Weekly BioML Digest [November 17, 2026]

Share
Weekly BioML Digest [November 17, 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)

1410 matched filters -> 20 selected after LLM relevance + novelty ranking.

  • 📰 The deep learning framework Hist2Prot infers single-cell spatial proteomics from histology to predict biomarkers and clinical outcomes
    Fan, Jiansong, Yang, Sen, ..., Pan, Xiang, Yin, Shanye — Nature Cancer, 2026-11-13
    Trains a deep model (Hist2Prot) to infer single-cell spatial proteomics directly from H&E whole-slide images, improving biomarker and outcome prediction across seven cohorts for precision oncology.

  • 📰 PIANO enables generative modeling of gene expression and scales multi-omics integration to 100 million cells.
    Ning Wang, Christopher Cardenas, Victor E. Nieto-Caballero, David Turner, ..., Z. Yao, Changkyu Lee, Jonathan W. Pillow, F. Krienen — Cell reports methods, 2026-11-12
    Introduces PIANO, a probabilistic generative autoencoder for single-cell multi-omics integration that scales to 100M cells and models covariate effects, enabling cross-species and spatial integration and large-scale perturbation analysis.

  • 🌿 Genetic variants shape single-cell chromatin accessibility and disease risk
    Xue, Angli, Fan, Jianan, ..., MacArthur, Daniel G., Powell, Joseph E. — Nature, 2026-11-12
    Maps 3.5M single-cell caQTLs and uses a graph neural network to infer peak–gene links, revealing cell-type/state-specific regulatory effects that improve disease-risk interpretation over eQTLs alone.

  • 📰 Deep learning of tautomer stability from crystallographic proton positions.
    Xiaolin Pan, Chao Han, Fengyang Han, Yingkai Zhang — Chemical science, 2026-11-11
    Trains a 2D graph neural network on 1.1M crystallographic proton placements to predict stable tautomers from topology alone, enabling rapid tautomer assignment and improved ligand annotation at scale.
    Affiliations: Department of Chemistry, New York University New York 10003 USA Yingkai.zhang@nyu.edu.; Simons Center for Computational Physical Chemistry at New York University New York New York 10003 USA.

  • 📰 PFM: perturbed flow matching for structure-based drug design.
    Yankai Yu, Guikun Xu, Zhuyang Xie, Yan Yang, Yongquan Jiang, Xiaobo Zhou, Kang Li — Bioinformatics (Oxford, England), 2026-11-10
    Proposes Perturbed Flow Matching for 3D pocket-conditioned molecule generation, coupling atom-type and coordinate paths to achieve state-of-the-art docking scores for structure-based drug design.
    Affiliations: School of Computing and Artificial Intelligence, Southwest Jiaotong University; School of Artificial Intelligence, Shanghai Jiao Tong University; ...

  • 📰 Pharmacophore-Conditioned Diffusion Model for Ligand-Based De Novo Drug Design.
    Amira Alakhdar, Barnabas Poczos, Newell Washburn, Kevin Cusack — Journal of chemical information and modeling, 2026-11-09
    PharmaDiff is a pharmacophore-conditioned 3D diffusion model that generates ligand structures aligning with predefined 3D hypotheses, outperforming ligand-based baselines and improving docking without target structures.
    Affiliations: Department of Chemistry, Carnegie Mellon University; Department of Machine Learning, Carnegie Mellon University; ...

  • 📰 Benchmarking and Experimental Validation of Machine Learning Strategies for Enzyme Engineering.
    Zishuo Zeng, Jiao Jin, Rufang Xu, Yuzhen Wu, Xiaozhou Luo — Journal of chemical information and modeling, 2026-11-09
    Establishes EnzyArena, a realistic benchmark for enzyme engineering showing zero-shot protein LMs generalize better than supervised kinetic predictors, with prospective wet-lab validation favoring zero-shot designs.
    Affiliations: State Key Laboratory of Quantitative Synthetic Biology, Shenzhen Institute of Synthetic Biology; Synceres Biosciences Co. Ltd. , 1100 Huanli Road; ...

