Weekly BioML Digest [October 26, 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)

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

• 🌿 Programmable design of functional proteins from natural language
  Dai, Fengyuan, You, Shiyang, ..., Lu, Hongyuan, Yuan, Fajie — Nature, 2026-10-21
  Pinal, a 16B-parameter protein LLM, generates functional proteins directly from natural-language descriptions, with experimental validation across enzymes and pathways. It links intent-to-sequence/structure and improves metabolic output without iterative optimization.

• 🌿 An operational perturbation proteomics-based virtual cell model
  Sun, Rui, Qian, Liujia, ..., Zhu, Yi, Guo, Tiannan — Nature, 2026-10-22
  ProteinTalks pretrains on 38M time-resolved proteomics to learn dynamics-aware latent states, transferring to drug efficacy, synergy, resistance, and patient stratification. It operationalizes a proteomics-based virtual cell for in silico drug discovery with strong cross-system generalization.

• 🌿 Learning millisecond protein dynamics from what is missing in NMR spectra
  Wayment-Steele, Hannah K., El Nesr, Gina, ..., Ovchinnikov, Sergey, Kern, Dorothee — Nature, 2026-10-22
  Dyna-1 leverages ESM-3 representations to predict μs–ms protein dynamics from 'missing' NMR assignments, linking exchange-broadening to conserved, functionally critical motions. It opens a scalable path to learn dynamics-function relationships from public NMR repositories.

• 📰 DNA-BFN: Bayesian Flow Networks for Generative Modeling of Regulatory DNA Sequences.
  Han Yu, Qixiu Du, Lei Wei, Xiaowo Wang — Bioinformatics (Oxford, England), 2026-10-19
  DNA-BFN is a Bayesian flow network that generates regulatory DNA by iteratively updating per-position nucleotide beliefs, preserving discrete constraints. It outperforms diffusion-style baselines across species, enables few-shot design, and samples rapidly.
  Affiliations: Ministry of Education Key Laboratory of Bioinformatics, Center for Synthetic and Systems Biology; College of Artificial Intelligence, Tsinghua University

• 📡 Recommendation system-inspired learning of crossmodal interaction patterns between chromatin accessibility and gene expression
  Wang, Xiaotang, Lin, Xuanwei, ..., Li, Hao, Zhang, Yongqi — Nature Communications, 2026-10-24
  CrossRec reframes scATAC–scRNA coupling as a recommender graph, using a bipartite transformer to learn interpretable peak–gene regulatory attention. It generalizes under data scarcity, discovers cell-type–specific rewiring, and improves cross-task single-cell inference.

• 📰 De Novo Design of a Protein Binder to Probe Gas Channel and Enhance the Oxygen Tolerance of [NiFe]-Hydrogenase.
  Xuan Sun, Qi Xiao, Nan Zhang, Wenjin Li, ..., Peiyu Jiang, Yucen Huang, Geng Wu, Liyun Zhang — Angewandte Chemie (International ed. in English), 2026-10-19
  An AI-guided de novo protein binder occludes hydrogenase O2 ingress channels, experimentally tripling oxygen tolerance and mapping gas pathways by cryo-EM. It showcases generative protein design to probe and control metalloenzyme transport without mutation.
  Affiliations: State Key Laboratory of Microbial Metabolism, School of Life Sciences & Biotechnology; State Key Laboratory of Medicinal Chemical Biology, College of Life Sciences; ...

• 📰 Deep mapping of structural perturbations to energetics enables precise TCR design.
  Shifu Luo, Songming Zhang, Ying Shi, Junjie Li, Jiaxin Cai, Ning-Yi Shao, Yi Pan, Jinyan Li — Cell reports methods, 2026-10-19
  mpTCRai predicts residue-level TCR–pMHC structural energetics via a sequential presentation–recognition model, correlating perturbation hotspots with affinity (r = −0.88). It enables rational affinity tuning and cross-reactivity risk screening for TCR engineering.
  Affiliations: Faculty of Health Sciences, University of Macau; Faculty of Computer Science and Artificial Intelligence, Shenzhen University of Advanced Technology; ...

