Weekly BioML Digest [October 19, 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)

1094 matched filters -> 20 selected after LLM relevance + novelty ranking.

• 🌿 Creating bottom-up RNA transfer vehicles from synthetic protein assemblies
  Schuhmacher, Maren Kirstin, Gruber, Christoph, ..., Giesert, Florian, Wurst, Wolfgang — Nature, 2026-10-15
  Uses generative protein design to build synthetic virus-like protein vehicles that deliver RNA with orders-of-magnitude higher efficiency than existing systems, with programmable tropism and in vivo validation in mouse and pig models.

• 📰 Atomic functional site scaffolding in biopolymers with RFdiffusion3.
  Jasper Butcher, Rohith Krishna, Raktim Mitra, Seth M Woodbury, ..., Maxwell Kazman, Lars L Schaaf, Frank DiMaio, David Baker — Science (New York, N.Y.), 2026-10-15
  Introduces RFdiffusion3, a free-atom generative method that scaffolds atomic functional sites across biopolymers and yields active de novo enzymes in 96-construct zero‑shot campaigns.
  Affiliations: Institute for Protein Design, University of Washington; Department of Biochemistry, University of Washington; ...

• 🌿 Computational design of metalloproteases
  Chen, Anqi, Wu, Kejia, ..., Hilvert, Donald, Baker, David — Nature, 2026-10-14
  Combines RFdiffusion-driven design with experimental validation to create highly active Zn-metalloproteases targeting disease proteins and to cage cytokines for bioorthogonal activation.

• 📡 Interpretable AI-guided discovery of degradable ionizable lipids with enhanced hepatic mRNA delivery in vivo
  Zhang, Zhen, Ren, Yongbo, ..., Kang, Feiyu, Yan, Hao — Nature Communications, 2026-10-12
  See above; interpretable discovery of degradable ionizable lipids enabling improved hepatic mRNA delivery.

• 📰 NeoBridge: Biophysics-informed neoantigen prioritization and clonal architecture profiling in pancreatic cancer.
  Alejandro Lana Mediavilla, Javier Pérez Cardoso, Jaime Álvarez-Benayas, Víctor Javier Sánchez-Arévalo Lobo — Cell genomics, 2026-10-15
  See above; NeoBridge hybrid model for neoantigen immunogenicity with cross‑cancer generalization.
  Affiliations: Department of the Higher School of Computer Science, Nebrija ARIES Research Group; Grupo de Oncología Cutánea, Servicio de Anatomía Patológica; ...

• 📰 AI-guided discovery of a brain-penetrant LILRB4 inhibitor active in microglial and 5xFAD models of Alzheimer's disease.
  Somaya A Abdel-Rahman, Moustafa T Gabr — Bioorganic chemistry, 2026-10-15
  Performs AI-guided ultra‑large screening to discover brain-penetrant LILRB4 inhibitors that restore microglial Aβ uptake and improve cognition and pathology in 5xFAD mice.
  Affiliations: Department of Radiology, Columbia University Irving Medical Center; Department of Radiology, Molecular Imaging Innovations Institute (MI3); ...

• 📰 Identification of FLT3-ITD inhibitors through AI-guided virtual screening, biochemical validation, DFT analysis, and molecular dynamics simulation.
  Cong Wang, Huijie Han, Xiaokun Zhang, Chaochun Wei, Hong Yan — Bioorganic chemistry, 2026-10-15
  Applies a GNN‑guided virtual screening and docking/MD pipeline to identify novel FLT3‑ITD inhibitors, with sub‑micromolar potency and favorable QM/MM and MD profiles.
  Affiliations: College of Chemistry and Life Science, Beijing University of Technology; Department of Chemistry, Tsinghua University; ...

• 📡 De novo design of lipopeptides with antimicrobial potency enabled by a chain-as-prompt deep learning framework
  Zhang, Fan, Yu, Xuchang, Wu, Yuzhu, Li, Lei — Nature Communications, 2026-10-17
  Presents a ‘chain‑as‑prompt’ deep learning framework that designs antimicrobial lipopeptides with 96% experimental hit rate, including a broad‑spectrum lead validated in mice.

