Weekly BioML Digest [September 16, 2026]

Machine Learning × Computational Biology paper compilation

Sorry for the late post this week, I had a battle with arXiv API.

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)

1198 matched filters -> 20 selected after LLM relevance + novelty ranking.

• 🌿 An operational perturbation proteomics-based virtual cell model
  Sun, Rui, Qian, Liujia, ..., Zhu, Yi, Guo, Tiannan — Nature, 2026-09-09
  ProteinTalks pretrains on 38M time-resolved proteomics measurements to learn transferable dynamical cell representations, enabling drug efficacy/synergy prediction, resistance probing, and patient stratification across cell lines and organoids.

• 🌿 Breaking timescales with generative sampling of conformational transitions
  Tang, Chenyu, Pandey, Mayank Prakash, ..., Dehez, François, Chipot, Christophe — Nature, 2026-09-09
  Gen-COMPAS couples a denoising diffusion model with committor-based filtering to reconstruct biomolecular transition pathways and free-energy landscapes from endpoints, bypassing predefined collective variables.

• 🔬 4D spatiotemporal landscape of mitochondrial phenotypes across cellular states unlocked through representation learning.
  Dhruv Agarwal, Zi-Chen Wang, Eric Arkfeld, Andre Modolo, ..., Gillian McMahon, Siddharth Nahar, Manav Doshi, Johannes Schöneberg — Cell, 2026-09-10
  MitoSpace is a self-supervised model trained on terabytes of 4D light-sheet data that learns interpretable mitochondrial morphology–dynamics embeddings, predicting membrane potential (R2=0.91) and generalizing zero-shot to unseen perturbations and organoids.

• 📰 PlantCAD2: A DNA foundation model for interpreting genomes across flowering plants.
  Jingjing Zhai, Aaron Gokaslan, Sheng-Kai Hsu, Szu-Ping Chen, ..., M Cinta Romay, Matt Pennell, Volodymyr Kuleshov, Edward S Buckler — Cell genomics, 2026-09-09
  PlantCAD2 (676M params) extends-context DNA modeling to 65 flowering plants, outperforming larger LMs on conservation and distal regulatory prediction, and fine‑tuning to cross‑species tasks.
  Affiliations: Institute for Genomic Diversity, Cornell University; Department of Computer Science, Cornell University; ...

• 📡 Functional alignment of protein language models via reinforcement learning
  Blalock, Nathaniel, Seshadri, Srinath, ..., Kulkarni, Ameya, Romero, Philip A. — Nature Communications, 2026-09-12
  RLXF aligns protein language models with functional objectives using reinforcement learning from experimental feedback, generating higher-functioning protein variants beyond evolutionary baselines across multiple protein families.

• 🏛️ Systematic discovery of circular permutations across the protein universe using CIRPIN.
  Aiden R Kolodziej, S Mazdak Abulnaga, Sergey Ovchinnikov — Proceedings of the National Academy of Sciences of the United States of America, 2026-09-08
  CIRPIN discovers circular-permutation relationships across 845 CATH topologies, enabling evolutionary insights into topological rearrangements such as multi‑form PDZ domains.
  Affiliations: Department of Biology, Massachusetts Institute of Technology; Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology; ...

• 📰 BeitAI-pHLA: Multiallele Peptide-HLA Class I Binding Prediction Using Protein Language Model and Multi-Instance Learning.
  Shigang Qiu, Yong Sun, Xiaofei Ye — Computational and structural biotechnology journal, 2026-01-01
  BeitAI-pHLA integrates protein language model embeddings with attention-based multi-instance learning to predict multiallele peptide–HLA class I binding and deconvolve motifs, achieving strong external AUPRC and neoepitope prioritization.
  Affiliations: Kindstar Biotech, Wuhan 430000; Kindstar Global Precision Medicine Institute, Wuhan 430000; ...

