• Daniil Litvinov, Biozentrum, University of Basel; SIB Swiss Institute of Bioinformatics, Switzerland
• Lorenzo Pantolini, Biozentrum, University of Basel; SIB Swiss Institute of Bioinformatics, Switzerland
• Peter Å krinjar, Biozentrum, University of Basel; SIB Swiss Institute of Bioinformatics, Switzerland
• Gerardo Tauriello, Biozentrum, University of Basel; SIB Swiss Institute of Bioinformatics, Switzerland
• Caitlyn McCafferty, Biozentrum, University of Basel, Switzerland
• Benjamin Engel, Biozentrum, University of Basel, Switzerland
• Torsten Schwede, Biozentrum, University of Basel; SIB Swiss Institute of Bioinformatics, Switzerland
• Janani Durairaj, Biozentrum, University of Basel; SIB Swiss Institute of Bioinformatics, Switzerland

Motivation: Protein complexes are central to cellular function, but experimental determination of their structures remains challenging. Structure
prediction methods require prior knowledge of stoichiometry - the number of copies of each protein entity within a complex. Current approaches rely
on computationally expensive brute-force methods that run structure prediction on multiple stoichiometry combinations, often with limited accuracy.

Results: We introduce Stoic, a method that uses protein language model embeddings to predict protein complex stoichiometry. Our approach
learns to identify interface residues that participate in protein-protein interactions, rather than relying on global sequence features. By integrating
these interface-aware embeddings into a graph neural network, Stoic achieves fast and accurate stoichiometry prediction for both homomeric and
heteromeric targets.

Availability: Source code for inference and training along with web versions are available in the repository at https://github.com/PickyBinders/stoic.

Contact: janani.durairaj@unibas.ch

• Luiz Felipe Piochi, INRIA, France
• Di Tang, Division of Infection Medicine (Lund University), Sweden
• Johan Malmström, Division of Infection Medicine (Lund University), SciLifeLab, Sweden
• Yasaman Karami, INRIA, France
• Hamed Khakzad, INRIA, France

Protein-protein interactions (PPIs) are central to cellular processes and host-pathogen dynamics across all domains of life, yet comprehensive interactome mapping remains challenging at the proteome scale. Experimental approaches provide only partial coverage, while existing computational methods often lack generalizability across species or are too resource-intensive for large-scale screening. Here, we introduce ppIRIS (protein-protein Interaction Regression via Iterative Siamese networks), a lightweight deep learning framework that integrates evolutionary and structural embeddings to predict PPIs directly from sequence. Evaluated on multi-species benchmarks, ppIRIS achieves state-of-the-art accuracy while enabling proteome-wide screening in minutes. Trained on curated bacterial datasets and applied to the Group A Streptococcus (GAS) proteome, ppIRIS identified functional clusters associated with virulence pathways, such as nutrient transport, stress response, and metal scavenging. Extending to cross-species prediction, ppIRIS recovered 56.2% of known GAS-human plasma interactions with enrichment in complement, coagulation, and protease inhibition pathways. Experimental validation confirmed novel predictions, demonstrating the applicability of ppIRIS for systematic discovery of bacterial and cross-species PPIs. The model together with a Google Colaboratory is freely available at github.com/lupiochi/ppIRIS.

• Sungjoon Park, LG AI Research, South Korea
• Soorin Yim, LG AI Research, South Korea
• Dongyun Kim, LG AI Research, South Korea
• Kiwoong Yoo, LG AI Research, South Korea
• Doyeong Hwang, LG AI Research, South Korea
• Kyungwook Lee, LG AI Research, South Korea
• Jongseong Jang, LG AI Research, South Korea
• Kiyoung Kim, LG AI Research, South Korea

