• Franca Fraternali, Franca Fraternali, ISMB-UCL

Effective adaptive immune responses rely on antibodies of different isotypes performing distinct effector functions. Understanding their structural diversity is crucial for engineering antibodies with optimal stability, binding, and therapeutic potential. In this keynote, I will present our integrative computational approaches to guide antibody design, which include isotype classification, chain compatibility prediction, 3D structural modeling, and analysis of allosteric communication.
In designing novel antibodies, effective pairing of antibody heavy and light chains is essential for effective function, yet the rules governing this remain unclear. I will introduce ImmunoMatch, a suite of AI models fine-tuned on full-length variable regions to predict cognate H–L chain pairs. Built on the AntiBERTa2 language model, ImmunoMatch outperforms CDR- and gene usage–based models, with further improvements from chain type–specific tuning. Applied to B cell repertoires and therapeutic antibodies, ImmunoMatch identifies chain pairing refinement as a hallmark of B cell maturation and uncovers key sequence features driving specificity.
Moving beyond the traditional focus on CDRs, we show that framework (FW) mutations can modulate antibody stability and effector function through long-range structural effects. Our analyses revealed that antibody language models (AbLMs) alone lack predictive power for FW mutagenesis. To improve on this, we adopted a structure-based approach, suggesting future directions such as fine-tuning AbLMs with in vitro FW-specific mutational data to improve their utility in antibody design. This shift can broaden the scope of rational engineering toward non-CDR regions and developability attributes, highlighting the need for a holistic view of antibody design.

• Meghana Kshirsagar, Meghana Kshirsagar, Microsoft
• Artur Meller, Artur Meller, Washington University at St. Louis
• Ian R. Humphreys, Ian R. Humphreys, University of Washington
• Samuel Sledzieski, Samuel Sledzieski, Computer Science and Artificial Intelligence Laboratory
• Yixi Xu, Yixi Xu, Microsoft
• Rahul Dodhia, Rahul Dodhia, Microsoft
• Eric Horvitz, Eric Horvitz, Microsoft
• Bonnie Berger, Bonnie Berger, Massachusetts Institute of Technology
• Gregory R Bowman, Gregory R Bowman, University of Pennsylvania
• Juan Lavista Ferres, Juan Lavista Ferres, Microsoft
• David Baker, David Baker, University of Washington
• Minkyung Baek, Minkyung Baek, Seoul National University

The majority of proteins must form higher-order assemblies to perform their biological functions, yet few machine learning models can accurately and rapidly predict the symmetry of assemblies involving multiple copies of the same protein chain. Here, we address this gap by finetuning several classes of protein foundation models, to predict homo-oligomer symmetry. Our best model named Seq2Symm, which utilizes ESM2, outperforms existing template-based and deep learning methods achieving an average AUC-PR of 0.47, 0.44 and 0.49 across homo-oligomer symmetries on three held-out test sets compared to 0.24, 0.24 and 0.25 with template-based search. Seq2Symm uses a single sequence as input and can predict at the rate of ~80,000 proteins/hour. We apply this method to 5 proteomes and ~3.5 million unlabeled protein sequences, showing its promise to be used in conjunction with downstream computationally intensive all-atom structure generation methods such as RoseTTAFold2 and AlphaFold2-multimer. Code, datasets, model are available at: https://github.com/microsoft/seq2symm.

• Zhidian Zhang, Zhidian Zhang, MIT
• Yo Akiyama, Yo Akiyama, MIT
• Yehlin Cho, Yehlin Cho, MIT
• Sergey Ovchinnikov, Sergey Ovchinnikov, MIT

Homo-oligomeric protein complexes—assemblies of identical subunits—are central to many biological processes and disease mechanisms. Accurate prediction of inter-subunit contacts within these assemblies remains a key challenge, especially as existing structure predictors like AlphaFold scale poorly with complex size. Protein language models (pLMs), trained on vast sequence databases, offer a scalable alternative by learning coevolutionary statistics from single-chain inputs.