• 📰 ProRB: a structure-free unified framework for joint prediction and design of protein-RNA interactions.
  Yiming Xue, Xiaojian Liu, Weimin Zhu, Shengfan Wang, Hong-Bin Shen, Xiaoyong Pan — Nucleic acids research, 2026-09-07
  ProRB unifies sequence-only prediction of protein–RNA affinity, bidirectional interface contacts, and protein-binding RNA design via adaptive cross-modal attention over language-model embeddings, revealing interpretable motif-centric binding logic.
  Affiliations: Institute of Image Processing and Pattern Recognition, Shanghai Jiao Tong University

• 📰 HighMorph: De Novo Cyclic Peptide Sequence Design via Protein-Protein Interaction Recapitulation.
  Minhui Lan, Chengyun Zhang, Wentong Wang, Haomeng Hu, Huitian Lin, Sen Cao, Jingjing Guo, Hongliang Duan — Journal of medicinal chemistry, 2026-09-10
  HighMorph combines Transformer-guided Monte Carlo tree search with atomic H-bond constraints from protein–protein complexes to design target-binding cyclic peptides, yielding micromolar actives against PD-L1 and KLK4.
  Affiliations: College of Pharmaceutical Sciences, Zhejiang University of Technology; Faculty of Applied Sciences, Macao Polytechnic University; ...

• 📰 CoCoBind: Consistency-Contrastive Multitask Learning for RNA-Ligand Interaction and Binding Site Prediction.
  Shihang Wang, Lin Wang, Wei Zhao, Yuanlin He, ..., Lin Huang, Huanxiang Liu, Yang Zhang, Xiaojun Yao — Journal of medicinal chemistry, 2026-09-10
  CoCoBind jointly predicts RNA–compound interactions and nucleotide-level binding sites using cross-modal attention with consistency-contrastive learning, improving site-aware recognition under distribution shift and localizing ligand-proximal pockets.
  Affiliations: Faculty of Applied Sciences, Macao Polytechnic University; Institute of Systems Medicine, Chinese Academy of Medical Sciences ; ...

• 📰 DMGRN: Enhancing Diffusion Models for Gene Regulatory Network Inference.
  Rongyuan Li, Jingli Wu, Chunfeng Chen, Gaoshi Li, ..., Junbo Xuan, Jinlu Liu, Zheng Deng, Daoqing Gong — IEEE transactions on computational biology and bioinformatics, 2026-09-10
  DMGRN casts single-cell GRN inference as denoising diffusion with a structural equation prior and gene-similarity alignment loss, improving early-precision on BEELINE benchmarks across noisy, sparse scRNA-seq data.

• 📰 ScGraphTrans: Pathway-Guided Graph Learning and Domain Adaptation for Cell Type Annotation in Single-Cell RNA-seq.
  Yue-Chao Li, Hai-Ru You, Meng-Meng Wei, Xin-Fei Wang, Yu Li, Zhi-An Huang, Yu-An Huang, Zhu-Hong You — IEEE transactions on computational biology and bioinformatics, 2026-09-10
  scGraphTrans integrates pathway pseudo-labels, graph structure learning, and domain adaptation to enhance cell-type annotation and infer disease-relevant ligand–receptor interactions from scRNA-seq without spatial priors.

• 📰 A Site‐Aware Representation Learning Framework For Unified Molecular Interaction Modeling and Generative Design
  Gang Luo, Qian-Qian Zhang, Chen-Hao Wang, Zhi-Heng Yi, Alex Jinpeng Wang, Jing Tang, Min Li — Advanced Science, 2026-09-11
  MolDBG is a unified, site-aware framework that couples binding-site identification with drug–target affinity prediction and affinity-conditioned molecular generation, enabling interpretable, pocket-specific design even for dynamic targets.

• 📰 Pharmacophore model guided 3D molecular generation through diffusion model
  Li, Bohao, Wu, Xinyu, ..., Shang, Jinsai, Chen, Hongming — Journal of Cheminformatics, 2026-09-09
  DiffPharm injects pharmacophore-graph control into a pretrained 3D diffusion prior (Graph ControlNet) to generate molecules that satisfy user-specified pharmacophore/substructure constraints, yielding an EGFR-mutant inhibitor.

• 📰 Generative Active Learning for Molecular Design with REINVENT: Balancing Binding Affinity and Synthetic Accessibility
  Marco Klähn, Hannes H. Loeffler, S. Wan, Alexey Voronov, Xi-Bei Zhang, Agastya P. Bhati, P. Coveney — Journal of Chemical Theory
  and Computation, 2026-09-08
  A generative active learning loop integrates REINVENT with physics-based free-energy (ESMACS) and synthesizability objectives to co-optimize binding affinity and synthetic accessibility for practical molecular design.