Motivation: Binding affinity governs how proteins interact and underlies essential biological processes. Computational approaches have been developed to simulate and predict protein binding, but the scarcity of high-quality data has imposed significant constraints. One consequence is that most methods focus on predicting mutational changes in binding affinity (ΔΔG), rather than binding affinity (ΔG) itself. This practice risks overfitting to skewed data distributions, limiting the generalizability of predictions. Recent advances in protein structure prediction have enabled computational modeling of protein conformations in mass, providing rich structural information from which binding interactions can be largely explained. However, leveraging these advances for effective prediction of binding affinity has yet to translate into reliable predictions.
Results: We present PreFold-dG, a model that estimates binding affinities of protein complexes utilizing intermediate embeddings from Boltz-2, an open-source foundation model for protein structure prediction. Our approach aggregates residue-level information weighted by interresidue distance, and predicts ΔG directly rather than its derivative, ΔΔG. PreFold-dG achieved state-of-the-art performance on well-established binding affinity prediction benchmarks and demonstrated robustness on independent test sets. Ablation studies suggest that all intermediate embeddings are utilized in the prediction, whereas their contributions to modeling ΔΔG and ΔG vary. We further validated our model through case studies on real-world broadly neutralizing antibody data with evolutionary relevance.
Availability: https://github.com/LGAI-Research/PreFold-dG.

• Nicolas Fuentes-Ugarte, Departamento de Biologia. Universidad de Chile, Chile
• Tiaren Ruiz-Rojas, Facultad de Ingenieria. Universidad San Sebastian, Chile
• Felipe García-Olave, Programa de doctorado en Biologia Computacional. Universidad San Sebastian, Santiago, Chile, Chile
• Alvaro Ruiz-Fernandez, Computational Biology Lab, Centro Cientifico y Tecnologico de Excelencia, Fundacion Ciencia & Vida, Santiago, Chile, Chile
• Jose Antonio Gatare, Facultad de Ingenieria. Universidad San Sebastian, Chile
• Victor Castro-Fernandez, Departamento de Biologia. Universidad de Chile, Chile
• Raul Araya-Secchi, Facultad de Ingenieria. Universidad San Sebastian, Chile

Prestin, a member of the SLC26A family, is essential for the electromotility of mammalian outer hair cells, converting voltage changes into mechanical work. In contrast, nonmammalian orthologues function as anion transporters. To investigate the molecular and structural basis of this functional divergence, we performed ancestral sequence reconstruction (ASR) of prestin across vertebrates, followed by structural modeling using AlphaFold2-multimer and molecular dynamics simulations. We identified more than 200 amino acid substitutions along the lineage that lead to placental mammals, with early substitutions concentrated in the transmembrane domain (TMD) and late substitutions clustering in the STAS domain, particularly in the intervening sequence (IVS). Structural modeling and simulation revealed that early substitutions modulate protein–lipid interactions and interhelical contacts. In placental mammals, the IVS-loop adopts a distinct conformation that places a negatively charged patch near the chloride access pathway, potentially affecting the ion dynamics and voltage responsiveness. These structural transitions occurred without major rearrangements of the global fold of prestin, supporting a notion in which the novel function evolved through distributed substitutions within a conserved scaffold. Our findings illustrate how molecular exaptation, and incremental structural remodeling enabled the repurposing of an ancestral anion transporter into a voltage-sensitive area-motor, providing a framework for understanding the molecular evolution of complex biophysical traits central to auditory neuroscience.

• Georg Auge, Hasso Plattner Institute, Digital Engineering Faculty, University of Potsdam Germany, Germany
• Matthis Clausen, Hasso Plattner Institute, Digital Engineering Faculty, University of Potsdam Germany, Germany
• Konstantin Ketterer, Hasso Plattner Institute, Digital Engineering Faculty, University of Potsdam Germany, Germany
• Jacob Schaefer, Hasso Plattner Institute, Digital Engineering Faculty, University of Potsdam Germany, Germany
• Nils Schmitt, Hasso Plattner Institute, Digital Engineering Faculty, University of Potsdam Germany, Germany
• Tom Altenburg, Hasso Plattner Institute, Digital Engineering Faculty, University of Potsdam Germany, Germany
• Yannick Hartmaring, Hasso Plattner Institute, Digital Engineering Faculty, University of Potsdam Germany, Germany
• Hendrik Raetz, Hasso Plattner Institute, Digital Engineering Faculty, University of Potsdam Germany, Germany
• Christoph N. Schlaffner, Hasso Plattner Institute, Digital Engineering Faculty, University of Potsdam Germany, Germany
• Bernhard Y. Renard, Hasso Plattner Institute, Digital Engineering Faculty, University of Potsdam Germany, Germany