• 📰 AI-guided identification of a covalent SMAD3 regulatory mechanism of celastrol in diabetic kidney disease.
  Pengchen Chen, Hailong Deng, Wei Wang, Laihong Gao, ..., Ying Xiao, Ran Wei, Lirong Liu, Bing Guo — Pharmacological research, 2026-10-20
  Uni-Mol–guided prioritization led to mechanistic discovery that celastrol covalently targets SMAD3 (Cys421), altering nuclear dynamics and SMAD3/SMAD4 assembly to suppress fibrotic signaling in DKD. It exemplifies AI-driven target triage coupled to chemical proteomics.
  Affiliations: Center for Clinical Laboratories, The Affiliated Hospital of Guizhou Medical University; State Key Laboratory of Discovery and Utilization of Functional Components in Traditional Chinese Medicine, Guizhou Medical University; ...

• 📰 [pMHCBind: predicting major histocompatibility complex class I-peptide interactions via protein large language models and hypergraph neural networks].
  Jiani Ma, Hang Yang, Xiaolan Wang, Yong Zhang, Yuanting Zheng — Sheng wu gong cheng xue bao = Chinese journal of biotechnology, 2026-10-25
  pMHCBind integrates ESM protein LMs with hypergraph neural networks to model higher-order MHC I–peptide contacts, improving presentation prediction over state-of-the-art. Interpretable outputs recover hydrogen bonds, salt bridges, and hydrophobic modes at the interface.
  Affiliations: School of Information and Control Engineering, China University of Mining and Technology; School of Nursing and Health Management, Shanghai University of Medicine and Health Sciences; ...

• 📰 Teaching diffusion models physics: reinforcement learning for physically valid diffusion-based docking.
  J Henry Broster, Bojana Popovic, Diana Kondinskaia, Charlotte M Deane, Fergus Imrie — Chemical science, 2026-10-21
  A reinforcement-learning objective teaches diffusion docking to respect physical constraints, yielding more chemically plausible and near-native poses. It improves generalization to dissimilar proteins over original diffusion-based methods and classical docking.
  Affiliations: Department of Statistics, University of Oxford Oxford OX1 3LB UK imrie@stats.ox.ac.uk.; Cambridge Crystallographic Data Centre Cambridge CB2 1EZ UK.

• 📰 LLM-powered construction of antibiotic resistance datasets and knowledge graphs for accelerated antibiotic resistance gene discovery
  Yang, Ziyi, Ren, Jie, ..., Liu, Wei, Ng, Man Tat Alexander — npj Antimicrobials and Resistance, 2026-10-22
  An LLM-powered pipeline mines >8,570 papers to build AResDB/AResKG (52M triples) of antibiotic resistance entities and relations. The resulting KG measurably boosts ML prediction of gene–antibiotic resistance, enabling scalable ARG discovery.

• 📰 Multimodal alignment improves generalizability of genomic biomarker prediction in computational pathology.
  Ekaterina Redekop, Eric Zimmermann, Ava P Amini, Alex X Lu, Neil Tenenholtz, James Hall, Lorin Crawford, Kristen A Severson — Cell reports methods, 2026-10-19
  MARBLE contrastively aligns histopathology representations with genomic biomarker embeddings from LLM/PLM, enabling data-efficient generalization to new biomarkers. Trained over >40k patients, it improves zero-/few-label biomarker prediction in computational pathology.
  Affiliations: Microsoft Research, Cambridge; Microsoft Research, Cambridge

• 📰 DNA augmentations for semi-supervised learning in genomic sequence classification.
  Hyunwoo Yoo, Gail L Rosen — NAR genomics and bioinformatics, 2026-12-01
  Empirical weak/strong augmentation recipes tailored to DNA (short masking + modest indels) stabilize pseudo-label SSL. They improve AMR gene-class and oncovirus-family classification under limited labels.
  Affiliations: Department of Electrical and Computer Engineering, Drexel University

• 📰 Sampling protein language models for functional protein design.
  Jeremie Theddy Darmawan, Yarin Gal, Pascal Notin — Cell systems, 2026-10-21
  Multi-mutation sampling strategies for protein LMs outperform single-mutant NLP analogs, better capturing epistasis for functional design. Benchmarks and a binder design validate practical gains for protein engineering.
  Affiliations: Department of Biochemistry, Yong Loo Lin School of Medicine; Department of Computer Science, University of Oxford; ...