In this work, we systematically investigate whether pLMs implicitly learn inter-subunit interaction signals, even when trained solely on monomeric sequences. We find that pLM-predicted contact maps often contain partial inter-subunit signal, but the prediction perfomance is consistently weaker than intra-subunit ones. Notably, we observe that larger pLMs recover more accurate inter-contacts, suggesting that model scaling enhances structural resolution of homo-oligomers.

Interestingly, missing inter-contact signals often correspond to interfaces without strong biophysical support outside of crystallography, raising the possibility that some predicted absences reflect genuine lack of physiological relevance. Our findings suggest that inter-subunit contact prediction from pLMs could serve as a computational filter for distinguishing biologically relevant homo-oligomers from crystallographic artifacts.

These findings open a new avenue for leveraging pLMs not only as structure predictors, but as tools for dissecting the evolutionary logic and physiological relevance of protein assemblies.

• Ilya Buyanov, Ilya Buyanov, Constructor University
• Anastasia Sarycheva, Anastasia Sarycheva, Tetra D GmbH
• Petr Popov, Petr Popov, Tetra D GmbH

Accurately predicting the impact of single-point mutations on protein stability is crucial for elucidating molecular
mechanisms underlying diseases in life sciences and advancing protein engineering in biotechnology. With recent advances in deep learning and protein structure prediction, deep learning approaches are expected to surpass existing methods for predicting protein thermostability. However, structure-based deep learning models, specifically convolutional neural networks, are affected by orientation biases, leading to inconsistent predictions with respect to the input protein orientation. In this study, we present OrgNet, a novel orientation-gnostic deep learning model employing three-dimensional convolutional neural networks to predict protein thermostability change upon point mutation. OrgNet encodes protein structures as voxel grids, enabling the model to capture fine-grained, spatially localized atomic features. OrgNet implements spatial transforms to standardize input protein orientations, thus eliminating orientation bias problem. When evaluated on established benchmarks, including Ssym and S669, OrgNet achieves state-of-the-art performance, demonstrating superior accuracy and robust performance compared to existing methods. OrgNet is available at https://github.com/i-Molecule/OrgNet.

• Adrian Mulholland

Molecular simulations contribute to practical protein design and engineering workflows. Equilibrium molecular dynamics (MD) simulations not only test and filter designs, but can predict binding, redox and other properties for engineering and optimization. This includes activation heat capacities determining enzyme temperature activity optima, and analysing causes of epistasis. A particular challenge is understanding and predicting mutations far from the active site that affect activity, often introduced by evolution. Dynamical-nonequilibrium (D-NEMD) simulations can predict distal sites relevant to modulating activity, cryptic binding sites, and allosteric effects. Simulations of chemical reactions in proteins with combined quantum mechanics/molecular mechanics (QM/MM) methods characterize crucial species in catalysis, including transition states and reaction intermediates, and how they are formed and stabilized. QM/MM models can be used as ‘theozyme’ templates for enzyme design. QM/MM calculations also allow prediction of spectroscopic and other electronic properties, assisting in design and optimization of photovoltaic proteins, e.g. in designed spectral tuning. Simulations can also analyse trajectories and effects of directed and natural evolution, providing insights for enzyme design and engineering. Examples include identifying the dynamical origins of heat capacity changes introduced by directed evolution of designer enzymes, and revealing e.g. how local electric fields are optimized for specific catalytic activities in beta-lactamase enzymes that cause resistance to ‘last resort’ antibiotics. Electric fields are vital features of many natural enzymes, including heme peroxidases in which they drive proton delivery. Electric field calculations and MD simulations can be combined effectively with AI tools for protein engineering in evolutionary enzyme design.