Motivation: An unprecedented amount of mass spectrometry-based proteomics data is publicly available through repositories such as the PRoteomics IDEntifications Database (PRIDE), and the field is increasingly leveraging machine learning approaches. However, the available data is not ready to be reused in a scalable way beyond the original acquisition purpose. Existing machine learning models commonly rely on a few manually curated datasets that require deep domain expertise and tedious technical work to construct. Importantly, these datasets have not been updated in recent years, so that newly published data remains inaccessible. We present usiGrabber, a scalable framework for assembling large proteomic datasets. usiGrabber is designed around portability and extensibility. It extracts spectra identification data from mzIdentML files, stores additional project-level metadata retrieved through the PRIDE API, indexes raw spectra using Universal Spectrum Identifiers (USIs), and offers download utilities to retrieve spectra data at scale.
Results: Within 49 hours, we parsed over 800 million peptide spectrum matches and corresponding USIs from over 1,200 projects. As a proof of concept, we used usiGrabber to construct a phosphorylation-specific training dataset of nearly 11 million spectra in under two days and used it to retrain a binary phosphorylation classifier based on the AHLF model architecture. With a balanced accuracy of 0.78, our model achieves comparable performance to the original model on an independent test set, showing that automated data extraction is an alternative to manual curation of static datasets.
Availability: All code is available at https://github.com/usiGrabber/usiGrabber; the data is available at https://zenodo.org/records/18853258.

• Shakiba Fadaei, University of Lausanne, Switzerland
• Fanny Krebs, University of Lausanne / SIB, Switzerland
• Vincent Zoete, University of Lausanne / SIB, Switzerland

Motivation: Protein kinases are key regulators of cellular signaling and are frequently implicated in human diseases. Although kinase domains are structurally conserved, predicting the effects of amino acid substitutions remains challenging as mutations often introduce subtle structural pertur-bations that are not captured by sequence-based or evolutionary methods. Existing supervised approaches further rely on pathogenicity annotations that are inconsistent across databases, thereby motivating the development of structure-based, label-independent frameworks for mutation effect prediction.
Results: We present a structure-based method using SE(3)-transformers to learn residue compati-bility with the local structural environment from experimentally resolved kinase 3D structures. Pro-teins are represented as atom-level graphs with physicochemical descriptors derived from the CHARMM force field and spatial connectivity. The model is trained on two self-supervised tasks given local structural context: masked residue atom reconstruction and masked residue classifica-tion. This formulation enables learning of geometric and physicochemical constraints without relying on pathogenicity labels. Evaluation using reconstruction loss, residue prediction accuracy, and comparison with BLOSUM substitution patterns indicate that the model captures biologically mean-ingful relationships between residue identity and 3D structural context. We interpret the scores as-signed to alternative amino acids as measures of structural fitness, where low-scoring residues are hypothesized to be less compatible with the local environment and more likely to induce deleterious effects on protein structure and activity.

• Yunji Kim, Seoul National University, South Korea
• Minkyung Baek, Seoul National University, South Korea

Motivation: De novo antibody design requires jointly determining the global binding orientation and shaping flexible CDR loops to engage a target epitope. Diffusion-based approaches such as RFantibody are capable of this joint task but frequently produce severe steric clashes requiring extensive post-hoc filtering. Flow-based methods such as IgFlow and FlowDesign offer more stable generation but remain restricted to pre-aligned frames, precluding true de novo design. Achieving structural integrity and epitope specificity simultaneously under this setting remains an open challenge. Results: We propose TiDE-Ab, a conditional SE(3) Flow Matching framework for de novo epitope-specific antibody design. By conditioning on unpaired antigen and antibody structures without any pre-aligned frame, TiDE-Ab inherits the structural stability of flow matching while enabling global binding pose search from scratch. To further improve epitope targeting, we introduce Time-Dependent Classifier-Free Guidance (TD-CFG), which replaces static conditioning with an adaptive schedule: strong guidance early to establish the global binding pose, followed by gradual relaxation for precise local CDR refinement. On 55 non-redundant benchmark complexes, TiDE-Ab outperforms RFantibody with higher epitope recall (0.93 vs. 0.88) and over 95% fewer steric clashes, with TD-CFG further improving backbone designability as measured by scRMSD and iPAE. In therapeutic case studies on TGF- and IL-17A, TiDE-Ab reproduced the binding profiles of clinical antibodies across isoform-selective and cross-reactive epitopes, where RFantibody consistently failed to produce viable candidates.
Availability: Source code, data, and pre-trained models are available at https://github.com/SNU-CSSB/TiDE-Ab