• 📰 Musical Peptide: Fragmentation of Musical Protein for AI-driven Bioactive Peptide Discovery Based on a New DeepPepQSAR Pipeline
  Zhou, Peng, Su, Jun, ..., Gan, Yulu, Zhang, Yunyi — Interdisciplinary Sciences: Computational Life Sciences, 2026-10-13
  See above; DeepPepQSAR + MPep framework producing multiple validated AMPs.

• 🔭 Pep2Net: an integrated peptide knowledge base and cross-attention model for multi-task prediction of peptides
  Zheng, Boyan, Yuan, Jiahao, ..., Chen, Haowen, Fu, Xiangzheng — Genome Biology, 2026-10-14
  See above; unified peptide DB + cross‑attention model enabling large‑scale multifunctional peptide prediction and design.

• 📰 DeepNPA: A protBert-based framework for residue-level annotation of neuropeptide precursors.
  Xin-Yu Li, Yang Yu, Li-Ping Liu, Meng-Ze Du, Qiang Tang, Juan-Juan Kang — iScience, 2026-10-16
  Develops DeepNPA, a ProtBert‑based, imbalance‑aware tool for residue‑level neuropeptide precursor annotation and cleavage site prediction with state‑of‑the‑art accuracy.
  Affiliations: School of Intelligent Medicine, Innovative Institute of Chinese Medicine and Pharmacy; State Key Laboratory of Southwestern Chinese Medicine Resources, Chengdu University of Traditional Chinese Medicine; ...

• 📰 Interpretable Protein−Nucleic Acid Binding Site Prediction via Domain-Guided Multimodal Learning
  Wen-Yu Xi, Ru-Heng Wang, Xiu-Cai Ye, Tetsuya Sakurai, Le-Yi Wei — ACS Synthetic Biology, 2026-10-16
  BiMTBind, a domain‑guided multimodal deep model, integrates sequence/structure with protein‑domain priors to improve residue‑level protein–nucleic acid binding site prediction and interpretability.

• 📰 MGCLSyn: multimodal graph contrastive learning for drug synergy prediction
  Yang, Ying, Ni, Jiancheng, ..., Wang, YanWei, Zhang, Chong — Journal of Computer-Aided Molecular Design, 2026-10-16
  MGCLSyn uses multimodal graph contrastive learning to integrate molecular graphs, drug similarities, fingerprints, and cell-line expression for superior drug–drug synergy prediction.

• 📰 Multiscale cross modal fusion and cascaded signal propagation for drug target interaction prediction.
  Bowen Wang, Xiaolan Xie, Haitao Zou, Shaoliang Peng — iScience, 2026-10-16
  SF‑DTI aligns graph‑based drug structures with pretrained semantic embeddings and fuses multi‑encoder protein features via cascaded cross‑modal propagation to improve DTI prediction in cold‑start and cross‑domain settings.
  Affiliations: College of Computer Science and Engineering, Guilin University of Technology; College of Computer Science and Electronic Engineering, Hunan University

• 📰 DHF-DTI: a dual-modality hybrid fusion framework for drug-target interaction prediction
  Chen, Dongjie, Liu, Yuemeng, ..., Yang, Zhihui, Zhao, Mingming — BMC Bioinformatics, 2026-10-16
  DHF‑DTI combines cross‑attention and bilinear fusion with supervised contrastive learning to integrate drug sequence/graph modalities, boosting DTI accuracy and generalization.

• 📰 BioInteract combines residue-level cross-attention with protein-language and physicochemical priors for interpretable drug–target interaction prediction
  Wang, Shengnan, Zhang, Qi, ..., Cheng, Jiangke, Huang, Zhi — Scientific Reports, 2026-10-15
  BioInteract provides an audited, leakage‑aware DTI benchmark with a cross‑attention model fusing ESM‑2 residue embeddings and molecular graphs, highlighting domain‑of‑applicability limits for transfer.