• 📰 Hypothesis-and-Refinement Learning of Organic Structures From Multimodal Spectroscopic Data.
  Chengchun Liu, Zhiyuan Yan, Li Yuan, Hao Li, Boxuan Zhao, Yonghong Tian, Bartosz A Grzybowski, Fanyang Mo — Angewandte Chemie (International ed. in English), 2026-09-11
  SpectroMol and MS‑Mol2Mol form a hypothesis–refinement system that proposes molecules from multimodal NMR signals and mass-constrained generation trained on 400M molecules, achieving high top‑1 accuracy and improving experimental structure elucidation.
  Affiliations: School of AI for Science, Peking University Shenzhen Graduate School; State Key Laboratory of Advanced Waterproof Materials, School of Materials Science and Engineering; ...

• 🖥️ MutexaGPT: an intuition-to-design translator for physics-based enzyme engineering
  Shao, Qianzhen, Zhong, Yinjie, ..., Xu, Han, Yang, Zhongyue J. — Nature Computational Science, 2026-09-10
  MutexaGPT is a multi-agent LLM platform that translates natural-language enzyme-engineering intuition into automated physics-based modeling workflows, delivering experimentally validated specificity and cold-activity improvements.

• 📰 Machine-Learning-Enabled Rapid Evolution of Photoenzymes for the Asymmetric Synthesis of gem-Difluorophosphonates.
  Hongkui Wang, Jiafan Xu, Jiahai Zhou, Yang Gu — Angewandte Chemie (International ed. in English), 2026-09-07
  An active-learning strategy combining focused mutagenesis with a protein LM (EVOLVEpro) rapidly evolves photoenzymes, discovering variants for gem‑difluorophosphonate synthesis with >99% yield and 98:2 e.r. from only 40 screens.
  Affiliations: Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences; University of Chinese Academy of Sciences, Beijing; ...

• 📰 UFold-X: an enhanced Dual & Dynamic U-Mamba model for long-range RNA secondary structure prediction
  Lai-Yi Fu, Jia-Chun Li, Rui-Qi Wang, He-Quan Sun, Dan-Yang Wu — Nucleic Acids Research, 2026-09-12
  UFold‑X fuses CNNs with a Mamba state-space branch and dynamic gating to model long-range RNA base pairing at scale, outperforming DL baselines on long RNAs and aligning with SHAPE reactivity via an auxiliary variant.

• 📰 Systematic benchmarking of AlphaFold and SWISS-MODEL kinase structures for structure-based drug discovery.
  Erick Bahena-Culhuac, Martiniano Bello — Journal of molecular graphics & modelling, 2026-09-11
  A systematic benchmark of AlphaFold2 and SWISS‑MODEL for kinases shows raw predicted structures can support docking, but dynamic validation (post-docking MD) is essential to reduce false positives in drug discovery.
  Affiliations: Laboratorio de Diseño y Desarrollo de Nuevos Fármacos e Innovación Biotecnológica, Sección de Estudios de Posgrado e Investigación; Laboratorio de Diseño y Desarrollo de Nuevos Fármacos e Innovación Biotecnológica, Sección de Estudios de Posgrado e Investigación

🧬 Preprints (arXiv + bioRxiv)

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

• 🧬 Learning Universal Representations of Intermolecular Interactions with ATOMICA
  Fang, A.; Desgagne, M.; Zhang, Z.; Zhou, A.; Loscalzo, J.; Pentelute, B. L.; Zitnik, M. — bioRxiv, 2026-09-07
  Introduces ATOMICA, an interaction-centered geometric deep model pretrained on 2M interfaces to learn transferable atom-to-interface representations across proteins, ligands, ions, and nucleic acids. It enables zero-shot pocket–ligand tasks and predicts cofactors in dark proteome with experimental heme validation.
  Affiliations: Harvard University

• 🧬 PANDA: Protein All-atom Nested-tree Denoising Architecture for End-to-End Generation
  Bai, J.; Jiang, H. — bioRxiv, 2026-09-13
  Presents PANDA, an all-atom denoising generative model that jointly samples sequence and coordinates in Cartesian space with coupled global/local tracks. Conditioning on motif distances boosts functional scaffolding, achieving state-of-the-art all-atom protein design success.
  Affiliations: College of Biotechnology, Tianjin University of Science and Technology

