Weekly BioML Digest [September 28, 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)

1100 matched filters -> 20 selected after LLM relevance + novelty ranking.

• 🌿 Miniaturizing and modifying natural proteins with Raygun
  Devkota, Kapil, Shonai, Daichi, ..., Soderling, Scott, Singh, Rohit — Nature, 2026-09-24
  Introduces Raygun, a generative AI that encodes proteins as length-agnostic probability distributions from language-model embeddings to miniaturize/expand sequences while preserving structure and function; validated by compact fluorescent proteins and enhanced EGFR-binding EGF variants.

• 🌿 Breaking timescales with generative sampling of conformational transitions
  Tang, Chenyu, Pandey, Mayank Prakash, ..., Dehez, François, Chipot, Christophe — Nature, 2026-09-24
  Gen-COMPAS couples a denoising diffusion model with committor-guided filtering to reconstruct rare biomolecular transition pathways without predefined collective variables, recovering transition states and free-energy landscapes from endpoint structures across diverse systems.

• 📡 Generative AI designs functional thiolation domains for reprogramming non-ribosomal peptide synthetases
  Bülbül, Emre F., Bang, Seounggun, ..., Klakow, Dietrich, Bozhüyük, Kenan A. J. — Nature Communications, 2026-09-22
  Uses generative protein models (ESM3, ProteinMPNN, EvoDiff) in design–build–test–learn cycles to create de novo thiolation domains that reprogram non-ribosomal peptide synthetases, yielding active hybrids and up to ~3× higher output than native parts.

• 📡 A structure-informed deep learning framework for modeling TCR-peptide-HLA interactions
  Cao, Kai, Li, Rui, ..., Uhler, Caroline, Xavier, Ramnik J. — Nature Communications, 2026-09-25
  Presents StriMap, a structure-informed deep learning framework integrating physicochemical, sequence, and structural features to predict TCR–peptide–HLA interactions, enabling neoepitope prioritization and experimentally validated microbial mimic discovery in autoimmunity.

• 📡 SpaCEy links spatial tissue patterns to clinical outcomes using explainable graph neural networks
  Rifaioglu, Ahmet Sureyya, Ervin, Egle Helene, ..., Tanevski, Jovan, Saez-Rodriguez, Julio — Nature Communications, 2026-09-26
  SpaCEy is an explainable graph neural network that models tissues as spatial graphs to link spatial proteomic patterns to clinical outcomes, identifying spatial markers and substructures associated with survival and progression across lung and breast cancer cohorts.

• 📰 The Fragility of Bioactivity Prediction: Rigorous Dataset Splits Expose the Illusion of ML Accuracy.
  Kisung Lee, Galymzhan Moldagulov, Bartosz A Grzybowski — Chemistry (Weinheim an der Bergstrasse, Germany), 2026-09-24
  Shows that rigorous OOD-aware splits make bioactivity prediction collapse toward random, with k-NN matching state-of-the-art ML across splits; calls for leakage-resistant evaluation to claim genuine generalization in molecular ML.
  Affiliations: Center for Algorithmic and Robotized Synthesis (CARS), Institute For Basic Science (IBS); Department of Chemistry, Ulsan National Institute of Science and Technology

• 📰 Deep learning of tautomer stability from crystallographic proton positions
  Xiao-Lin Pan, Chao Han, Feng-Yang Han, Ying-Kai Zhang — Chemical Science, 2026-09-23
  Trains a 2D graph neural network on >1.1M crystallographically resolved proton placements to predict stable tautomers directly from topology, enabling large-scale tautomer assignment and improving PDBbind ligand plausibility.

• 📡 Generating protein hydrogels with customizable stress relaxation behavior via deep learning-driven entanglement design
  Deng, Puqing, Wu, Yutong, ..., Sun, Fei, Gao, Hanyu — Nature Communications, 2026-09-21
  TangleDiff is a diffusion-based generative framework for homodimeric entangled protein design with programmable binding energy; designed dimers form hydrogels with stress relaxation tuned by predicted affinities and achieve >70% in-silico foldability.