• 💻 Microenvironment Awareness Pragmatic Analysis for Generative Enzyme Function Reasoning.
  Jing Hao, Ke Yan, Bin Liu — Bioinformatics, 2026-10-14
  See above; generative microenvironment-aware EC prediction with chain‑of‑thought style outputs.

• 📰 AI-EXAFS: Physics-Informed Prediction of EXAFS from Atomic Structure.
  Javier Heras-Domingo, Andrea Ruiz-Ferrando, Adam H Clark, Stephan Pollitt, Sharon Mitchell, Javier Pérez-Ramírez, Núria López — Advanced science (Weinheim, Baden-Wurttemberg, Germany), 2026-10-18
  AI‑EXAFS uses a physics‑informed GNN to predict EXAFS spectra directly from 3D atomic structures across diverse transition‑metal complexes, enabling rapid structure–spectrum inference.
  Affiliations: Institute of Chemical Research of Catalonia (ICIQ-CERCA), The Barcelona Institute of Science and Technology (BIST); Departament de Química Inorgànica i Orgànica (Secció de Química Inorgànica), Institut de Química Teòrica i Computacional (IQTC); ...

• 📰 Hybrid Graph Neural Network Ensemble with Stacked Meta-Learning for Laboratory-Conditioned Chromatographic Retention-Time Prediction
  Sheng-Nan Wang, Qi Zhang, Xu-Cheng Liu, Zi-Xuan Wang, Jiang-Ke Cheng, Hao-Long Hong — The Journal of Physical
  Chemistry B, 2026-10-15
  See above; hybrid GNN meta‑learner with lab embeddings for RT generalization across sites.

• 📰 Binary interaction parameters of a million mixtures for PCP-SAFT using deep learning.
  Carl Hemprich, Gernot Bauer, Gustavo Chaparro, André Bardow, Philipp Rehner — Digital discovery, 2026-10-14
  See above; DL‑predicted SAFT k_ij parameters extending mixture coverage by >10^6 pairs.
  Affiliations: Energy and Process Systems Engineering, ETH Zurich 8092 Zurich Switzerland chemprich@ethz.ch gchaparro@ethz.ch abardow@ethz.ch prehner@ethz.ch.; Institute of Thermodynamics and Thermal Process Engineering, University of Stuttgart 70569 Stuttgart Germany gernot.bauer@itt.uni-stuttgart.de.

🧬 Preprints (arXiv + bioRxiv)

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

• 🧬 Complex enzyme active site scaffolding by iteratively detuned catalytic guidance
  Qian, T.; Wang, W.; Gao, S.; Muraishi, A.; ...; Wang, Y.; Huang, X.; DiMaio, F.; Baker, D. — bioRxiv, 2026-10-16
  Introduces CaGE, a constraint-guided sequence design loop that conditions AlphaFold3/RF3 diffusion with classifier-free and training-free motif guidance to satisfy complex catalytic geometries, yielding experimentally active de novo enzymes with multi-residue/cofactor active sites.
  Affiliations: Institute for Protein Design, University of Washington; Howard Hughes Medical Institute, University of Washington

• 🧬 Protein Deimmunization with a Generative Diffusion Model
  Chen, Z.; Yu, C.-c.; Zhou, C.; Jiang, X.; Nguyen, L.; Myhrvold, C.; Jaakkola, T. S.; Jiang, K. — bioRxiv, 2026-10-16
  Presents MESSI, a discrete diffusion model that designs structure-consistent protein sequences while eliminating predicted MHC-II epitopes, producing deimmunized therapeutics with preserved or improved activity in luciferase, mini-binder, and uricase.
  Affiliations: Princeton University

• 🧬 A high-resolution protein-RNA binding atlas for structure-aware interaction prediction and post-transcriptional regulatory modeling
  Karimzadeh, M.; Fish, L.; Karner, H.; Linder, J.; ...; Wang, B.; Najafabadi, H. S.; Leonetti, M.; Goodarzi, H. — bioRxiv, 2026-10-13
  ecCLIP builds a high-resolution atlas of RBP–RNA binding and Borzuki/Pythia learn sequence- and structure-aware predictors, improving RNA half-life prediction and experimentally validating structure-dependent binding via massively parallel reporter CLIP.
  Affiliations: Arc Institute, University of California

• 📄 ProTok: Unified Protein Tokenization via Modular Multi-Codebook Quantization
  Romain Hardy, Xiaoman Zhang, Pranav Rajpurkar — arXiv, 2026-10-13
  Proposes ProTok, a unified multi-codebook tokenizer over sequence, structure, and function that enables discrete multimodal protein representations supporting reconstruction, retrieval, and joint backbone–sequence generation.