• Dongjun Guo, University College London, United Kingdom
• Deborah Dunn-Walters, University of Surrey, United Kingdom
• Franca Fraternali, University College London, United Kingdom
• Joseph Ng, University College London, United Kingdom

The development of stable antibodies formed by compatible heavy (H) and light (L) chain pairs is crucial in both in vivo maturation of antibody-producing cells and ex vivo designs of therapeutic antibodies. We present ImmunoMatch, a machine-learning framework trained on paired H and L sequences from human B cells to identify molecular features underlying chain compatibility. ImmunoMatch distinguishes cognate from random H–L pairs and captures differences associated with lambda and kappa light chains, reflecting B cell selection mechanisms in the bone marrow. We apply ImmunoMatch to reconstruct paired antibodies from spatial VDJ sequencing data and study the refinement of H–L pairing across B cell maturation stages in health and disease. We find further that ImmunoMatch is sensitive to sequence differences at the H-L interface. These insights provide a computational lens into the broader biological principles governing antibody assembly and stability.

• Pengfei Zhang, School of Computing and Augmented Intelligence & Biodesign Institute, Arizona State University, United States
• Xiaoyi He, School of Computing and Augmented Intelligence & Biodesign Institute, Arizona State University, United States
• Fredo Guan, School of Computing and Augmented Intelligence & Biodesign Institute, Arizona State University, United States
• Hao Mei, School of Computing and Augmented Intelligence & Biodesign Institute, Arizona State University, United States
• Gloria Grama, Biodesign Institute, Arizona State University, United States
• Seojin Bang, Biodesign Institute, Arizona State University, United States
• Heewook Lee, School of Computing and Augmented Intelligence & Biodesign Institute, Arizona State University, United States

Epitope-conditioned T cell receptor (TCR) generation extends protein language modeling to therapeutic design of TCRs, with recent works demonstrating improved performance using reinforcement learning (RL) post-training utilizing surrogate models. However, such frameworks suffer from Goodhart's Law: optimizing imperfect surrogate rewards leads to reward hacking and biologically invalid sequences. We present a plug-and-play reward design framework that mitigates reward hacking without altering the generator and its training pipeline. The framework integrates multiple complementary strategies: heuristic biological priors to reject degenerate sequences, model ensembling to reduce per-model biases, and binding-specificity objectives to suppress cross-epitope overfitting and promote target-focused binding behavior. When applied to the RL-based fine-tuning process of TCR generators, these rewards stabilize optimization, preserve sequence diversity, and yield generations that are more biologically aligned with real TCR sequences. The resulting models produce more authentic epitope-specific receptors, demonstrating that our Goodhart-resistant reward design substantially improves reliability and controllability in biologically grounded sequence generation.

• Alif Bin Abdul Qayyum, Texas A&M University, United States
• Yingtong Zhou, Texas A&M University, United States
• Xiaoning Qian, Texas A&M University, United States
• Byung-Jun Yoon, Texas A&M University, United States

Motivation:
Protein Language Models (PLMs) have revolutionized protein engineering by capturing the evolutionary constraints inherent in natural protein sequences. However, precisely steering these models to engineer novel proteins with targeted functionalities remains challenging due to the inherent difficulty in modifying their latent representations considering the design objectives. Recently, activation steering of PLMs has emerged as a potent, training-free intervention for directing PLM outputs. However, the requirement for high-quality labeled datasets limits its application. In data-scarce or out-of-distribution (OOD) regimes, researchers must rely on surrogate models for label prediction; however, deterministic surrogates fail to account for the underlying uncertainty, often yielding steering vectors that result in suboptimal protein design.
Results:
To address this, we propose PROSOUNDS (PROtein Sequence Optimization through UNcertainty-weighteD Steering), a PLM-based protein design framework that integrates uncertainty quantification into the activation steering logic. By weighting the steering activation calculation process based on uncertainty estimates of the surrogate predictions, PROSOUNDS enables robust protein optimization through precise mutational design even in the absence of ground-truth labels. Comprehensive performance evaluation reveals that PROSOUNDS consistently outperforms deterministic alternatives across three different protein property optimization tasks.
Availability and Implementation:
The datasets and implementation code for PROSOUNDS are available at https://github.com/TeresaZhouTamu/PRO-SOUNDS.