• 📰 AlloPool is a deep learning framework that infers protein allostery from molecular dynamics simulations.
  Matthieu Marfoglia, Miguel A Pedraza-Joya, Lucas Guirardel, Aisima Chatzi Souleiman, Patrick Barth — PLoS biology, 2026-09-01
  AlloPool, a temporal graph neural network, prunes residue contacts from MD to reveal minimal, time-resolved allosteric networks that govern conformational transitions, outperforming prior ML in reconstructing dynamics and mapping perturbation effects.
  Affiliations: Interfaculty Institute of Bioengineering, École Polytechnique Fédérale de Lausanne; Ludwig Institute for Cancer Research Lausanne, Lausanne

• 🤖 Guiding large language models to predict edit sequences for molecular synthesizability optimization
  Li, Junren, Lai, Luhua — Nature Machine Intelligence, 2026-09-23
  SynCraft reframes synthesizability optimization as LLM-predicted edit sequences that minimally modify molecules to cross synthesis cliffs, outperforming projection/filters and rescuing high-scoring candidates while preserving pharmacophores.

• 📰 CRAFT: A Chemical Reaction-Aware Transformer for Synthesizable Molecular Optimization via Multimodal Learning
  Run-Fu Yu, Tong-Mao Ma, Zian Song, Qi-Yao Yin, Shuang Wang — Journal of Medicinal Chemistry, 2026-09-23
  CRAFT ensembles reaction-aware Transformer policies with MCTS to optimize molecules under explicit reaction constraints, generating high-scoring, synthesizable drug-like candidates and transparent forward-synthesis routes (e.g., DRD2/AKT1/CXCR4).

• 📰 DesignMaster: A Multi-Conditional Diffusion Framework for Rational PROTAC Design.
  Binze Shi, Yi Hao, Jie Liu, Tong Pan, ..., Michael Roy, Ashley P Ng, Xuequn Shang, Fuyi Li — Bioinformatics (Oxford, England), 2026-09-24
  DesignMaster is a multi-conditional diffusion model with E(3)-equivariant graph Transformers that conditions linker length and physicochemical properties to generate PROTACs with improved validity, recovery, and conformational agreement to crystal references.
  Affiliations: School of Computer Science, Northwestern Polytechnical University; South Australian immunoGENomics Cancer Institute (SAiGENCI), Adelaide University; ...

• 📰 A unified framework for the systematic detection of non-canonical proteins in standard proteomics workflows using a Ribo-seq informed transcriptomic language model
  Provencher, Nicolas, Leblanc, Sébastien, Jacques, Jean-François, Roucou, Xavier — Scientific Reports, 2026-09-25
  Retrains a transcriptomic language model with 40k Ribo-seq TISs to predict 48k non-canonical ORFs and builds a size-controlled Swiss-Prot/ncProt database, enabling routine MS workflows to detect >500 nc-proteins with high concordance.

• 📰 MEGPNM as a multiscale edge-aware GAT network with hybrid pooling predicting permeability of non-peptidic macrocycles
  He, Shida, Wei, Lesong, ..., Zou, Quan, Zhang, Feng — Communications Chemistry, 2026-09-24
  MEGPNM employs multiscale edge-aware graph attention with Jumping Knowledge to predict permeability of non-peptidic macrocycles, outperforming deep and fingerprint baselines and highlighting structural motifs linked to chameleon behavior.

• 📰 Physics informed machine learning predicts binding free energy changes in protein-RNA complexes upon mutations
  Xu, Wang, Zhang, Houfang, ..., Zhao, Yunjie, Peng, Yunhui — Communications Biology, 2026-09-26
  Pred-MutPRI is a physics-informed ML for ΔΔG in protein–RNA complexes, combining thermodynamic cycle consistency, atom-level interaction graphs, masked ESM-2 entropy, and AlphaFold3 strain features to achieve state-of-the-art blind-test performance.

• 📰 A general deep learning method based on substrate information for genome-scale mining of natural product biosynthetic enzymes
  Li, Yangyang, Feng, Yuntong, ..., Han, Wenzheng, Liu, Hao — Journal of Cheminformatics, 2026-09-23
  ConESI integrates substrate-aware deep learning with metric-discriminative contrastive learning to mine genome-scale biosynthetic enzymes directly from sequences, stably retrieving cross-species/cross-substrate enzyme–substrate pairs.