• 📰 Protein-Protein Interface Residue Prediction by Integrating Protein Language Representations and Structural Graph Attention.
  Zhen Hou, Wei He, Weinan Cao, Jiarui Li, Hongquan Li, Shuo Sun, Hua Yang — ACS omega, 2026-10-20
  EvoStruct-GAT combines ESM-2 language features, DSSP structure, and distance-aware graph attention to rank partner-agnostic protein–protein interface residues. Under homology-robust splits, it improves AUROC/AUPRC and provides interpretable residue saliency.
  Affiliations: School of Information Science and Engineering, Shanxi Agricultural University; Shanxi Key Lab for Modernization of TCVM, College of Veterinary Medicine; ...

• 📰 Hierarchically Gated Multiplex Heterogeneous Graph Neural Network with Self-Distillation for Protein Stability Prediction.
  Shugang Zhang, Nianwen Xing, Wenxu Wang, Zhiqiang Wei — IEEE journal of biomedical and health informatics, 2026-10-21
  HierGate models multiplex heterogeneous residue relations with hierarchical gating and self-distillation to predict mutation-induced protein stability changes. It outperforms prior methods and aids cancer-mechanism interpretation and directed evolution.

• 📰 Modeling microbiome modulation of tumor metabolic networks to predict synergistic therapies.
  Annie J Badenoch, Zeyang Pang, Carolina H Chung, Aaron Robida, ..., Ritish Natesan, Layth Kakish, Jiahe Li, Sriram Chandrasekaran — Cell reports methods, 2026-10-19
  A microbiome-aware ML + genome-scale metabolic modeling framework predicts drug synergies in colorectal cancer under pathogen/probiotic co-culture. It uncovers Fn-dependent pathway liabilities validated by assays and targeted perturbations.
  Affiliations: Gilbert S. Omenn Department of Computational Medicine and Bioinformatics, University of Michigan; Department of Biomedical Engineering, University of Michigan; ...

• 📰 GONet: A Molecular Interaction-Guided Graph Learning Framework for Multi-Omics Cancer Classification.
  Yi-Yang Chen, Yun-Ning Lu, Da Qi, Yong Bai, Gui-Xue Hou, Si-Qi Liu — GigaScience, 2026-10-22
  GONet is a hierarchical GNN integrating mRNA/miRNA/protein profiles with interaction networks, using DGI and gated fusion to classify cancer subtypes/stages. It achieves strong AUROC across 10 cancer types and generalizes to CPTAC proteomics.

• 📰 scHGNN: A single-cell gene regulatory network inference method based on heuristic graph neural networks
  Zhang, Tianjiao, Teng, Xiaopeng, ..., Dong, Benzhi, Wang, Guohua — Functional & Integrative Genomics, 2026-10-23
  scHGNN infers single-cell GRNs by balancing local and global topology via heuristic graph neural modules, improving link prediction on diverse gold standards. It validates TF targets (e.g., TP53/OTX2/FOXP1) with high hTFtarget support.

• 📰 SMILES-Based Deep Learning Enables Discovery of Chemically Modified Ultrashort Antimicrobial Peptides.
  Gang-Ao Hu, Piao-Ru Wu, Meng-Xin Li, Shuo Che, Mohamed Seif, Mahmoud Emam, Hong Wang, Bin Wei — Journal of medicinal chemistry, 2026-10-22
  SMAMP encodes peptides as molecular graphs from SMILES to capture noncanonical residues and modifications, enabling AMP design beyond 20 amino acids. It designs ultrashort, modified peptides with potent antibacterial activity and low cytotoxicity.
  Affiliations: College of Pharmaceutical Science & Collaborative Innovation Center of Yangtze River Delta Region Green Pharmaceuticals, Zhejiang Key Laboratory of Green; Food Toxicology and Contaminants Department, Food Industries and Nutrition Research Institute; ...

🧬 Preprints (arXiv + bioRxiv)

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

• 📄 RiboBridge: protein-conditioned de novo RNA binder structure design
  Zhou Zhang, Hanqun Cao, Minghao Sun, Odin Zhang, Yang Zhang, Pheng Ann Heng, Wujun Li, Tianfan Fu — arXiv, 2026-10-22
  Introduces a structure-first generative framework that decouples RNA conformational generation from protein–RNA assembly using cross-frame interaction geometry, enabling de novo design of protein-conditioned RNA binders that best recover native complexes on a held-out benchmark.