• Mehmet Efe Akça, BoÄŸaziçi Ãœniversitesi, Turkey
• Gökçe UludoÄŸan, Bogazici University, Turkey
• Arzucan Ozgur, Bogazici University, Turkey
• Inci BaytaÅŸ, Bogazici University, Turkey

Motivation: Accurate prediction of protein function is essential for elucidating molecular mechanisms and advancing biological and therapeutic discovery. Yet experimental annotation lags far behind the rapid growth of protein sequence data. Computational approaches address this gap by associating proteins with Gene Ontology (GO) terms, which encode functional knowledge through hierarchical relations and textual definitions. However, existing models often emphasize one modality over the other, limiting their ability to generalize, particularly to unseen or newly introduced GO terms that frequently arise as the ontology evolves, and making the previously trained models outdated.
Results: We present STAR-GO, a Transformer-based framework that jointly models the semantic and structural characteristics of GO terms to enhance zero-shot protein function prediction. STAR-GO integrates textual definitions with ontology graph structure to learn unified GO representations, which are processed in hierarchical order to propagate information from general to specific terms. These representations are then aligned with protein sequence embeddings to capture sequence–function relationships. STAR-GO achieves state-of-the-art performance and superior zero-shot generalization, demonstrating the utility of integrating semantics and structure for robust and adaptable protein function prediction.
Availability: Code and pre-trained models are available at https://github.com/boun-tabi-lifelu/stargo

• Samuele Firmani, Computational Health Center, Helmholtz Center Munich, Germany
• Felix Steinbauer, Technical University of Munich, Germany
• Gjergji Kasneci, Technical University of Munich, Germany
• Marc Horlacher, Computational Health Center, Helmholtz Center Munich, Germany
• Annalisa Marsico, Computational Health Center, Helmholtz Center Munich, Germany

Motivation: RNA-binding proteins (RBPs) regulate post-transcriptional processes, yet many remain undiscovered because RNA-binding activity often occurs outside canonical RNA-binding domains (RBDs), including within intrinsically disordered regions (IDRs) or through protein complexes.
Computational methods can help identify novel RBPs, but approaches relying solely on sequence-derived features or ignoring the cellular interaction context are limited in capturing the complexity of RNA-binding behavior. To date, no framework rigorously integrates both sequence information and protein interaction context for RBP prediction. Results: We introduce RIBEX, a multimodal framework that combines protein language model (pLM) embeddings with protein interactome topology to improve RBP prediction and interpretation. Specifically, we integrate sequence representations with graph-derived positional encodings (PE) from the human STRING protein-protein interaction (PPI) network. PE are computed using Personalized PageRank, reduced with principal component analysis, and fused with pooled sequence embeddings through FiLM conditioning, while Low-Rank Adaptation (LoRA) enables parameter-efficient task adaptation. Across both an annotation-based benchmark and experimental RNA Interactome Capture (RIC) dataset, PE consistently improves predictive performance, indicating that interactome topology provides complementary information beyond sequence features. LoRA adaptation of ESM2-650M further yields larger gains than simply scaling frozen backbone size. RIBEX outperforms state-of-the-art methods such as RBP-TSTL and HydRA, particularly on challenging subsets including proteins lacking canonical RBDs and those enriched in IDRs. For interpretability, we combine sequence-level computational alanine scanning with network-level positional-encoding ablation and inverse-PCA mapping, recovering known RNA-binding domains, IDR-associated contributions, and functional interactome communities linked to RBP predictions.

• Florence Tama, RIKEN Center for Computational Science & Nagoya University, Japan

Understanding biomolecular function requires insight into both structure and dynamics. High-speed atomic force microscopy (HS-AFM) enables direct visualization of biomolecular motions at the single-molecule level under near-physiological conditions, producing two-dimensional (2D) topographical images in real time. However, its limited resolution prevents direct determination of three-dimensional (3D) atomic structures, necessitating computational approaches to bridge this gap.
We present NMFF-AFM, a flexible fitting method that integrates normal mode analysis with AFM image fitting to reconstruct 3D atomic models from HS-AFM data. The method leverages low-frequency normal modes to capture large-scale, functionally relevant conformational changes while deforming a known protein structure to match AFM images. Validation using simulated AFM data from proteins with known dynamics demonstrated robustness and reliability. To facilitate broader use, NMFF-AFM was implemented in the BioAFMviewer platform, creating a streamlined workflow from raw HS-AFM movies to atomistic modeling and visualization. This user-friendly integration enables direct analysis of experimental data across diverse systems. Applications to a single protein domain, a multi-protein complex, and a megadalton-scale filament highlight the method's versatility. Furthermore, applying NMFF-AFM to time-resolved HS-AFM data allowed reconstruction of atomistic molecular movies. Overall, NMFF-AFM and BioAFMviewer provide an efficient framework for extracting atomistic dynamics from HS-AFM data, enabling large-scale analysis and advancing mechanistic understanding of biomolecular function.