• 📰 A Machine Learning Framework Enables Supervised Treatment Response Prediction from Tumor Transcriptomics across Cancer Types.
  L. Pal, E. Gertz, N. Nair, S. Mukherjee, ..., Danh-Tai Hoang, S. Hannenhalli, Alejandro A. Schäffer, Eytan Ruppin — Cancer research, 2026-09-24
  EXPRESSO is a supervised transcriptomic framework trained across 91 cohorts to predict treatment response to six frontline therapies, consistently outperforming 20 published signatures and generalizing in 22 independent validations.

• 📰 FluxAtlas is a pan-cancer metabolic flux atlas that identifies tumor-selective dependencies and prognostic biomarkers.
  M. Kohli, Aurelien Tripp, Anastasiia Shcherbakova, George Poulogiannis — Cell reports, 2026-09-24
  FluxAtlas integrates enzyme constraints and diffusion limits to infer pan-cancer metabolic fluxes from >10k tumors, revealing tissue-selective dependencies and building ML survival predictors that identify biotin uptake as a prognostic biomarker.

• 📰 TACTIC: A transfer learning framework to predict drug interactions in emerging pathogens.
  Carolina H Chung, David C Chang, Nicole M Rhoads, Prajna Lalitha, ..., Mercy A Okezue, Bhanuz Dechayont, Ashlee D Brunaugh, Sriram Chandrasekaran — Cell reports methods, 2026-09-21
  TACTIC transfers drug-combination learning across 12 bacteria to predict interactions for understudied pathogens, validating novel synergies against M. abscessus, P. aeruginosa, and S. aureus from ~600k predicted pairs.
  Affiliations: Department of Biomedical Engineering, University of Michigan; Department of Biomedical Engineering, University of Michigan; ...

• 📰 Large-scale discovery platform enables identification of peptides targeting drug-resistant candidiasis.
  Bahar Behsaz, Abhinav Adduri, Mustafa Guler, Osama G Mohamed, ..., Pieter C Dorrestein, Jason Clement, Ashootosh Tripathi, Hosein Mohimani — Cell reports methods, 2026-09-21
  NPDiscover fuses genome mining, metabolomics, and ML to prioritize NRPs, discovering edaphochelin A that kills multidrug-resistant Candida auris/glabrata via respiratory disruption and demonstrating in vivo safety/efficacy.
  Affiliations: Chemia Biosciences Inc, Pittsburgh; Computational Biology Department, School of Computer Science; ...

🧬 Preprints (arXiv + bioRxiv)

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

• 🧬 Computational design of metalloproteases
  Chen, A.; Wu, K.; Choi, H.; Venkatesh, P.; ...; Oller-Salvia, B.; Hyman, A. A.; Hilvert, D.; Baker, D. — bioRxiv, 2026-09-21
  Uses RoseTTAFold Diffusion 2 to design de novo zinc metalloproteases with >10^8–1010-fold peptide-bond hydrolysis acceleration, enabling selective cleavage of disease targets and bio‑orthogonal control of protein function.
  Affiliations: University of Washington

• 🧬 Integrating structural homology with deep learning to achieve highly accurate protein-protein interface prediction for the human interactome
  Xiong, D.; Torres, M.; Murray, D.; Zhang, Z.; ...; Naravane, A. C.; Fragoza, R.; Honig, B.; Yu, H. — bioRxiv, 2026-09-21
  Integrates 3D structural homology with geometric deep learning (PIONEER2) to predict partner-specific protein–protein interfaces at proteome scale, outperforming AlphaFold3 on low-confidence complexes and guiding mutation impact across the human interactome.
  Affiliations: Cornell University

• 🧬 Speciesformer learns conserved cellular states for cross-species generative virtual cell modeling
  Wang, J.; Dong, J.; Li, G.; Fang, C.; Liu, L.; Gao, X. — bioRxiv, 2026-09-23
  Introduces a cross-species generative single-cell foundation model (Speciesformer) that aligns genes into an evolution-informed space and generates/conditions cell states and perturbation responses across 11 species and 154 tissues.
  Affiliations: King Abdullah University of Science and Technology