• 🧬 OmniScore: Universal Scoring of Diverse Biomolecular Complexes via Equivariant Geometry-Aware Discrete Representation Learning
  Bui, T.-C.; Hong, Y.; Song, J.; Tao, Y.; Chung, I.; Park, J. G.; Ko, J.; Lee, J. — bioRxiv, 2026-10-14
  OmniScore is an equivariant, geometry-aware encoder pretrained across proteins, nucleic acids, small molecules, and complexes, providing universal embeddings that improve complex quality assessment and interaction prediction tasks.
  Affiliations: Seoul National University

• 📄 ADT: All-Atom Diffusion Tokenizer for 3D Molecular Structures
  Xuantuo Huang, Encheng Su, Aoran Wang, Pan Tan, ..., Yizhou Wang, Chen Tang, Shixiang Tang, Mingxiao Li — arXiv, 2026-10-16
  All-Atom Diffusion Tokenizer (ADT) discretizes 3D molecular structures with atom-level tokens for small molecules and residue-level all-atom tokens for proteins, coupling diffusion-based dequantization to enable compact autoregressive generation with accurate all-atom reconstruction.

• 📄 OctMole: Multiscale 3D Molecular Generation with Hierarchical Spatial Flows
  Yu-Tong Jin, Yi-Yang Tian, Qiwei Ye, Peng-Shuai Wang — arXiv, 2026-10-12
  OctMole introduces hierarchical octree flows that jointly learn sparse occupancy and atom assignments, enabling efficient multiscale 3D molecular generation with strong chemical validity and flexible spatial conditioning.

• 📄 MIRRA: Fast, High-Fidelity Molecular Generation with Chemistry-Informed Discrete Flow Matching
  Ge Sun, Juan J. de Pablo — arXiv, 2026-10-12
  MIRRA is a chemistry-informed discrete flow-matching generator over large vocabularies of chemical building blocks, refreshing only needed positions to accelerate sampling while improving distributional fidelity and property-targeted hit rates.

• 📄 LLM-to-Flow Distillation: Teaching Natural Language Goals to Scientific Generators
  Riccardo De Santi, Federico Di Gennaro, Angelo Gnazzo, Sophia Tang, ..., Fanny Yang, İsmail İlkan Ceylan, Michael M. Bronstein, Andreas Krause — arXiv, 2026-10-12
  LLM-to-Flow Distillation aligns domain-specific generative flows to natural-language scientific goals by using an LLM teacher to set adaptive curricula and DPO-based preference distillation, enabling peptide and small-molecule generators to satisfy high-level discovery objectives without explicit rewards.

• 📄 Blitz: Accelerating the Diffusion Modules of AlphaFold3 and Boltz-2
  Vinay Saji Mathew, Soundar Kumara, Gretta Kellogg, William K. M. Lai — arXiv, 2026-10-16
  Blitz distills AlphaFold3 and Boltz-2 diffusion modules into eight-step samplers via geometric noise schedules and multistep consistency training, retaining or exceeding teacher-level structure/interface accuracy at >20× fewer denoising steps.

• 📄 CloneOT: predicting clone-specific perturbation responses with lineage-supervised neural inverse optimal transport
  Yanshuo Chen, Wei Chen, Heng Huang — arXiv, 2026-10-12
  CloneOT learns a clone-informed transport cost and map via neural inverse optimal transport, using lineage barcodes to predict clone-specific single-cell perturbation responses and recover drug-tolerant programs.

