Weekly BioML Digest [September 21, 2026]

Machine Learning × Computational Biology paper compilation

Hey! It's your weekly digest of machine learning papers in CompBio and Drug Discovery.

Feedback? Email me at biomldigest@gmail.com.

📚 Peer-Reviewed Journals (Top 20)

1282 matched filters -> 20 selected after LLM relevance + novelty ranking.

• 🔬 Predicting cellular responses to perturbation across diverse contexts with State.
  Abhinav K Adduri, Dhruv Gautam, Beatrice Bevilacqua, Mohsen Naghipourfar, ..., Alexander Dobin, Dave P Burke, Hani Goodarzi, Yusuf H Roohani — Cell, 2026-09-17
  Introduces State, a model trained on 167M single-cell profiles to predict genetic, signaling, and chemical perturbation effects across cellular contexts, enabling generalization to unseen cell states and robust DE gene identification. Provides Cell-Eval for standardized benchmarking of perturbation-effect models.
  Affiliations: Arc Institute, Palo Alto; Arc Institute, Palo Alto; ...

• 📡 Deep learning coupled with scalable domain-specific structural validation expands RNA virus discovery from metatranscriptomes
  Luo, Gaoyang, Zang, Zelin, ..., Li, Stan Z., Ju, Feng — Nature Communications, 2026-09-16
  Presents Rider, a two-stage RNA virome discovery pipeline that combines compact protein language models with sliding-window structure prediction and RdRp domain validation to detect highly divergent RNA viruses from >10,000 metatranscriptomes. Recovers >99% of known viruses and discovers novel clades, including in IBD cohorts.

• 📰 The Virtual Biotech: A multi-agent AI framework for therapeutic discovery and development.
  Harrison G Zhang, Peter Eckmann, Jiacheng Miao, Andrew B Mahon, James Zou — Science (New York, N.Y.), 2026-09-17
  Demonstrates a multi-agent AI 'Virtual Biotech' that integrates evidence across target discovery, safety, modality selection, and clinical development to inform drug-development decisions. Validates findings by large-scale trial annotation, multimodal integration for therapy proposals, and post-hoc failure analysis.
  Affiliations: Department of Biomedical Data Science, Stanford University; Department of Computer Science, Stanford University; ...

• 🔬 4D spatiotemporal landscape of mitochondrial phenotypes across cellular states unlocked through representation learning.
  Dhruv Agarwal, Zichen Wang, Eric Arkfeld, Andre Modolo, ..., Gillian McMahon, Siddharth Nahar, Manav Doshi, Johannes Schöneberg — Cell, 2026-09-17
  MitoSpace uses self-supervised learning on terabytes of 4D lattice light-sheet microscopy to learn representations capturing mitochondrial morphology-dynamics and drug responses, achieving R2=0.91 for membrane potential prediction. Shows zero-shot generalization to unseen perturbations and organoids, enabling 4D phenotypic screening.
  Affiliations: Department of Pharmacology, University of California; Department of Biochemistry and Molecular Biophysics, University of California; ...

• 📰 Evaluating SARS-CoV-2 antibody resilience via prediction and design of escape viral variants.
  Marian Huot, Pierre Rosenbaum, Cyril Planchais, Hugo Mouquet, Rémi Monasson, Simona Cocco — Cell systems, 2026-09-16
  EscapeMap integrates deep mutational scanning and a coronavirus generative sequence prior to design antibody-escape RBD variants under ACE2 viability constraints. Validates predicted escape routes and identifies antibody combinations resilient to simultaneous escape.
  Affiliations: Laboratory of Physics of the École Normale Supérieure, CNRS UMR 8023 and PSL Research; Institut Pasteur, Université Paris Cité; ...

• 📰 Mapping the combinatorial coding between olfactory receptors and perception with deep learning.
  Seyone Chithrananda, Judith Amores, Kevin K Yang — Cell systems, 2026-09-16
  MolOR uses cross-attention between a molecular GNN and protein language model embeddings to predict odorant–olfactory receptor activation profiles and link them to human odor percepts. Enables discovery of ligands for orphan ORs and design of odorants with desired qualities.
  Affiliations: Microsoft Research, Cambridge; Stanford University, Stanford; ...

• 📰 Multiobjective learning and design of bacteriophage specificity.
  Naia Novy, Phil Huss, Sarah Evert, Philip A Romero, Srivatsan Raman — Cell systems, 2026-09-16
  Applies multiobjective deep learning to engineer T7 phage receptor-binding proteins with predefined host specificity and broad generality. Achieves high experimental hit rates across 26 tasks, revealing that opposite specificities can be separated by few mutations.
  Affiliations: University of Wisconsin-Madison, Biochemistry; University of Wisconsin-Madison, Biochemistry; ...