• 🧬 Reconstructing human gene regulatory evolution using deep learning
  Yamada, T.; Yuan, X.; Zeberg, H.; Soltys, V.; Lana Alberro, M.; Sarropoulos, I.; Paabo, S.; Kaessmann, H. — bioRxiv, 2026-10-23
  Uses a sequence-to-function deep learning model (AlphaGenome) to predict cis-regulatory effects of ~17M human-lineage substitutions across 305 tissues, mapping gains that explain human gene expression evolution and linking changes to TF motif creation and archaic hominins.
  Affiliations: Center for Molecular Biology of Heidelberg University (ZMBH)

• 🧬 GeoSinkhorn Flow: Geometry-Aware Flow Matching for Perturbation-Conditioned Dynamics in Single cell Data Phenoscapes
  XU, K.; Wilkinson, A.; Shuang, N.; Huguet, G.; Gerstein, M.; Wolf, G.; Tape, C. J.; Krishnaswamy, S. — bioRxiv, 2026-10-22
  Proposes GeoSinkhorn Flow, a geometry-aware, perturbation-conditioned flow-matching model that converts static phenoscapes into dynamic, geodesic-consistent trajectories, generalizing to held-out drugs, genes, and time points in single-cell datasets.
  Affiliations: Yale University

• 📄 Multi-Objective Protein Optimization via Modular Diffusion Steering
  Meir Roketlishvili, Felix Huber, Wenda Chu, Yisong Yue, Raul Astudillo — arXiv, 2026-10-20
  Presents MODS, an inference-time diffusion steering framework with modular scalarized rewards and ensemble surrogates to discover Pareto-optimal protein sequences, achieving strong hypervolume gains on experimental protein optimization tasks.

• 📄 recAMPLIFY: Large-Scale Pretraining of Recursive Protein Language Models
  Saurav Jha, Qianli Liao, Robert Vernon, Sarath Chandar, Quentin Fournier — arXiv, 2026-10-19
  Introduces recAMPLIFY, a recursive masked protein language model with weight-tied cycles and adaptive gating (GMR) that improves contact prediction and downstream protein tasks at fewer parameters via test-time depth adaptation.

• 🧬 Viral host assignment across the tree of life with VHAMSTeR
  Grigson, S. R.; Neri, U.; Roux, S. — bioRxiv, 2026-10-24
  Combines fine-tuned genomic language model embeddings with genome architecture and homology features to classify viral hosts across five domains, outperforming prior methods and revealing novel RNA bacteriophages and ecological patterns.
  Affiliations: DOE Joint Genome Institute

• 📄 NUViT: Vision Transformers for Fourier Space Data
  Jeffrey Gu, Robert C. Heeter, Ellen D. Zhong — arXiv, 2026-10-23
  Redesigns Vision Transformers for Fourier-space data (NUViT) via non-uniform tokenization and positional encoding, improving MRI annotation and cryo-EM protein heterogeneity analysis over standard backbones.

• 📄 Curriculum Learning for Generative Flexible Docking
  Wenyin Zhou, Hossein Azizpour — arXiv, 2026-10-20
  Shows that curriculum learning over apo-to-holo conformational shifts markedly boosts generative flexible docking with diffusion/flow models, yielding higher top-1 pose success and PoseBusters validity on PDBBind benchmarks.

• 🧬 ConforFlux: Particle-Guided Trunk Repulsion for Diverse Protein Conformations
  Suzuki, S.; Amagasa, T. — bioRxiv, 2026-10-19
  Introduces ConforFlux, an inference-time repulsion scheme that couples parallel protein-structure prediction trajectories to diversify conformations, improving multi-state coverage across Boltz-2, OpenFold3, and Protenix without retraining.
  Affiliations: University of Tsukuba

• 📄 AI Mindmeld: A Latent Interconnect Protocol
  Siavash Golkar, Jake Kovalic, Shirley Ho — arXiv, 2026-10-21
  Learns lightweight connectors (LIP) that inject specialist protein and small-molecule representations directly into frozen LLM internal states, enabling zero-shot, compositional scientific reasoning across new biochemical tasks.