  • 📰 Direct Identification and Quantification of Peptide Biomarkers in 20% Human Serum Using a Solid-State Nanopore Platform and Machine Learning.
    Yuhua Cai, Yuhan Liu, Ismahane Khaled, Benjamin Cressiot, Laurent Bacri, Juan Pelta — Journal of the American Chemical Society, 2026-11-11
    Combines polymer-engineered solid-state nanopores with machine learning to directly identify and quantify closely related peptide biomarkers in 20% human serum, resolving bradykinin-family isoforms at single-molecule level.
    Affiliations: Université Paris-Saclay, Univ. Évry; Université Paris-Saclay, Univ. Évry

  • 📡 Absolute quantification and profiling of cellular tRNAs by nanopore sequencing and a deep learning classifier
    Ueda, Hiroki, Maeda, Mai, ..., Miyauchi, Kenjyo, Suzuki, Tsutomu — Nature Communications, 2026-11-09
    Develops a deep learning classifier on raw nanopore RNA signals to accurately identify tRNA species and detect modifications, enabling absolute quantification and single-molecule modification calling without basecalling.

  • 📰 DRAGON-3D: an electron-density-aware generative model for dual-target drug molecule generation.
    Zhong-Xing Zhou, Hong-Xing Zhang, Shi Feng, Yan Zhao, Zhen-Yang Liu, Jun-Bao Huang, Qingchuan Zheng — Physical chemistry chemical physics : PCCP, 2026-11-11
    Introduces DRAGON-3D, a dual-target de novo generator that fuses 3D electron-density representations in a bi-Net β-VAE to produce generalizable dual-target ligands using only single-target data.
    Affiliations: Institute of Theoretical Chemistry, College of Chemistry; Institute of Frontier Medical Science, School of Pharmaceutical Sciences

  • 📰 Exploring transition metal complexes with large language models.
    Yunsheng Liu, Jacob W Toney, Joseph M Cavanagh, Kunyang Sun, ..., Paul A Graggs, Dean Toste, Heather J Kulik, Teresa Head-Gordon — Chemical science, 2026-11-11
    Fine-tunes Llama-3.2 with coordination-aware SMILES to create TMC-Llama, an LLM that generates novel, physically realistic transition-metal complexes validated by DFT and experimental-like distributions.
    Affiliations: Kenneth S. Pitzer Theory Center, University of California Berkeley CA 94720 USA thg@berkeley.edu.; Department of Chemistry, University of California Berkeley CA 94720 USA.; ...

  • 📰 Evo2VED: a database of DNA large language model-derived scores for variant interpretation.
    Rui Zhu, Xiaopu Zhou, Ivy Liang, Marla Mendes, ..., Haixu Tang, Stephen J O'Brien, Stephen W Scherer, Lucila Ohno-Machado — Nucleic acids research, 2026-11-13
    Releases Evo2VED, a public database of genomic DNA-LM variant effect scores with tools linking scores to GWAS, annotations, and trait heritability, supporting large-scale noncoding variant interpretation.
    Affiliations: Department of Biomedical Informatics & Data Science, School of Medicine; The Centre for Applied Genomics, The Hospital for Sick Children; ...

  • 📰 Tuned Fitness Landscapes for Benchmarking Model-Guided Protein Design.
    Neil Thomas, Atish Agarwala, David Belanger, Yun S Song, Lucy J Colwell — Genetics, 2026-11-14
    Presents SLIP, a framework to build tunable synthetic protein fitness landscapes from Potts models for benchmarking ML-guided directed evolution under controllable difficulty.
    Affiliations: University of California, Berkeley; Google DeepMind, Cambridge; ...

  • 📰 GPCRAgent: an intelligent multi-agent platform integrating knowledge graphs and large language models for GPCR-targeted drug discovery.
    Wen-Ke Xiao, Qing-Qing Zhou, Zhi-Wen Yang, Zhi-Xiang Wu, ..., Zhi-Hui Zhao, De-Xian Wang, Chong-Yang Shen, Wei Chen — Journal of advanced research, 2026-11-10
    GPCRAgent integrates knowledge graphs with multi-agent LLM reasoning to prioritize GPCR targets and ligands with traceable paths, supported by docking/MD and organoid evidence (e.g., LPAR1 fibrosis case study).