• Ali Baran Taşdemir, Ali Baran Taşdemir, Hacettepe University
• Ayşe Berçin Barlas, Ayşe Berçin Barlas, İzmir Biomedicine and Genome Center
• Abdurrahman Olğaç, Abdurrahman Olğaç, Gazi University
• Ezgi Karaca, Ezgi Karaca, İzmir Biomedicine and Genome Center
• Tunca Doğan, Tunca Doğan, Hacettepe University

Designing novel proteins with both structural stability and targeted molecular function remains a central challenge in computational biology. While recent generative models such as diffusion and flow-matching offer promising capabilities for protein backbone generation, functional controllability is still limited. In this work, we introduce FlowProt, a classifier-guided flow-matching generative model designed to create protein backbones with domain-specific functional properties. As a case study, we focus on the catalytic domain of human DNA methyltransferase DNMT3A, a 286-residue protein essential in early epigenetic regulation.

FlowProt builds on the FrameFlow architecture, predicting per-residue translation and rotation matrices to reconstruct 3D backbones from noise. A domain classifier, trained to distinguish DNMT proteins from others, guides the model during inference using gradient-based feedback. This enables FlowProt to steer generation toward DNMT-like structures. We evaluate backbone quality using self-consistency metrics (scRMSD, scTM, pLDDT) and domain relevance using ProGReS, sequence similarity, and SAM-binding potential.

FlowProt consistently generates high-confidence structures up to 286 residues—the exact length of DNMT3A—with low scRMSD, high scTM, and strong functional similarity. We further validate our designs through structure-based alignment and cofactor-binding analysis with Chai-1, demonstrating high-confidence SAM-binding regions in the generated models.

To our knowledge, FlowProt is the first method to integrate flow-matching with classifier guidance for domain-specific backbone design. As future work, we aim to assess DNA-binding potential and further refine functional capabilities via molecular dynamics simulations and benchmarking against state-of-the-art protein design models.

• Remo Rohs, Remo Rohs, University of Southern California
• Jesse Weller, Jesse Weller, University of Southern California

The rapid expansion of crystal structure data and libraries of readily synthesizable molecules has recently opened up new areas of chemical space for drug discovery. Combined with advancements in virtual ligand screening, these expanded libraries are making an impact in early-stage drug discovery. However, traditional virtual screening methods are still only able to explore a small fraction of the near-infinite drug-like chemical space. Generative deep learning techniques address these limitations by leveraging existing data to learn the key intra- and inter-molecular relationships in drug-target interactions. We present DrugHIVE, a deep hierarchical variational autoencoder that surpasses leading autoregressive and diffusion-based models in both speed and performance on standard generative tasks. Our model generates molecules in a rapid single-shot fashion, making it highly scalable and orders of magnitude faster than other top approaches requiring slow, multi-step inference. DrugHIVE’s hierarchical architecture provides enhanced control over molecular generation, enabling substantial improvements in virtual screening efficiency and automating various drug design processes such as de novo generation, molecular optimization, scaffold hopping, linker design, and high-throughput pattern replacement. We demonstrate an improved ability to optimize drug-like properties, synthesizability, binding affinity, and selectivity of molecules through evolutionary latent space search using both experimentally resolved and AlphaFold predicted receptor structures. Recently, we used DrugHIVE to design novel compounds as prospective therapeutics for the important P53 cancer target. These promising new compounds have been synthesized and are currently undergoing experimental testing.

• Xiangru Tang, Xiangru Tang, Yale University
• Xinwu Ye, Xinwu Ye, Yale University
• Fang Wu, Fang Wu, Stanford University
• Daniel Shao, Daniel Shao, Yale University
• Dong Xu, Dong Xu, University of Missouri
• Mark Gerstein, Mark Gerstein, Yale University