• Domenico Scaramozzino, Karolinska Institutet, Sweden
• Byung Ho Lee, Karolinska Institutet, Sweden
• Laura Orellana, Karolinska Institutet, Sweden

Protein conformational changes are the cornerstone of biological function. While conformers captured experimentally represent metastable states, the pathways connecting them have been elusive for experiments and simulations alike. Nowadays, cryogenic Electron Microscopy is providing rich structural data on proteins trapped in different states for increasingly large systems, but these are out of scope for most computational methods which exhibit an N^2 dependence on size. Based on our previous eBDIMS algorithm, here we present eBDIMS2, an optimized version with quasi-linear size dependence, able to simulate on a desktop computer particularly complex transitions for megadalton protein assemblies, like the rotary motion of ATP synthases. Not only eBDIMS2 pathways spontaneously visit experimental intermediates but also overlap with enhanced and microsecond Molecular Dynamics simulations requiring extensive supercomputing resources. By integrating Elastic Networks with Brownian Dynamics, eBDIMS2 allows an unprecedented exploration of conformational changes of sub-mesoscopic systems previously inaccessible.

• Anupam Banerjee, Laufer Center for Physical and Quantitative Biology, Stony Brook University, United States
• Anthony Bogetti, Laufer Center for Physical and Quantitative Biology, Stony Brook University, United States
• Ivet Bahar, Laufer Center for Physical and Quantitative Biology, Stony Brook University, United States

Understanding the effects of missense mutations or single amino acid variants (SAVs) on protein function is crucial for elucidating the molecular basis of diseases/disorders and designing rational therapies. We introduce here Rhapsody-2, a machine learning tool for discriminating pathogenic and neutral SAVs, significantly expanding on a precursor limited by the availability of structural data. With the advent of AlphaFold2 as a powerful tool for structure prediction, Rhapsody-2 is trained on a significantly expanded dataset of 117,525 SAVs corresponding to 12,094 human proteins reported in the ClinVar database. Adopting a broad set of descriptors composed of sequence evolutionary, structural, dynamic, and energetics features in the training algorithm, Rhapsody-2 achieved an AUROC of 0.94 in 10-fold cross-validation when all SAVs of a particular test protein (mutant) were excluded from the training set. Benchmarking against a variety of testing datasets demonstrated the high performance of Rhapsody-2. While sequence evolutionary descriptors play a dominant role in pathogenicity prediction, those based on structural dynamics provide a mechanistic interpretation. Notably, residues involved in allosteric communication and those distinguished by pronounced fluctuations in the high-frequency modes of motion or subject to spatial constraints in soft modes usually give rise to pathogenicity when mutated. Overall, Rhapsody-2 provides an efficient and transparent tool for accurately predicting the pathogenicity of SAVs and unraveling the mechanistic basis of the observed behavior, thus advancing our understanding of genotype-to-phenotype relations.

• Ute F. Röhrig, SIB Swiss Institute of Bioinformatics, Switzerland
• Marine Mathieu-Bugnon, SIB Swiss Institute of Bioinformatics, Switzerland
• Vincent Zoete, SIB Swiss Institute of Bioinformatics and UNIL University of Lausanne, Switzerland

Motivation: Molecular docking is a pillar of structure-based drug design and shows advantages in structure prediction of small-molecule ligand–protein complexes over co-folding methods for novel ligands and novel binding pockets. Here, we describe substantial improvements of our physics-based docking algorithm Attracting Cavities, which is widely used through the SwissDock webserver.
Results: AC 3.0 includes enhanced sampling features, new functionalities, and technical improvements. These lead to better sampling at lower execution times and higher versatility. Comparison with AutoDock Vina demonstrates better docking results on multiple test sets.
Availability: AC 3.0 will be made freely accessible through the SwissDock webserver (www.swissdock.ch).