• 🧬 Sequence architecture of RNA reveals key molecular signatures driving RNA phase separation
  Chatterjee, S.; Bahadur, R. P. — bioRxiv, 2026-10-24
  Builds an interpretable hybrid model combining handcrafted RNA descriptors with RNA-FM embeddings and stacking to predict LLPS-associated RNAs (AUC-ROC 0.935), revealing sequence signatures mechanistically linked to condensate formation.
  Affiliations: Indian Institute of Technology Kharagpur

• 🧬 DREM: integrating resistance and domain probabilities to prioritize plant immune receptors
  Dublino, R.; Carandente, V.; Sperli, G.; Ercolano, M. — bioRxiv, 2026-10-22
  Develops DREM, an attention-equipped bidirectional LSTM framework that classifies plant resistance proteins and their domains from sequence alone, recovering atypical receptors missed by motif-based tools and ranking candidates genome-wide.
  Affiliations: University of Naples "FedericoII"

• 📄 Spatially Informed and Direction Sensitive Representation Learning for Spatial Transcriptomics
  Azim Dehghani Amirabad, Ali Khodabandeh Yalabadi, Artem Moskalev, Yun-Ching Chen, Mehdi Pirooznia, Scott Oloff, Rui Liao — arXiv, 2026-10-20
  Introduces Tessera, a spatial transcriptomics representation learner that couples spatially informed gene selection with direction-aware, translation-invariant neighborhood encoding, improving generalization across tissues and technologies.

• 📄 VGL-stGen: Learning a shared multimodal model for lung histology and spatial transcriptomics
  Haenara Shin, Dongjoo Lee, Sungwoo Bae, Kwon Joong Na, Daeseung Lee, Hongyoon Choi — arXiv, 2026-10-20
  Trains a shared vision–language model on paired lung histology and spatial transcriptomics with single-cell references to support 13 tasks, enabling conditional gene expression completion and cell-type labeling from H&E.

• 📄 ReMedI: Inference-Time Refinement and Self-Distillation for Controllable Molecular Optimization
  Jia Yi Zhang, Shivanshu Shekhar, Allison McLevige, Tong Zhang — arXiv, 2026-10-23
  Proposes ReMedI, an inference-time Metropolis–Hastings refinement with self-distillation that turns molecular language models into controllable, multi-objective lead optimizers, beating RL baselines on property improvements with minimal retraining.

• 📄 QProtFuse: Query-Guided Layer Fusion for Retrieval-Augmented Protein Function Prediction
  Mianzhi Pan, Peishuo Liu, Yawen Ouyang, Botian Wang, Yiming Rong, Hao Zhou, Jianbing Zhang — arXiv, 2026-10-19
  QProtFuse adaptively fuses multi-layer protein language model features via query-guided weighting and joint contrastive training, boosting retrieval of proteins and annotations and improving RAG-based natural-language function descriptions.

• 🧬 Do protein representations encode variant function? Diagnosing residue-level protein attributions via embedding-space gradients
  Barghout, R. A.; Wu, S.; Chau, W.; Barraza-Chavez, J. M.; Driza, P.; Mahadevan, R.; Sanchez-Lengeling, B. — bioRxiv, 2026-10-22
  Introduces IGLOO, a diagnostic framework using embedding-space gradients and synthetic landscapes to assess residue-level attributions of protein LM fitness predictors, identifying training choices that significantly improve attribution fidelity.
  Affiliations: University of Toronto

• 📄 Escaping the Phylogenetic Trap: Persistent Homology Enables Cross-Species Ion Channel Classification via Multimodal Graph Neural Networks
  Arezoo Bybordi, Favour Achimba, Mariam Gelashvili, Jessy Ramirez, Weigang Qiu, Mandë Holford, Anita Raja — arXiv, 2026-10-21
  Demonstrates that persistent-homology features fused with ESM-2 and a disulfide-aware GNN enable cross-species ion-channel target classification of venom peptides under leave-one-taxon-out splits, outperforming other geometric priors.

• 📄 WL-SeqVAE: Sequence-Based Graph Generation Using Variational Autoencoders and Weisfeiler-Leman Colors
  Iakovos Evdaimon, Giannis Nikolentzos, Achilleas Tsimichodimos, Christos Xypolopoulos, Michalis Vazirgiannis — arXiv, 2026-10-23
  Maps graphs to Weisfeiler–Leman color sequences for VAE-based autoregressive generation, achieving competitive molecular quality while avoiding quadratic adjacency parameterization.

• 📄 Trajectory Matching: Discrete Flow Matching with Structured Actions
  Niklas Kemper, Marten Lienen, David Lüdke, Stephan Günnemann — arXiv, 2026-10-23
  Trajectory matching conditions discrete flow matching on full action sequences, enabling non-factorized, structured edits such as fragment-based molecule generation and improving molecular validity and transfer over factorized DFMs.