• 📄 CocoRNA: Collective RNA Design with Cooperative Multi-agent Reinforcement Learning
  Tianmeng Hu, Biao Luo, Ke Li — arXiv, 2026-10-16
  CocoRNA applies cooperative multi-agent reinforcement learning with a centralized critic to decompose RNA inverse folding, achieving faster and more reliable design that generalizes to unseen secondary structures.

• 📄 Learning the Shared Latent Structure of Cryo-EM Inverse Problems
  Nithin Chalapathi, Theo Sternlieb, Ethan Han, Sonya M. Hanson, Pilar Cossio, Aditi S. Krishnapriyan, Ritwik Gupta — arXiv, 2026-10-15
  EMber pretrains vision transformers on >50M experimental cryo-EM particles plus simulated projections, learning dense spatial features that improve conformational discrimination, population estimation, and reconstruction across cryo-EM inverse problems.

• 📄 zPocket: Better Pocket Finding with Frozen Co-Folding Features
  Tally Portnoi, Allison M Keys, Benjamin R DiFrancesco, Nate Gruver — arXiv, 2026-10-13
  zPocket trains lightweight heads on frozen co-folding trunk features to predict ligandable pockets, outperforming baseline pocket finders on a challenging, out-of-distribution PocketWatch benchmark and enabling proteome-scale scans.

• 🧬 Scalable Enzyme Reaction Retrieval Enables Metagenomic Biocatalyst Discovery
  Luo, D.; Jiang, M.; Wang, B. — bioRxiv, 2026-10-18
  EzHit formulates bidirectional enzyme–reaction retrieval with geometric confidence via Mahalanobis distance, scaling to trillions of associations to annotate metagenomic proteins and experimentally validate novel biocatalysts.
  Affiliations: Xiamen University

• 📄 RetroPlanner: Multi-Route Expansion and Multi-Objective Reasoning for LLM-Based Retrosynthesis Planning
  Ye Rin Kim, Yongjun Jeong, Seongjun Yun, Hanbum Ko, Soohyun Oh, Sungbin Lim, Sungwoong Kim — arXiv, 2026-10-12
  RetroPlanner is an LLM-based retrosynthesis agent that expands and jointly ranks multiple routes by plausibility, consistency, and cost, steering search toward higher-quality synthetic plans while maintaining success rate.

• 📄 MolHop: Scaffold-Hopping Molecular Design with Agentic Reinforcement Learning
  Yanyi Su, Chenyan Wang, Yulei Gong, Yeyun Chen, Xiaohui Yang, Jun Cheng — arXiv, 2026-10-14
  MolHop frames scaffold hopping as agentic RL over a hierarchical scaffold navigation and editing action space, enabling LLM agents to discover functionally similar molecules with distinct cores.

• 📄 STAMP as a topology-aware molecular language for fragment-level learning and editing
  Yuxuan Ren, Jiatong Li, Ziqiao Meng, Yatao Bian — arXiv, 2026-10-13
  STAMP defines a topology-aware molecular language that explicitly encodes fragment identities, attachment sites, and tree structure, enabling executable fragment-level editing and improved activity-cliff prediction and optimization.

• 📄 PairIT: Autoregressive Transformers for Low-Data Molecule Optimization
  Hannah Lawrence, Bodhi P. Vani, Ji Won Park, Natasa Tagasovska, Samuel Don Stanton, Andrew Martin Watkins, Michael Maser, Ewa M. Nowara — arXiv, 2026-10-13
  PairIT trains autoregressive transformers on matched molecular pairs to learn edit operations conditioned on prompts for direction and magnitude, supporting multi-property, seed-constrained optimization in low-data regimes.

• 📄 PMO-Dock: Benchmarking Docking, Specificity, and Generalization in Molecular Optimization
  Gor Simonyan, Tatevik Abrahamyan, Narek Abelyan, Tigran Fahradyan, Hrant Khachatrian — arXiv, 2026-10-13
  PMO-Dock provides a standardized docking-based molecular optimization benchmark with hit generation, lead optimization, and specificity tasks under strict oracle budgets, revealing no universally best method and highlighting generalization/sensitivity trade-offs.