• Dani Korpela, Department of Computer Science Aalto University, Broad Institute of MIT and Harvard, Finland
• Alexandru Dumitrescu, Department of Computer Science Aalto University, Broad Institute of MIT and Harvard, Finland
• Martin Strazar, National Institute of Biology Ljubljana, Broad Institute of MIT and Harvard, Slovenia
• Rui Li, Broad Institute of MIT and Harvard, United States
• Ramnik Xavier, Broad Institute of MIT and Harvard, United States
• Daniel Graham, Broad Institute of MIT and Harvard, United States
• Harri Lähdesmäki, Department of Computer Science Aalto University, Finland

Motivation: Peptide-MHC II binding drives adaptive immunity, yet discovery of novel binder peptides remains challenging due to open binding grooves of MHC-II that accommodate variable-length peptides. While discriminative models perform well, they are unfeasible for generation via enumeration due to vast peptide space (20^13 ~ 8 * 10^16 for peptides of length 13 amino acids). Generative AI approaches could accelerate binder design to enable vaccines targeted to particular MHC-II alleles or optimize other peptide chemical properties.
Results: We introduce PepGen, the first protein language model for MHC II peptide generation building on Generalized Language Modeling. PepGen conditions on alleles, arbitrary partial peptides including putative TCR- interacting motifs, and continuous binding affinity. Across multiple benchmarks including infilling and de novo generation, PepGen outperformed frequency sampling, Gibbs clustering, and autoregressive baselines. Adjusted log-probabilities enable good classification performance. Experimental validation confirmed that the SARS-CoV-2 peptide TEGALNTPKDHIGTR binding the HLA-DQA101:03-DQB106:03 allele can be redesigned to bind the HLA-DQA101:02-DQB105:02 allele. PepGen generated three TCR-motif-preserving binders gaining up to 70% of original MFI. Overall, PepGen provides scalable, motif-constrained MHC II peptide re-design and de novo generation, validated through thorough benchmarks and functional assays.
Availability and implementation: Code and data is available at https://github.com/DaniTheOrange/PepGen

• Kai Chen, School of Computer Science and Engineering, Central South University, China
• Zhijian Huang, School of Computer Science and Engineering, Central South University, China
• Yinbo Wang, School of Computer Science and Engineering, Central South University, China
• Siyuan Shen, School of Computer Science and Engineering, Central South University, China
• Jinmiao Song, Xinjiang Key Laboratory of Intelligent Computing and Smart Applications, Xinjiang University, China
• Lei Deng, School of Computer Science and Engineering, Central South University, China

Motivation: Proteolysis-targeting chimeras (PROTACs) enable targeted protein degradation by recruiting an E3 ubiquitin ligase to a protein of interest (POI) and forming a ternary complex. Despite their therapeutic promise, rational PROTAC design remains challenging, as degradation efficacy depends on subtle and highly structure-dependent interactions among the POI, the E3 ligase, and the bifunctional molecule.
Results: We propose Pocket-PROTACs, a pocket-aware attention-based framework for predicting PROTAC-induced protein degradation from a triplet of POI, E3 ligase, and PROTAC. Pocket-PROTACs encodes protein sequences using a pre-trained protein language model and represents PROTACs with a geometry-aware graph neural network over an ensemble of three-dimensional conformers. Both POI–PROTAC and E3 ligase–PROTAC interactions are explicitly modeled through a residue–atom cross-attention mechanism that captures fine-grained interaction patterns. To improve model interpretability, we introduce a pocket-aware module that incorporates structural context to guide residue-level relevance estimation, enabling multi-level attribution analysis. Experiments on two benchmark datasets show that Pocket-PROTACs consistently outperforms fingerprint-based baselines and recent deep learning methods. The learned relevance maps highlight localized interaction patterns on both the POI and the E3 ligase that are qualitatively consistent with known pocket-level features. A case study on KLHDC2-engaging BET PROTACs further demonstrates that our model accurately predicts degradation behavior and provides biologically meaningful, attention-based interpretations, offering practical support for PROTAC design and experimental investigation.
Availability: Source code and datasets are available at https://github.com/Adochew/Pocket-PROTACs.
Contact: leideng@csu.edu.cn
Supplementary information: Supplementary data are available at Bioinformatics online.