Inverse protein folding, which aims to design amino acid sequences for desired protein structures, is fundamental to protein engineering and therapeutic development. While recent deep-learning approaches have made remarkable progress, they typically represent biochemical properties as discrete features associated with individual residues. Here, we present BC-Design, a framework that represents biochemical properties as continuous distributions across protein surfaces and interiors. Through contrastive learning, our model learns to encode essential biochemical information within structure embeddings, enabling sequence prediction using only structural input during inference—maintaining compatibility with real-world applications while leveraging biochemical awareness. BC-Design achieves 88% sequence recovery versus state-of-the-art methods’ 67% (a 21% absolute improvement) and reduces perplexity from 2.4 to 1.5 (39.5% relative improvement) on the CATH 4.2 benchmark. Notably, our model exhibits robust generalization across diverse protein characteristics, performing consistently well on proteins of varying sizes (50-500 residues), structural complexity (measured by contact order), and all major CATH fold classes. Through ablation studies, we demonstrate the complementary contributions of structural and biochemical information to this performance. Overall, BC-Design establishes a new paradigm for integrating multimodal protein information, opening new avenues for computational protein engineering and drug discovery.

• Xinyi Zhou, Xinyi Zhou, The Chinese University of Hong Kong
• Guibao Shen, Guibao Shen, The Hong Kong University of Science and Technology
• Yingcong Chen, Yingcong Chen, The Hong Kong University of Science and Technology
• Guangyong Chen, Guangyong Chen, Hangzhou Institute of Medicine
• Pheng Ann Heng, Pheng Ann Heng, The Chinese University of Hong Kong

Motivation: Structure-based protein design is crucial for designing proteins with novel structures and functions, which aims to generate sequences that fold into desired structures. Current deep learning-based methods primarily focus on training and evaluating models using sequence recovery-based metrics. However, this approach overlooks the inherent ambiguity in the relationship between protein sequences and structures. Relying solely on sequence recovery as a training objective limits the models’ ability to produce diverse sequences that maintain similar structures. These limitations become more pronounced when dealing with remote homologous proteins, which share functional and structural similarities despite low sequence identity.
Results: Here, we present DivPro, a model that learns to design diverse sequences that can fold into similar structures. To improve sequence diversity, instead of learning a single fixed sequence representation for an input structure as in existing methods, DivPro learns a probabilistic sequence space from which diverse sequences could be sampled. We leverage the recent advancements in in-silico protein structure prediction. By incorporating structure prediction results as training guidance, DivPro ensures that sequences sampled from this learned space reliably fold into the target structure. We conduct extensive experiments on three sequence design benchmarks and evaluated the structures of designed sequences using structure prediction models including AlphaFold2. Results show that DivPro can maintain high structure recovery while significantly improve the sequence diversity.

• Dmitry Molodenskiy, Dmitry Molodenskiy, European Molecular Biology Laboratory Hamburg
• Valentin Maurer, Valentin Maurer, European Molecular Biology Laboratory Hamburg
• Dingquan Yu, Dingquan Yu, European Molecular Biology Laboratory Hamburg
• Grzegorz Chojnowski, Grzegorz Chojnowski, European Molecular Biology Laboratory Hamburg
• Stefan Bienert, Stefan Bienert, Biozentrum
• Gerardo Tauriello, Gerardo Tauriello, Biozentrum
• Konstantin Gilep, Konstantin Gilep, European Molecular Biology Laboratory Hamburg
• Torsten Schwede, Torsten Schwede, Biozentrum
• Jan Kosinski, Jan Kosinski, European Molecular Biology Laboratory Hamburg

AlphaPulldown2 streamlines protein structural modeling by automating workflows, improving code adaptability, and optimizing data management for large-scale applications. It introduces an automated Snakemake pipeline, compressed data storage, support for additional modeling backends like AlphaFold3 and AlphaLink2, and a range of other improvements. These upgrades make AlphaPulldown2 a versatile platform for predicting both binary interactions and complex multi-unit assemblies.