  • 📰 HEIWA: Hybrid Explicit-Implicit Water Attention Affects Force Field Accuracy in Structure-Based Protein Language Model ESM3.
    Ngai Hei Ernest Ho, Chao Tung Yang, I-Son Ng, Jo-Shu Chang — Journal of chemical information and modeling, 2026-11-09
    HEIWA augments ESM3 with hybrid explicit–implicit water attention, improving residue-level biophysical consistency and generalization (e.g., thermophilicity patterns) by injecting solvent-aware interactions.
    Affiliations: Department of Chemical Engineering, National Cheng Kung University; Department of Computer Science, Tunghai University; ...

  • 📰 EZpred: Improving Deep Learning‐Based Enzyme Function Prediction Using Unlabeled Sequence Homologs
    Cheng-Xin Zhang, Quan-Cheng Liu, L. Freddolino — Advanced Science, 2026-11-10
    EZpred leverages unlabeled homolog sequence features from ESM-C to boost deep learning EC-number prediction, achieving large gains for enzymes with <30% identity to training data.

  • 📰 ThermoMutDB 2.0: semiautomated literature curation and structural enrichment of thermostability data for protein mutations.
    Gratchela Dutra Rodrigues, Alisson Clementino-Silva, Taffarel Brant-Ribeiro, Maria Luiza Martins De Mendonça, ..., A. Góes-Neto, D. Pires, F. S. Kremer, Joicymara Santos Xavier — Nucleic acids research, 2026-11-10
    ThermoMutDB 2.0 introduces LLM-assisted literature triage and mutant structural enrichment for curated protein thermostability measurements, enabling scalable, structure-contextualized stability data.

  • 📰 Instrument-Agnostic Machine Learning Framework for Accurate Prediction of Functional Groups from Tandem Mass Spectrometry.
    Connor Beveridge, Prageeth R Wijewardhane, Matthew Muhoberac, Nicolás M Morato, ..., Jeffrey R Gilbert, Christopher J Welch, Shane Tichy, Gaurav Chopra — Analytical chemistry, 2026-11-10
    Builds instrument-agnostic ML models that predict functional groups directly from MS/MS spectra, transferring across platforms without database searches to accelerate structural elucidation.
    Affiliations: Department of Chemistry, Purdue University; Lilly Oncology @ Lilly, Lilly Oncology; ...

  • 📰 Biologically informed Structure–Gene–Process integration for anticancer drug combination synergy prediction
    He, Xuan, Du, Huanhuan, ..., Chen, Xingchi, Zhu, Hongbo — Journal of Cheminformatics, 2026-11-13
    SGP-Syn links drug structure to cell-line gene expression and GO processes via a Structure–Gene–Process deep model, improving drug-combination synergy prediction with multilevel biological interpretability.

  • 📰 Handshake: Partner-specific protein-protein binding site prediction at scale using ProstT5 and cross-chain attention.
    Nurit Haspel — PLoS computational biology, 2026-11-13
    HandShake is a sequence-only, partner-specific PPI interface predictor using ProstT5 embeddings, LoRA, and cross-chain attention, achieving strong interface/contact detection without 3D inputs.
    Affiliations: Department of Computer Science, University of Massachusetts Boston

🧬 Preprints (arXiv + bioRxiv)

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

  • 🧬 Biomolecular design through latent reasoning with RFD4-Proteina
    Corley, N.; Didi, K.; Kazman, M. W.; Brent, R. I.; ...; Li, Y.; Li, Z. M.; Muraishi, A.; Muratspahic — bioRxiv, 2026-11-15
    Introduces a physically grounded, multimodal latent generative model with intrinsic/extrinsic reasoning that designs proteins, binders, glycans, and nucleic acids; demonstrates state-of-the-art in silico and experimental successes including a first de novo enzyme for API synthesis.
    Affiliations: Institute for Protein Design, University of Washington; Department of Biochemistry, University of Washington; ...