• 🧬 HyphAeon: Attention on Evolution Across Deep Time Transforms Comparative Genomics
  Kosakovsky Pond, S. L.; Weaver, S.; Callan, D.; Zehr, J. D.; ...; Clark, N. L.; Makova, K. D.; Martin, D. P.; Nekrutenko, A. — bioRxiv, 2026-09-10
  HyphAeon is a phylogeny-informed transformer that amortizes episodic selection detection at genome scale, matching maximum-likelihood test accuracy >1000× faster. Its latent geometry enables artifact correction, epistatic sector discovery, and temporal sweep tracking.
  Affiliations: Temple University

• 🧬 Generative Language Modeling for Antibody CDR Grafting and Alignment-driven De Novo Design
  Gonzalez Hernandez, F.; Turnbull, O. M.; Sultana, M.; Roldan-Martin, L.; ...; Diethe, T.; Croasdale-Wood, R.; Deane, C.; Oglic, D. — bioRxiv, 2026-09-10
  GenCDR frames antibody frameworks as prompts and jointly generates all CDRs as responses with LLaMA-based autoregressive models, enabling clean reward attribution and alignment. It attains top CDR recovery and steers de novo designs toward binding and developability.
  Affiliations: AstraZeneca

• 🧬 Inverse FoldDir: Structure-conditioned Protein Sequence Design by Dirichlet Flow Matching
  TARTICI, A.; Stojkovic, M.; Tian, A.; Jewett, M. C.; Altman, R. B.; Wittmann, B. J. — bioRxiv, 2026-09-11
  Inverse FoldDir applies Dirichlet flow matching on the amino-acid simplex for structure-conditioned inverse folding with controllable inpainting and soft priors. It surpasses ESM-IF1 on CATH recovery and yields experimentally validated nanobody variants.
  Affiliations: Stanford University

• 📄 Fixed-Dimensional Latent Flow for Generating Variable-Size 3D Molecules
  Weichi Yao, Cameron Gruich, Bryan R. Goldsmith, Yixin Wang — arXiv, 2026-09-08
  EF-TALFM uses fixed-dimensional latent flow matching to sample variable-size 3D molecules, followed by autoregressive decoding of atom types, coordinates, and chemical states. It achieves high validity/novelty and doubles DFT-verified hit rates via an internal ranking head.

• 📄 PocketVE: Stable and Property-Guided Structure-Based Drug Design with Variance-Exploding Diffusion
  Peining Zhang, Jinbo Bi — arXiv, 2026-09-08
  PocketVE is a protein-pocket-conditioned VE diffusion model with classifier-free multi-property guidance and pocket regularization. It markedly improves 3D validity and strain while preserving docking and molecular-property quality on CrossDocked.

• 🧬 Interaction Profiles as a Universal Language for Generative Molecular Design with ShEPhERD-2
  Abeywardane, K. A.; Walker, K.; Coley, C. W. — bioRxiv, 2026-09-12
  ShEPhERD-2 is a 3D generative model conditioned on explicit interaction profiles (shape, electrostatics, directional pharmacophores) as a chemotype-agnostic design specification. It realizes precise control for fragment merging, selectivity, dual-target, and modality hopping without retraining.
  Affiliations: Massachusetts Institute of Technology

• 📄 scDEFT: A deep learning framework for drug-effect prediction and counterfactual reasoning
  Murthy Devarakonda — arXiv, 2026-09-09
  scDEFT learns a drug-conditioned operator on single-cell representations to predict state shifts and responder status, then attributes drivers under cell-composition control. On multi-cohort IBD data, it achieves AUROC 0.70 pre-treatment and supports counterfactual cohort prediction.

• 🧬 CellART: a unified framework for extracting single-cell information from high-resolution spatial transcriptomics
  Chen, Y.; Liu, Y.; Wang, Z.; Zeng, Y.; ...; Chen, H.; Wang, J.; Xiao, J.; Yang, C. — bioRxiv, 2026-09-08
  CellART jointly performs cell segmentation and cell-type annotation from high-resolution spatial transcriptomics by fusing images, ST counts, and scRNA-seq via deep probabilistic modeling. It scales to millions of spots and reveals tumor–immune interactions and transient cancer states.
  Affiliations: HKUST

• 📄 MIRAGE: Measuring Interpolation and Redundancy in Affinity GEneralization
  Mehdi Yazdani-Jahromi, Sanjay Padhi, Ivan Garibay — arXiv, 2026-09-13
  MIRAGE is a benchmark that stratifies targets by historical protein-family support to disentangle interpolation from true generalization in structure-based affinity and pose prediction. It exposes support-driven performance inflation and recommends reporting excess over support-insensitive baselines.