• Michel van Kempen, Michel van Kempen, Max Planck Institute for Multidisciplinary Sciences
• Johannes Soeding, Johannes Soeding, Max Planck Institute for Multidisciplinary Sciences

Fast protein structure search methods, such as Foldseek, are essential to make use of the vast amount of structural information generated by structure prediction methods. In Foldseek, the key idea is to represent structures as sequences of discrete tokens from a structural alphabet, enabling fast searches through structure databases using efficient sequence comparison methods. Foldseek uses the 3Di alphabet for structure representation. However, its structure representation, comprising 20 states, describes only a limited aspect of the overall structure, resulting in lower search sensitivity compared to methods like TMalign or Dali, which use the full structure. To further improve structure search sensitivity, we present a new structural alphabet as an extension to the established 3Di alphabet. Instead of replacing 3Di, our new alphabet was trained to encode structural information complementary to the 3Di states. The combination of the two alphabets allows to balance search sensitivity and speed: the 3Di alphabet alone is used for the most time-critical tasks, while the final alignments benefit from additional structural information from both alphabets, increasing the entire search performance. On the SCOPe dataset, extending the 3Di alphabet with 12 states of the new alphabet increases search sensitivity at the superfamily level by 22%, compared to 13% when adding the amino acid alphabet instead. Moreover, adding the new alphabet as a third alphabet to Foldseek improves its search sensitivity by 4.5%.

• Justas Dapkunas, Justas Dapkunas, Institute of Biotechnology
• Kliment Olechnovic, Kliment Olechnovic, Institute of Biotechnology
• Ceslovas Venclovas, Ceslovas Venclovas, Institute of Biotechnology

To accomplish their functions in living organisms, proteins usually interact with various biological macromolecules, including other proteins and nucleic acids. Despite recent progress in structure prediction, only part of these interactions can be predicted accurately, and modeling those involving nucleic acids is especially hard. Therefore, improved computational methods for analysis and prediction of biomolecular interactions are in high demand. The development of such methods largely depends on the availability of reliable data. However, the experimental data in the Protein Data Bank (PDB) are noisy and hard to interpret. To facilitate the analysis of the biomolecular interactions, we developed the PPI3D web resource that is based on a database of clustered non-redundant sets of biomolecular complexes, interaction interfaces and binding sites. The structures are clustered based on both sequence and structure similarity, thus retaining the alternative interaction modes. All protein-protein, protein-peptide and protein-nucleic acid interaction interfaces and binding sites are pre-analyzed by means of Voronoi tessellation. The data are updated every week to keep in sync with the PDB. The users can query the data by different criteria, select the interactions of interest, download the desired data subsets in tabular format and as coordinate files, and use them for detailed investigation of protein interactions or for training the machine learning models. We expect that the PPI3D clusters will become a useful resource for researchers working on diverse problems related to biomolecular interactions. PPI3D is available at http://bioinformatics.ibt.lt/ppi3d/.

• Fatima Shahid, Fatima Shahid, Faculty of Science and Technology
• Neeladri Sen, Neeladri Sen, Department of Structural and Molecular Biology
• Christine Orengo, Christine Orengo, Department of Structural and Molecular Biology
• Su Datt Lam, Su Datt Lam, Universiti Kebangasaan Malaysia

Plants face significant environmental stress such as pathogens, salinity, drought, and extreme temperatures. To survive, they evolve diverse adaptive mechanisms. Genome-wide association studies (GWAS) are widely used to identify genes linked to stress resistance, but often generate too many variants to interpret easily.

This study maps GWAS-derived missense mutations to rice protein structures to prioritise those with functional impact. Despite limited experimentally determined plant protein structures, resources like the AlphaFold Protein Structure Database and The Encyclopedia of Domains (TED) offer high-quality models and domain annotations. We focused on TED domains with reliable structure—excluding those with low pLDDT scores, disorder, poor packing, or non-globular features.

Stress-resistance mutations from the GWAS Atlas were then mapped to these domains. Functional sites were predicted using P2Rank and AlphaFill, and proximity of mutations to these sites was analysed. Among 149 mutations mapped to 113 TED domains, 14 were predicted as non-deleterious by MutPred2—potential gain-of-function variants. 70 mutations were near predicted functional sites.