  • 📄 ProTEUS: A mutation-resolved world model for closed-loop protein evolution
    Yikun Zhang, Minghao Guo, Guangxue Xu, Peter Yichen Chen, Wojciech Matusik, Wengong Jin, Jingjing Chai, Kaiyi Jiang — arXiv, 2026-11-09
    Presents ProTEUS, a mutation-resolved world model using a MutationFormer to predict functional effects and plan mutations in closed loop; experimentally evolves enzymes to >10× activity with a fast physics-constrained structure predictor.

  • 🧬 DeepRNAi: Predicting on- and off-target siRNA activity and establishing a benchmark
    Masse, N. Y.; Lau, T. T. Y.; Wintersinger, J.; Slobodyanyuk, M.; ...; Lazar, N. H.; Findlay, S. D.; Celaj, A.; Frey, B. J. — bioRxiv, 2026-11-13
    DeepRNAi is a deep model trained on 16M siRNA–transcript pairs that predicts on- and off-target knockdown from sequence, approaching experimental accuracy and generalizing across species for safer RNAi therapeutic design.
    Affiliations: Deep Genomics Inc.

  • 🧬 Joint protein-DNA language modeling enables prediction and engineering of DNA recognition specificity
    Xiong, Y.; Ren, J.; Wei, X.; Li, X.; Li, J.; Wei, W. — bioRxiv, 2026-11-10
    Joint protein–DNA language models pretrained on paired sequences enable zero-shot and fine-tuned prediction of DNA-recognition specificity, mutation effects, and motif inference directly from protein sequences, supporting genome editing and TF engineering.
    Affiliations: Peking University

  • 📄 CupOFMoCA: Coupled Objective-Guided Discrete Flows for Molecular Conjugate Assembly
    Ruoxi Rosie Zhang, Ziang Li, Jiatao Gu, Pranam Chatterjee — arXiv, 2026-11-11
    CupOFMoCA uses objective-guided discrete flows to co-generate coupled components of molecular conjugates (e.g., PROTACs, PDCs) under assembly constraints, improving conjugation validity and predicted activity over staged pipelines.

  • 🧬 Vermeer: Autoregressive generative modeling of microscopy predicts protein localization
    Kambhampati, S.; Zimmermann, E.; Hayir, E.; Yang, K. K.; Chen, F.; Lu, A. X. — bioRxiv, 2026-11-12
    Vermeer is a channel-adaptive autoregressive generator that synthesizes protein-localization microscopy images conditioned on protein sequence and cellular morphology, transferring zero-shot across imaging channel configurations.
    Affiliations: Harvard University

  • 🧬 Leveraging state-of-the-art AI structure predictors to generate atomic resolution ensembles of intrinsically disordered proteins
    Schnapka, V.; Portal, N.; Bonomi, M. — bioRxiv, 2026-11-11
    bAIes leverages AI-predicted inter-residue distance distributions from multiple structure predictors to generate atomistic ensembles for intrinsically disordered proteins that match NMR/SAXS, bridging DL signals with physics-based sampling.
    Affiliations: Institut Pasteur

  • 🧬 MicroDUET: a dual-headed genomic foundation model for pathogen surveillance
    Barber, M.; Suster, C. J. E.; Lei, J.; Getchell, M.; Wulandari, S.; Sintchenko, V.; Pronyk, P. M. — bioRxiv, 2026-11-13
    MicroDUET adapts a genomic foundation model with dual adapter heads to disentangle phylogeny and function, enabling fast lineage/clade retrieval and resistance prediction for scalable pathogen surveillance.
    Affiliations: University of Oxford

