suggests sterically hindered cross-recognition by QW9_S3T-B53. We observe populations of cross-reactive TCRs for B57, but not B53 and also find greater peptide-HLA stability for B57 in comparison to B53. These data demonstrate differential impacts of HLAs on TCR cross-recognition and antigen presentation of a naturally arising variant, with important implications for vaccine design.

Immunologic recognition of peptide antigens bound to class I major histocompatibility complex (MHC) molecules is essential to both novel immunotherapeutic development and human health at large. Current methods for predicting antigen peptide immunogenicity rely primarily on simple sequence representations, which allow for some understanding of immunogenic features but provide inadequate consideration of the full scale of molecular mechanisms tied to peptide recognition. We here characterize contributions that unsupervised and supervised artificial intelligence (AI) methods can make toward understanding and predicting MHC(HLA-A2)-peptide complex immunogenicity when applied to large ensembles of molecular dynamics simulations. We first show that an unsupervised AI method allows us to identify subtle features that drive immunogenicity differences between a cancer neoantigen and its wild-type peptide counterpart. Next, we demonstrate that a supervised AI method for class I MHC(HLA-A2)-peptide complex classification significantly outperforms a sequence model on small datasets corrected for trivial sequence correlations. Furthermore, we show that both unsupervised and supervised approaches reveal determinants of immunogenicity based on time-dependent molecular fluctuations and anchor position dynamics outside the MHC binding groove. We discuss implications of these structural and dynamic immunogenicity correlates for the induction of T cell responses and therapeutic T cell receptor design.

The application of deep learning to generative molecule design has shown early promise for accelerating lead series development. However, questions remain concerning how factors like training, data set, and seed bias impact the technology's utility to medicinal and computational chemists. In this work, we analyze the impact of seed and training bias on the output of an activity-conditioned graph-based variational autoencoder (VAE). Leveraging a massive, labeled data set corresponding to the dopamine D2 receptor, our graph-based generative model is shown to excel in producing desired conditioned activities and favorable unconditioned physical properties in generated molecules. We implement an activity-swapping method that allows for the activation, deactivation, or retention of activity of molecular seeds, and we apply independent deep learning classifiers to verify the generative results. Overall, we uncover relationships between noise, molecular seeds, and training set selection across a range of latent-space sampling procedures, providing important insights for practical AI-driven molecule generation.

Proteins in complex with small molecule ligands represent the core of structure-based drug discovery. However, three-dimensional representations are absent from most deep-learning-based generative models. We here present a graph-based generative modeling technology that encodes explicit 3D protein-ligand contacts within a relational graph architecture. The models combine a conditional variational autoencoder that allows for activity-specific molecule generation with putative contact generation that provides predictions of molecular interactions within the target binding pocket. We show that molecules generated with our 3D procedure are more compatible with the binding pocket of the dopamine D2 receptor than those produced by a comparable ligand-based 2D generative method, as measured by docking scores, expected stereochemistry, and recoverability in commercial chemical databases. Predicted protein-ligand contacts were found among highest-ranked docking poses with a high recovery rate. This work shows how the structural context of a protein target can be used to enhance molecule generation.

We here present a streamlined, explainable graph convolutional neural network (gCNN) architecture for small molecule activity prediction. We first conduct a hyperparameter optimization across nearly 800 protein targets that produces a simplified gCNN QSAR architecture, and we observe that such a model can yield performance improvements over both standard gCNN and RF methods on difficult-to-classify test sets. Additionally, we discuss how reductions in convolutional layer dimensions potentially speak to the "anatomical" needs of gCNNs with respect to radial coarse graining of molecular substructure. We augment this simplified architecture with saliency map technology that highlights molecular substructures relevant to activity, and we perform saliency analysis on nearly 100 data-rich protein targets. We show that resultant substructural clusters are useful visualization tools for understanding substructure-activity relationships. We go on to highlight connections between our models' saliency predictions and observations made in the medicinal chemistry literature, focusing on four case studies of past lead finding and lead optimization campaigns.