• 🔭 Genolator enables protein function interpretation using a multimodal large language model fusing genomic and structural interpretation with natural language interaction
  Danner, Martin, Islam, Tanhim, ..., Kurth, Ingo, Krause, Jeremias — Genome Biology, 2026-09-16
  Genolator is a multimodal LLM that fuses DNA/protein sequence and structural embeddings with natural language queries, fine-tuned on GO-derived Q&A to answer protein function, localization, and process questions. Outperforms general LLMs and smaller domain models, with interpretable internal representations.

• 📡 Learning chemical-induced gene expression perturbations with WAVE
  Lv, Tianhang, Chen, Bojin, ..., Liao, Jie, Fan, Xiaohui — Nature Communications, 2026-09-18
  WAVE (β-VAE) predicts chemical-induced bulk and single-cell gene expression from molecular structure and cellular basal states, enabling in silico perturbation screens. Demonstrates utility for drug discovery and repurposing, including LUAD candidate prioritization with experimental validation.

• 📡 scTransMIL bridges patient-level disease states and single-cell transcriptomics for cancer screening and heterogeneity inference
  Tang, Zhenchao, Wang, Fang, ..., Chen, Calvin Yu-Chian, Yao, Jianhua — Nature Communications, 2026-09-14
  scTransMIL links patient-level cancer labels to single-cell transcriptomes via transformer-based multi-instance learning, accurately classifying cancer states and tissue-of-origin while identifying tumor-associated cell populations. Attention maps enable full-transcriptome biomarker discovery from minimally annotated data.

• 📰 SpaMOAL is a deep learning method that enables accurate spatial domain identification from multi-omics data.
  Jin-Xian Wang, Yu-Ying Huo, Rui Zhao, Yan Pan, Jianqiang Wu, Han Wang, Xiang-Yu Li — PLoS biology, 2026-09-16
  SpaMOAL performs graph contrastive learning that integrates spatial coordinates, histology, and multi-omics to learn clustering-friendly embeddings for spatial domain identification. Consistently outperforms existing methods across paired spatial multi-omics datasets.

• 💻 scBalFlow: A Staged Flow Matching Framework for Imbalanced Single-Cell Drug Perturbation Prediction.
  H. Lyu, Jia-Wei Luo — Bioinformatics, 2026-09-16
  scBalFlow is a two-stage framework that first predicts perturbation response intensity (with Gaussian-augmented inference for imbalance), then uses Flow Matching to synthesize strongly responsive single-cell profiles. Significantly improves prediction under severe class imbalance on SciPlex and McFarland benchmarks.

• 📰 Accurate RNA-Ligand Binding Site Prediction Based on a Multi-Channel Graph Neural Network
  Li, Na, Niu, Jingran, ..., Wei, Dongqing, Man, Rongjun — Interdisciplinary Sciences: Computational Life Sciences, 2026-09-19
  BC-GNN is a multi-channel graph neural net with boundary-aware propagation and microenvironment-aware recalibration for nucleotide-level RNA–ligand binding site prediction. Outperforms re-evaluated baselines on RNAmigos2 with leakage-controlled splits.

• 💻 PatchEpi: Patch-Aware Equivariant Learning Improves Structure-Based Epitope Prediction.
  Si-Cheng Wen, Fei Li, Yue Qian — Bioinformatics, 2026-09-16
  PatchEpi reframes B-cell epitope prediction as surface-patch learning with equivariant geometric networks and boundary-contrastive objectives, aligning learned patches with true antibody footprints. Achieves SOTA performance on homology-controlled splits.

• 📰 HelixDTA: Dual-Branch Sequence–Structure Learning with Complete Target Structures for Robust and Interpretable Drug–Target Affinity Prediction
  Shang Lou, Xu-Hua Li, Yu-Jie Peng, Yuan Yuan, ..., Hong Qian, Tao Ren, Hong-Cang Gu, Fan Zhang — Journal of Chemical Information
  and Modeling, 2026-09-16
  HelixDTA uses a dual-branch architecture to jointly learn from sequence context and complete target structures for drug–target affinity prediction, improving cold-start generalization and interpretability. Demonstrates structure-aware candidate prioritization in a WRN case study.