• 📄 Predicting Collision Cross Sections with GRACE: Geometric Residual Adduct Conditioning via Early-fusion
  Parthasarathy Suryanarayanan, Susanta Das, Shreyans Sethi, Kenneth M. Merz, Joseph A. Morrone — arXiv, 2026-09-10
  GRACE predicts ion-mobility CCS with a 3D encoder that early-fuses adduct identity and learns residuals over a physics baseline. It sets state-of-the-art errors across random, scaffold, and adduct-sensitive splits and holds on external datasets.

• 📄 Predicting directional flexibility in proteins
  Vsevolod Viliuga, Leif Seute, Matteo Tadiello, Nicolas Wolf, Frauke Gräter, Arne Elofsson — arXiv, 2026-09-08
  BackFlip-2 is an SE(3)-equivariant GNN that directly predicts directional backbone flexibility and pairwise dynamic correlations from a single structure. It matches much larger ensemble-generation models while being orders of magnitude faster.

• 📄 Dynamic language model representations for multi-objective reaction optimisation
  Joshua W. Sin, David Ming Segura, Bojana Ranković, Siu Lun Chau, ..., Kurt Püntener, Maximilian J. Notheis, Raphael Bigler, Philippe Schwaller — arXiv, 2026-09-10
  Learns task-adaptive reaction representations dynamically from textual condition descriptions within a multi-objective Bayesian optimization loop. It outperforms descriptor and one-hot baselines and prospectively finds gram-scale optimal conditions in complex catalysis.

• 📄 Multimodal deep learning from spectra for small-molecule structure identification: enhancing robustness with mixed-condition training
  Bowen Gao, Lei Zhu, Yiying Wang, Wenjie Yu — arXiv, 2026-09-13
  Proposes mixed-condition training and mixture-of-experts fusion for multimodal spectra-based structure identification under missing/degraded MS/IR/NMR. It substantially boosts reranking robustness and single-modality performance without sacrificing complete-input accuracy.

• 🧬 PharmCast: rapid generation of three-dimensional pharmacophore fingerprints from two-dimensional structure without conformer generation
  Muskal, S. M.; McGregor, M. J. — bioRxiv, 2026-09-07
  PharmCast predicts full 3D pharmacophore fingerprints directly from SMILES via a feedforward network, removing conformer generation. It enables millisecond-scale, scaffold-hopping similarity with high fingerprint fidelity to conformer-based references.
  Affiliations: Eidogen-Sertanty, Inc.

• 🧬 How do Co-Folding Models Organize Structural Information?
  Park, M.; Kim, S.; Moon, S.; Kim, H.; Jeon, G.; Kim, W. Y. — bioRxiv, 2026-09-08
  Dissects co-folding (Boltz-1) representations, revealing distinct single/intra/inter-chain streams and a Mix–Compress–Refine trajectory with diffusion reconciling inconsistent constraints. The findings suggest partition-aware trunks for future complex prediction models.
  Affiliations: Department of Chemistry, KAIST

• 🧬 TAPAS: Learned integration of AlphaFold3 confidence and geometric features for TCR-pMHC binding prediction
  Kim, H. Y.; Han, H. J.; Kim, D. — bioRxiv, 2026-09-09
  TAPAS integrates AlphaFold3 interface confidence, geometric features, and sequence embeddings in a tabular learner for TCR–pMHC binding. It consistently matches or exceeds the strongest single AF3 metric across multiple benchmarks via complementary feature fusion.
  Affiliations: Korea Advanced Institute of Science and Technology

• 📄 Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets
  Judith Bernett, Anton Spannagl, Joel Ås, Markus List, David B. Blumenthal — arXiv, 2026-09-09
  Audits shortcut biases in PPI datasets and introduces an optimization-based pipeline for similarity-aware splitting and bias-minimizing negative sampling. The approach reduces non-biological shortcuts and improves reliability of learned PPI predictors.

• 📄 A Trust-Network-Based Federated Learning Framework for Multi-Center Aging Clock Prediction
  Chunxu Zhang, Bo Li, Wenliang Wang, Yang Liu, ..., Yo-ichi Nabeshima, Akinori Yamamura, Bo Yang, Qiang Yang — arXiv, 2026-09-09
  TNFL is a trust-network federated framework with an age-aware mixture-of-experts and generative replay for multi-center aging clocks. It enables accurate, interpretable predictions and reveals coordinated higher-order protein subnetworks associated with aging.