We present a simple, modular graph-based convolutional neural network that takes structural information from protein-ligand complexes as input to generate models for activity and binding mode prediction. Complex structures are generated by a standard docking procedure and fed into a dual-graph architecture that includes separate subnetworks for the ligand bonded topology and the ligand-protein contact map. Recent work has indicated that data set bias drives many past promising results derived from combining deep learning and docking. Our dual-graph network allows contributions from ligand identity that give rise to such biases to be distinguished from effects of protein-ligand interactions on classification. We show that our neural network is capable of learning from protein structural information when, as in the case of binding mode prediction, an unbiased data set is constructed. We next develop a deep learning model for binding mode prediction that uses docking ranking as input in combination with docking structures. This strategy mirrors past consensus models and outperforms a baseline docking program (AutoDock Vina) in a variety of tests, including on cross-docking data sets that mimic real-world docking use cases. Furthermore, the magnitudes of network predictions serve as reliable measures of model confidence.

Objective: Viral infections influence intracellular peptide repertoires available for presentation by HLA-I. Alterations in HLA-I/peptide complexes can modulate binding of killer immunoglobuline-like receptors (KIRs) and thereby the function of natural killer (NK) cells. Although multiple studies have provided evidence that HLA-I/KIR interactions play a role in HIV-1 disease progression, the consequence of HIV-1 infection for HLA-I/KIR interactions remain largely unknown. Design: We determined changes in HLA-I presented peptides resulting from HIV-1-infection of primary human CD4+ T cells and assessed the impact of changes in peptide repertoires on HLA-I/KIR interactions. Methods: Liquid chromatography-coupled tandem mass spectrometry to identify HLA-I presented peptides, cell-based in-vitro assays to evaluate functional consequences of alterations in immunopeptidome and atomistic molecular dynamics simulations to confirm experimental data. Results: A total of 583 peptides exclusively presented on HIV-1-infected cells were identified, of which only 0.2% represented HIV-1 derived peptides. Focusing on HLA-C∗03 : 04/KIR2DL3 interactions, we observed that HLA-C∗03 : 04-presented peptides derived from noninfected CD4+ T cells mediated stronger binding of inhibitory KIR2DL3 than peptides derived from HIV-1-infected cells. Furthermore, the most abundant peptide presented by HLA-C∗03 : 04 on noninfected CD4+ T cells (VIYPARISL) mediated the strongest KIR2DL3-binding, while the most abundant peptide presented on HIV-1-infected cells (YAIQATETL) did not mediate KIR2DL3-binding. Molecular dynamics simulations of HLA-C∗03 : 04/KIR2DL3 interactions in the context of these two peptides revealed that VIYPARISL significantly enhanced the HLA-C∗03 : 04/peptide contact area to KIR2DL3 compared with YAIQATETL. Conclusion: These data demonstrate that HIV-1 infection-induced changes in HLA-I-presented peptides can reduce engagement of inhibitory KIRs, providing a mechanism for enhanced activation of NK cells by virus-infected cells.

Nucleic acid aptamers hold great promise for therapeutic applications due to their favorable intrinsic properties, as well as high-throughput experimental selection techniques. Despite the utility of the systematic evolution of ligands by the exponential enrichment (SELEX) method for aptamer determination, complementary in silico aptamer design is highly sought after to facilitate virtual screening and increased understanding of important nucleic acid-protein interactions. Here, with a combined experimental and theoretical approach, we have developed two optimal epithelial cellular adhesion molecule (EpCAM) aptamers. Our structure-based in silico method first predicts their binding modes and then optimizes them for EpCAM with molecular dynamics simulations, docking, and free energy calculations. Our isothermal titration calorimetry experiments further confirm that the EpCAM aptamers indeed exhibit enhanced affinity over a previously patented nanomolar aptamer, EP23. Moreover, our study suggests that EP23 and the de novo designed aptamers primarily bind to EpCAM dimers (and not monomers, as hypothesized in previous published works), suggesting a paradigm for developing EpCAM-targeted therapies.