• 🧬 Coevolutionary mining of prokaryotic non-coding elements with a genome language model
  Li, D. B.; Brixi, G.; Kim, A. S.; Fiamenghi, M. B.; ...; Deisseroth, K.; Wilkinson, M. E.; Fischbach, M. A.; Hie, B. — bioRxiv, 2026-09-24
  Presents Minerva, a genome language model framework that mines coevolution directly from DNA to predict ncRNA base-pairing, protein contacts, and repeats without alignments, uncovering widespread unannotated bacterial noncoding interactions.
  Affiliations: Stanford University

• 🧬 Reaction-aware hypergraph learning of chemical representations with LARK
  Qiao, J.; Liu, Y.; Jin, J.; Wang, D.; Zou, Q.; Su, R.; Wei, L. — bioRxiv, 2026-09-21
  LARK learns reaction-aware chemical representations via a role‑aware hypergraph Transformer pretrained on ~0.6M reactions, transferring to reaction prediction and single‑molecule property tasks by localizing transformation‑relevant substructures.
  Affiliations: Shandong University

• 📄 CRNDiff: Count-Native Diffusion Framework via Chemical Reaction Networks
  Yuxuan Qiu, Praful Gagrani, Tetsuya J Kobayashi — arXiv, 2026-09-25
  CRNDiff models count-native single-cell RNA-seq with a chemical-reaction-network diffusion process, enabling closed-form noising, FFBS sampling, and rare-subpopulation steering that improves conditional fidelity on scRNA-seq generation.

• 🧬 DeepSCENIC: transfer learning from sequence-to-function models enables causal gene regulatory network inference
  Partel, G.; De Winter, S.; Konstantakos, V.; Blaauw, C. H.; Aerts, S. — bioRxiv, 2026-09-22
  DeepSCENIC transfers sequence-to-function enhancer models (e.g., Enformer) to single-cell multiome to infer causal TF–enhancer–gene networks and simulate perturbations, improving enhancer–gene links and capturing state transitions across species.
  Affiliations: VIB Center of AI & Computational Biology, VIB-KU Leuven Center for Brain and Disease Research

• 🧬 State-selective inhibition of alpha-synuclein aggregation by de novo oligomer-binding proteins
  Zhang, Y.; Han, H. L.; Ortigosa-Pascual, L.; Miles, U. Z.; ...; Halfmann, R.; McShan, A. C.; Sahtoe, D. D.; Knowles, T. P. J. — bioRxiv, 2026-09-25
  Deep learning–designed de novo binders capture α‑synuclein’s NAC region in β‑strand conformation, selectively inhibiting fibril formation and secondary nucleation with substoichiometric effects in vitro and reduced seeding in cells.
  Affiliations: Yusuf Hamied Department of Chemistry, University of Cambridge

• 📄 PFArena: Benchmarking Language Models for Protein Modification
  Yawen Ouyang, Xinbo Zhang, Ziyuan Ma, Yixin Wu, ..., Wei-Ying Ma, Dahua Lin, Bowen Zhou, Hao Zhou — arXiv, 2026-09-24
  PFArena benchmarks PLMs, LLMs, and LLM agents for protein modification across single‑mutant generation and multi‑mutant ranking, revealing regime shifts where PLMs excel without target data and LLMs/agents gain with fitness evidence.

• 📄 TopU-LBVS: A Realistic Multi Target Benchmark for Ligand Based Virtual Screening
  Surbhi Kumar, Yuhe Zhou, Varun Shiralkar, Niu Huang, Baris Coskunuzer — arXiv, 2026-09-24
  TopU-LBVS provides a hard‑negative, multi‑target benchmark with standardized protocols, showing sharp performance drops from random to property‑matched decoys and enabling reproducible comparison of LBVS and molecular representations.