• 📰 Protein function-conditioned language models for variant effect prediction and controllable design.
  Shaowen Zhu, Yue Cao, Yihong Yang, Yang Shen — Biophysical journal, 2026-09-15
  Func2Seq/Func2Prot conditions autoregressive protein language modeling on GO-derived function embeddings to score variant effects and generate function-constrained sequences and backbones. Improves variant prediction, especially for small/low-diversity families, and preserves functional sites in controllable design.
  Affiliations: Department of Electrical and Computer Engineering, Texas A&M University; Department of Electrical and Computer Engineering, Texas A&M University; ...

• 📰 Training a force field for proteins and small molecules from scratch.
  Alexandre Blanco-González, Thea K Schulze, Evianne Rovers, Joe G Greener — Chemical science, 2026-09-17
  Garnet is a GNN-based, continuously typed force-field learner trained from quantum, condensed-phase, and protein NMR data that assigns all FF parameters without legacy priors. Matches modern FFs across small molecules and proteins and supports alternative potentials (e.g., double exponential).
  Affiliations: Medical Research Council Laboratory of Molecular Biology Cambridge CB2 0QH UK jgreener@mrclmb.ac.uk.

• 📰 DASH: A Pocket-Aware and Objective-Aware Framework for Million-Scale Structure-Based Molecular Generation.
  Baohua Zhang, Huangchao Xu, Xiaoning Wang, Longfei Li, Zhong Jin — Journal of chemical information and modeling, 2026-09-14
  DASH converts protein-conditioned diffusion outputs into million-scale, objective-aware molecular libraries via pocket-complexity-adaptive sampling and configurable desirability scoring. Demonstrates scalable production with annotated SDFs and EGFR-oriented computational prioritization.
  Affiliations: Computer Network Information Center, Chinese Academy of Sciences ; School of Computer Science and Technology, University of Chinese Academy of Sciences; ...

• 📰 LinkLlama: Enabling a Large Language Model for Chemically Reasonable Linker Design.
  Kunyang Sun, Yingze Wang, Justin Purnomo, Joseph M Cavanagh, Giovanni Battista Alteri, Teresa Head-Gordon — Journal of chemical information and modeling, 2026-09-14
  LinkLlama fine-tunes Meta Llama 3 to generate chemically reasonable fragment linkers from natural-language constraints (geometry, Lipinski, rotatable bonds), doubling the fraction of valid, drug-like designs. Validated via docking/MD for scaffold hopping and PROTAC linkers.
  Affiliations: Kenneth S. Pitzer Theory Center and Department of Chemistry, University of California; Department of Bioengineering, University of California; ...

• 📰 PockLigGPT: Pocket-Sequence-Conditioned Molecular Generation with GPTs and RL.
  Pablo Varas Pardo, Guillermo Marcos-Ayuso, Eugenia Ulzurrun, David Ríos Insua, Nuria E Campillo — Journal of chemical information and modeling, 2026-09-14
  PockLigGPT conditions GPT-based molecular generation on pocket amino acid sequences and refines with docking-guided RL, balancing chemical plausibility and target specificity. Achieves competitive docking while maintaining favorable drug-like profiles; AD case studies support utility.
  Affiliations: AItenea Biotech S.L. , Madrid28014; Instituto de Ciencias Matemáticas (ICMAT-CSIC) , Madrid28049; ...

🧬 Preprints (arXiv + bioRxiv)

62 matched filters -> 20 selected after LLM relevance + novelty ranking.

• 🧬 Latent generative search unlocks de novo design of untapped biomolecular interactions at scale
  Didi, K.; Reidenbach, D.; Penner, M.; Ravichandran, S.; ...; Kucukbenli, E.; Vahdat, A.; Ogden, P.; Kreis, K. — bioRxiv, 2026-09-18
  Introduces latent generative search that steers a sequence-structure co-design model at inference to generate high-affinity de novo protein binders, including first-in-class binders to free carbohydrates, validated by million-scale phage display screens.
  Affiliations: NVIDIA Corporation, Santa Clara

• 📄 TorchCraft: Unified binder design by inverting an all-atom structure predictor
  TorchCraft Team, Yu Liu, Zhouhanyu Shen, Zhengyi Li, ..., Qilin Yu, Xiayan Qin, Yucheng Zhang, Mingchen Chen — arXiv, 2026-09-17
  Inverts a frozen all-atom predictor (AlphaFold 3 via TorchFold) to directly optimize sequence logits for unified binder design, experimentally yielding binding minibinders, framework-conditioned VHHs, cyclic peptides, and ligand-binding proteins without post hoc redesign.