We present a simple, modular graph-based convolutional neural network that takes structural information from protein-ligand complexes as input to generate models for activity and binding mode prediction. Complex structures are generated by a standard docking procedure and fed into a dual-graph architecture that includes separate sub-networks for the ligand bonded topology and the ligand-protein contact map. This network division allows contributions from ligand identity to be distinguished from effects of protein-ligand interactions on classification. We show, in agreement with recent literature, that dataset bias drives many of the promising results on virtual screening that have previously been reported. However, we also show that our neural network is capable of learning from protein structural information when, as in the case of binding mode prediction, an unbiased dataset is constructed. We develop a deep learning model for binding mode prediction that uses docking ranking as input in combination with docking structures. This strategy mirrors past consensus models and outperforms the baseline docking program in a variety of tests, including on cross-docking datasets that mimic real-world docking use cases. Furthermore, the magnitudes of network predictions serve as reliable measures of model confidence

Protein unfolding dynamics are bound by their degree of entropy production, a quantity that relates the amount of heat dissipated by a nonequilibrium process to a system's forward and time-reversed trajectories. We here explore the statistics of heat dissipation that emerge in protein molecules subjected to a chemical denaturant. Coupling large molecular dynamics datasets and Markov state models with the theory of entropy production, we demonstrate that dissipative processes can be rigorously characterized over the course of the urea-induced unfolding of the protein chymotrypsin inhibitor 2. By enumerating full entropy production probability distributions as a function of time, we first illustrate that distinct passive and dissipative regimes are present in the denaturation dynamics. Within the dissipative dynamical region, we next find that chymotrypsin inhibitor 2 is strongly driven into unfolded states in which the protein's hydrophobic core has been penetrated by urea molecules and disintegrated. Detailed analyses reveal that urea's interruption of key hydrophobic contacts between core residues causes many of the protein's native structural features to dissolve.

HIV controllers (HCs) are individuals who can naturally control HIV infection, partially due to potent HIV-specific CD8+ T cell responses. Here, we examined the hypothesis that superior function of CD8+ T cells from HCs is encoded by their T cell receptors (TCRs). We compared the functional properties of immunodominant HIV-specific TCRs obtained from HLA-B*2705 HCs and chronic progressors (CPs) following expression in primary T cells. T cells transduced with TCRs from HCs and CPs showed equivalent induction of epitope-specific cytotoxicity, cytokine secretion, and antigen-binding properties. Transduced T cells comparably, albeit modestly, also suppressed HIV infection in vitro and in humanized mice. We also performed extensive molecular dynamics simulations that provided a structural basis for similarities in cytotoxicity and epitope cross-reactivity. These results demonstrate that the differential abilities of HIV-specific CD8+ T cells from HCs and CPs are not genetically encoded in the TCRs alone and must depend on additional factors.

The molecular motor exploited by bacteriophage φ29 to pack DNA into its capsid is regarded as one of the most powerful mechanical devices present in viral, bacterial, and eukaryotic systems alike. Acting as a linker element, a prohead RNA (pRNA) effectively joins the connector and ATPase (adenosine triphosphatase) components of the φ29 motor. During DNA packing, this pRNA needs to withstand enormous strain along the capsid's portal axis-how this remarkable stability is achieved remains to be elucidated. We investigate the mechanical properties of the φ29 motor's three-way junction (3WJ)-pRNA using a combined steered molecular dynamics and atomic force spectroscopy approach. The 3WJ exhibits strong resistance to stretching along its coaxial helices, demonstrating its super structural robustness. This resistance disappears, however, when external forces are applied to the transverse directions. From a molecular standpoint, we demonstrate that this direction-dependent stability can be attributed to two Mg clamps that cooperate and generate mechanical resistance in the pRNA's coaxial direction. Our results suggest that the asymmetric nature of the 3WJ's mechanical stability is entwined with its biological function: Enhanced rigidity along the portal axis is likely essential to withstand the strain caused by DNA condensation, and flexibility in other directions should aid in the assembly of the pRNA and its association with other motor components.