• 🧬 Discovery of Selective Small-Molecule Ligands of SV2C by AI-Enhanced Virtual Screening and Experimental Validation
  Brueckner, A. C.; Martin, M. F.; Khuttan, S.; Shields, B.; ...; Salahpour, A.; Bucher, M. L.; Coleman, J. A.; Miller, G. W. — bioRxiv, 2026-09-22
  An AI‑enhanced virtual screening pipeline combining CNN rescoring with homology/MD conformational ensembles discovers SV2C‑selective chemotypes with low‑μM affinity and isoform selectivity validated by TSA/SPA assays.
  Affiliations: SandboxAQ, Palo Alto

• 🧬 A Taxonomy-Informed Sparse DNA Foundation Model for Microbial Genomics
  Li, A. Z.; Wang, S.; Cheng, S.; Du, Y.; Liu, R. — bioRxiv, 2026-09-24
  MicroGlot is a taxonomy‑informed sparse MoE DNA foundation model with hyperbolic taxon embeddings that captures microbial phenotypes and taxonomy efficiently, improving probing/fine‑tuning performance across diverse microbes.
  Affiliations: The University of Texas at San Antonio

• 🧬 Vir2vec: A Genome-Wide Viral Embedding
  Rancati, S.; Arozarena Donelli, P.; Nicora, G.; Bergomi, L.; ...; Bellazzi, R.; Boucher, C.; Parimbelli, E.; Marini, S. — bioRxiv, 2026-09-24
  Vir2vec trains viral genome LMs on 565k genomes and introduces vGUE, achieving SOTA on five of seven viral understanding tasks with zero‑shot generalization and codon‑level SHAP attributions for host‑adaptation mutations.
  Affiliations: University of Florida

• 📄 Atelier: Learning Local Self-Supervised Features for CryoEM Volumes via Hypernetworks
  Phillip Lo, Sudarshan Babu, Dari Kimanius, Aly A. Khan — arXiv, 2026-09-24
  Atelier amortizes implicit neural representations for cryoEM via a transformer hypernetwork, yielding continuous local feature fields that boost downstream voxel‑level annotations across diverse complexes.

• 📄 Budgeted Quotient-Residual Guidance for Frozen Pocket-Conditioned Molecular Diffusion
  Xinyu Wang, Jinbo Bi, Minghu Song — arXiv, 2026-09-25
  Introduces quotient‑residual guidance for pocket‑conditioned molecular diffusion, activating distance/contact objectives at inference without retraining and improving ligand design validity and edit tasks on CBGBench.

• 🧬 MoTRUST: adaptive semantic protection for mosaic single-cell integration and molecular recovery
  Liu, Q.; Xu, Y. — bioRxiv, 2026-09-24
  MoTRUST unifies neural mosaic single‑cell integration with adaptive semantic protection and conditional residual diffusion, ranking first across 30 tasks and improving ATAC→RNA recovery while preserving point predictions.
  Affiliations: University of Science and Technology Beijing

• 📄 Neural Transport Nested Sampling
  David Yallup, Will Handley — arXiv, 2026-09-24
  Neural Transport Nested Sampling couples flow‑matching drift with a MH‑corrected Langevin kernel inside nested sampling to draw Boltzmann samples and estimate partition functions, outperforming neural baselines on Lennard‑Jones clusters.

• 📄 Multivariate conformal uncertainty propagation in multitask atomistic simulation: Successes and pitfalls
  Katharine Fisher, Michael Herbst, James Kermode, Youssef Marzouk — arXiv, 2026-09-25
  Develops multivariate conformal calibration for atomistic ML (energies, forces, stresses) that propagates calibrated uncertainty to elastic constants and vacancy energies, capturing correlations and error cancellation in materials workflows.

• 📄 When Riemann flows with Wasserstein: Generative Modeling of Probability Distributions on Manifolds
  Doron Haviv, Edward De Brouwer, Rishabh Anand, Rex Ying, Aïcha Bentaieb, Gabriele Scalia, Hector Corrada Bravo — arXiv, 2026-09-22
  Riemannian Wasserstein Entropic Flow Matching generates distributions on manifolds using optimal‑transport velocities, enabling whole single‑cell sample generation in hyperspherical spaces and protein conformations on the torus.

• 📄 Response-state Learning for Transferable Vibrational Spectroscopic Characterization with Electron Prior
  Zetong Li, Zhuosong Xie, Hengyu Fan, Jiaao Yu, ..., Zheng Lu, Liming Xu, Juanni Wu, Honglin Li — arXiv, 2026-09-24
  SENK, an SO(3)‑equivariant neural Kalman framework with electronic priors, jointly predicts Hessians and response tensors to deliver calibrated IR/Raman spectra that transfer from small molecules to biomolecular systems.