• 🧬 Ourotide: decoding the hierarchical peptide recognition for generative design
  Shen, Y.; Zhang, J.; Wu, Z.; Xing, Z.; ...; Zhou, Q.; Han, F.; Jiang, N.; Chen, X. — bioRxiv, 2026-09-16
  Maps a previously unrecognized hierarchical peptide-binding mode and trains Ourotide with physics-guided data mining and conditional geometric flow matching to improve peptide backbone accuracy, interface recovery, and affinity prediction up to 65 residues.
  Affiliations: Department of Medicinal Chemistry, School of Pharmacy

• 📄 Ensemble-Conditioned Molecular Design
  Ross Irwin, Alessandro Tibo, Jon Paul Janet, Simon Olsson — arXiv, 2026-09-14
  Reframes 3D molecular design as conditioning on conformational ensembles and aggregate properties, composing vector fields at inference to enable multi-mode and state-selective ligand generation, improving dual-target and active-state-selective designs.

• 🧬 Genome-scale perturbation signatures from primary human CD4+ T cells improve genetics-based prioritization of immune drug targets
  Han, A. L.; Slotnik, M.; Fox, D. A.; Dhindsa, R. S.; Gudjonsson, J. E.; Kahlenberg, J. M.; Welch, J. — bioRxiv, 2026-09-16
  IGNITE integrates human genetics with genome-scale perturb-seq in primary CD4+ T cells using semi-supervised ML to prioritize immune drug targets, substantially enriching in-trial targets and elevating tractable candidates such as ELOVL6 and RUVBL1.
  Affiliations: University of Michigan Medical School, Ann Arbor

• 🧬 miRstring: An RNA language model enables mature miRNA decoding and artificial small RNA design across species
  Peng, R.; Li, X.; fang, t.; Yu, X. — bioRxiv, 2026-09-20
  miRstring, an RNA language model, decodes mature miRNA duplex boundaries from precursors across 414 species and guides artificial miRNA scaffold design, with attention aligning to endonuclease cleavage sites and experimental repression validation.
  Affiliations: Shanghai Jiao Tong Unviversity

• 🧬 A diffusion model of viral evolution predicts mutation fitness and evolutionary trajectories
  Wu, J.; Ding, X.; Wu, A. — bioRxiv, 2026-09-18
  Models viral evolution with a diffusion process where forward noise mimics mutation and reverse denoising captures selection, predicting mutational fitness and future trajectories from sequences alone and outperforming state-of-the-art generative baselines.
  Affiliations: Suzhou Institute of Systems Medicine, Chinese Academy of Medical Sciences & Peking Union Medical College

• 🧬 Physical priors improve performance of structure-based binding affinity models
  Kaminow, B.; Payne, A. M.; MacDermott-Opeskin, H. I.; Chodera, J. D.; Singh, S. — bioRxiv, 2026-09-18
  Shows that physics-informed priors and E(3) biases in decomposed 3D structure-based affinity models (mtenn) improve accuracy and generalization over ligand-only baselines, matching industry standards and excelling on unseen targets like COVID Moonshot.
  Affiliations: Memorial Sloan Kettering Cancer Center

• 📄 Search at the Cost of Sampling: Nearly-Instant Latent Space Bayesian Optimization
  Donney Fan, Colin Doumont, Aleksandra Kalisz, Paul Duckworth, Jacob R. Gardner, Henry Moss, Geoff Pleiss — arXiv, 2026-09-16
  Proposes a nearly closed-form latent-space Bayesian optimization with linear surrogates on spherical domains, delivering 100x speedups while matching or improving molecular generation performance in de novo design pipelines.

• 🧬 Machine learned potentials with electrostatic embedding accurately capture Kemp eliminase reactivity
  Lear, A.; Chan, E. W.; Zinovjev, K.; van der Kamp, M. W.; Bunzel, H. A.; Mulholland, A. J. — bioRxiv, 2026-09-16
  Trains electrostatic ML/MM-embedded machine-learned potentials (MACE+EMLE) to compute enzyme reaction barriers for Kemp eliminase variants, recovering experimentally observed barrier differences with QM-level accuracy at a fraction of the cost.
  Affiliations: University of Bristol

• 🧬 Lacuna: Cryptic Binding Pocket Discovery via Conformational Ensemble Analysis
  Moore, C. — bioRxiv, 2026-09-17
  Lacuna discovers cryptic binding pockets by generating conformational ensembles and clustering pocket detections across states, with optional PLM-assisted ranking and learned surface detectors, achieving top recovery on CryptoBench in seconds per chain.
  Affiliations: Texas A and M University

• 📄 EnSol: an environment-aware graph neural network for molecular solubility prediction
  Thao Nguyen, Saman Shafaei, Zhengyi Zhang, Huimin Zhao, Heng Ji — arXiv, 2026-09-17
  EnSol is an environment-aware probabilistic GNN that couples solute and solvent graphs via cross-attention, modulates solvent features by temperature, and predicts full solubility distributions with a mixture density head, outperforming SOTA with experimental validation.