To explore potential impacts on protein interactions, AlphaFold 3 was used to model 24 protein complexes, and mCSM-PPI2 estimated changes in binding affinity. Some mutations enhanced protein-protein interactions. We calculated predicted changes in binding affinity following mutations using mCSM-PPI2. Several interesting cases demonstrated increased binding to interacting partners, which will be discussed in the talk.

This is the first study using AlphaFold models to investigate stress-resistance mutations in plants, providing insights into their functional impact and supporting future breeding strategies vital for food security amid climate change.

• Sevastianos Korsak, Sevastianos Korsak, Faculty of Mathematics and Information Science
• Krzysztof H Banecki, Krzysztof H Banecki, Faculty of Mathematics and Information Science
• Karolina Buka, Karolina Buka, Centre of New Technologies
• Piotr Górski, Piotr Górski, Faculty of Physics
• Dariusz Plewczynski, Dariusz Plewczynski, Centre of New Technologies

Modeling DNA replication poses significant challenges due to the intricate interplay of biophysical processes and the need for precise parameter optimization. In this study, we explore the interactions among three key biophysical factors that influence chromatin folding: replication, loop extrusion, and compartmentalization. Replication forks, which act as moving barriers to loop extrusion factors, contribute to the dynamic reorganization of chromatin during S phase. Notably, replication timing is known to correlate with the phase separation of chromatin into A and B compartments. Our approach integrates three components: (1) a numerical model that uses single-cell replication timing data to simulate fork propagation; (2) a stochastic Monte Carlo simulation capturing loop extrusion dynamics, CTCF and fork barriers, and epigenetic state spreading via a Potts Hamiltonian; and (3) a 3D OpenMM simulation that reconstructs chromatin structure based on the resulting state trajectories. In this work, we model the dynamic evolution of chromatin states using co-evolutionary graphs, in which both node and link states evolve stochastically and interactively. These graphs are translated into 3D chromatin structures: links correspond to harmonic bonds representing physical loops, while node states determine compartmental interactions modeled via block-copolymer attractive forces. We reconstruct 3D chromatin trajectories across the cell cycle by incorporating biologically grounded force-field parameters that vary between cell cycle phases to reflect experimentally observed changes in chromatin organization. Our framework, to our knowledge the first to dynamically integrate these three biophysical factors, provides new insights into chromatin behavior during replication and reveals how replication stress impacts chromatin organization.

• Clément Bernard, Clément Bernard, Universite-Paris-Saclay
• Guillaume Postic, Guillaume Postic, Universite-Paris-Saclay
• Sahar Ghannay, Sahar Ghannay, LISN-CNRS / Université Paris-Saclay
• Fariz Tahi, Fariz Tahi, Universite-Paris-Saclay

Predicting the 3D structure of RNA is an ongoing challenge that has yet to be completely addressed despite continuous advancements. RNA 3D structures rely on distances between residues and base interactions but also backbone torsional angles. Knowing the torsional angles for each residue could help reconstruct its global folding, which is what we tackle in this work. We present a novel approach for directly predicting RNA torsional angles from raw sequence data. Our method draws inspiration from the successful application of language models in various domains and adapts them to RNA.
We have developed a language-based model, RNA-TorsionBERT, incorporating better sequential interactions for predicting RNA torsional and pseudo-torsional angles from the sequence only. Through extensive benchmarking, we demonstrate that our method improves the prediction of torsional angles compared to state-of-the-art methods. In addition, by using our predictive model, we have inferred a torsion angle-dependent scoring function, called TB-MCQ, that replaces the true reference angles by our model prediction. We show that it accurately evaluates the quality of near-native predicted structures, in terms of RNA backbone torsion angle values. Our work demonstrates promising results, suggesting the potential utility of language models in advancing RNA 3D structure prediction.
The source code is freely available on the EvryRNA platform: https://evryrna.ibisc.univ-evry.fr/evryrna/RNA-TorsionBERT.