• 📄 GPCR Ligand Bioactivity Prediction with Physics-Informed Dual-State Query Learning
  Shuo Zhang, Huifeng Zhang, Rongqi Hong, Jian K. Liu — arXiv, 2026-09-15
  DSQ embeds the Monod-Wyman-Changeux allosteric model within a neural network using disentangled active/inactive latent queries and a neural MWC gate, improving GPCR efficacy-aware bioactivity prediction, especially for agonists.

• 📄 When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows
  Gabriel Bénédict, Melanie Buechler, Gerard Riera-Solà, Chloé de Ancos, Yves Gaetan Nana Teukam, Moritz Freidank — arXiv, 2026-09-16
  EditJumps provides the first open implementation of continuous-time edit-based generative models for antibody optimization, unifying prior methods and revealing critical hyperparameter and evaluation sensitivities in generative antibody design.

• 🧬 AI-Powered Discovery of Novel RNA Viruses from the Permafrost of a 14,300-Year-Old Pleistocene Wolf
  Zaheer ud Din, S.; Wu, Q. — bioRxiv, 2026-09-18
  Combines protein language model homology search with AlphaFold2 structural validation to mine ancient metatranscriptomes, discovering two deeply divergent RNA mycoviruses preserved in a 14,300-year-old wolf specimen.

• 🧬 MMAD-Risk: Multivariate Mixed Survival Analysis for the Prediction of Age-Dependent Disease Risks from Plasma Proteomes
  Hilger, A. M.; Soeding, J. — bioRxiv, 2026-09-17
  MMAD-Risk introduces a scalable multivariate mixed accelerated failure-time model trained with amortized variational inference on 3,000 plasma proteins to predict age-dependent risks for 271 diseases, outperforming Cox models and compressing to a 10-protein panel.
  Affiliations: Max Planck Institute for Multidisciplinary Sciences

• 📄 Decoding Extrahepatic Targeting of Lipid Nanoparticles with Interpretable Machine Learning
  Asal Mehradfar, Mohammad Shahab Sepehri, Owen Antholine, Varun Shankar, Glen S. Kwon, Salman Avestimehr, Morteza Rasoulianboroujeni — arXiv, 2026-09-15
  Builds an interpretable ML framework using RDKit descriptors and formulation variables to predict hepatic vs extrahepatic LNP accumulation, with SHAP analyses identifying chemical and compositional design rules for extrahepatic RNA delivery.

• 🧬 Integrating complementary biological information for multi-objective enzyme engineering
  Blalock, N.; Sosa, Y.; Heuschkel, J.; Li, R.; ...; Song, J.; Pefaur, N.; Kingsley, L. J.; Romero, P. A. — bioRxiv, 2026-09-17
  Develops an ML-guided multi-objective enzyme engineering framework that integrates sparse functional data with evolutionary and structural priors to co-optimize Gre2 catalytic performance, protein yield, and thermal stability under process-relevant conditions.
  Affiliations: Duke University

• 🧬 TCRdenoise - an unsupervised similarity-based approach for denoising of TCR-pMHC specificity data
  Lund, J. M.; Deleuran, S. N.; Nielsen, M. — bioRxiv, 2026-09-15
  TCRdenoise applies unsupervised sequence-similarity clustering with a modified silhouette objective to denoise TCR–pMHC datasets, boosting NetTCR performance and corroborating labels with AlphaFold 3 interface confidence.
  Affiliations: Technical University of Denmark

• 📄 Multi-View Molecular Representation Learning with Hierarchical Graphs and Contextualized Fingerprints
  Gwang-Hyeon Yun, Jong-Hoon Park, Bing Hu, Helen Chen, Anita Layton, Young-Rae Cho — arXiv, 2026-09-14
  HiFi-Mol pretrains a fragment-aware hierarchical graph encoder and a contextualized fingerprint encoder, then fuses them to improve scaffold-split molecular property prediction across MoleculeNet tasks.