  • 🧬 Can Folding Models Tell Binders from Bluffers? Evidence from POISK: The Patent-Derived Antibody Dataset
    Tupikina, D.; Roncoli, A.; Bujotzek, A.; Abanades, B. — bioRxiv, 2026-11-14
    Introduces POISK, a large patent-derived antibody–antigen dataset, and benchmarks modern folding models for binder discrimination, revealing useful yet limited screening power and releasing a community resource for method development.
    Affiliations: Prescient Design - AI for Drug Discovery, Computational Science Center Of Excellence

  • 📄 Binding Mode Matters: Hotspot-Aware Drug Discovery via Explorative Preferences
    Dingshuo Chen, Hao Yang, Kuangqi Zhou, Zhixun Li, Qiang Liu, Xiangxiang Zeng, Shu Wu, Yaosen Min — arXiv, 2026-11-11
    BindMol frames target-based drug design as multi-objective RL over protein interaction hotspots, discovering diverse high-affinity molecules with distinct binding modes beyond scalar docking optimization.

  • 📄 Mechanism-Level Chemical Reaction Simulation with Electron Bookkeeping Transformer
    Sarah Cao, Shitong Luo, Connor W. Coley — arXiv, 2026-11-12
    Electron Bookkeeping Transformer models reaction mechanisms by tokenizing electrons and generating arrow-pushing trajectories, improving elementary-step and pathway prediction while enforcing mass/charge conservation.

  • 🧬 Disentangled generative modeling of single-cell perturbation responses across dynamic cellular states
    Wang, Y.; Zhao, Y.; Li, S.; Yang, D. — bioRxiv, 2026-11-10
    PertVI is a disentangled deep generative model for single-cell perturbation data that separates baseline state, batch, and treatment effects, enabling accurate recovery of state-gated transcriptional responses and phenotypic shifts.
    Affiliations: UNIVERSITY OF PITTSBURGH

  • 🧬 RNA Small-Molecule Binding-Site Prediction by Integrating RNA Language Models and Structural Geometry
    Sharma, N.; Mitra, P. — bioRxiv, 2026-11-11
    RSMBind integrates pretrained RNA language-model embeddings with graph geometric features to predict RNA small-molecule binding sites at nucleotide resolution, outperforming prior methods on multiple benchmarks.
    Affiliations: Indian Institute of Technology, Kharagpur

  • 🧬 Unsupervised Deep Learning for Cryo-EM/ET Particle Classification
    Uddin, M. R.; Min, S.; Wu, Q.; Zhan, X.; Chen, M.; Xu, M. — bioRxiv, 2026-11-13
    An unsupervised autoencoder framework jointly learns pose and structure to cluster cryo-EM/ET particle images, separating protein classes and contamination without labels and detecting native complexes in tomograms.
    Affiliations: Carnegie Mellon University

  • 🧬 Rapid generation of antibody conformational ensembles using machine learning
    Rao, P. V.; Sommer, D.; Davidson, D.; Alberstein, R. G.; ...; Bonneau, R.; Mahajan, S. P.; Watkins, A. M.; Dreyer, F. A. — bioRxiv, 2026-11-09
    Cadense is a generative model that produces antibody Fv conformational ensembles from sequence, recovering apo/holo states and improving docking by capturing CDR loop dynamics at orders-of-magnitude lower cost than MD.
    Affiliations: Genentech

  • 📄 Pyrite: Building and Benchmarking Protein-Ligand Pose Ranking Agents
    Megha Srivastava, Michael H. Lim, Allison M. Keys, Pranav Murugan, Nate Gruver — arXiv, 2026-11-10
    Pyrite builds a pose-ranking benchmark and LLM-based agents that synthesize and select interpretable features to compare candidate protein–ligand poses, surpassing expert chemists and prior automated baselines.

  • 📄 HyPLM Technical Report
    Haowei Lin, Cai Zhou, Xiangxin Zhou — arXiv, 2026-11-09
    HyPLM-Flash is a 650M-parameter protein language model that matches or exceeds larger models on mutation and contact tasks while running 2× faster, advancing the Pareto frontier for PLMs in accuracy vs. efficiency.

  • 📄 HyInv Technical Report
    Haowei Lin, Cai Zhou, Xiangxin Zhou — arXiv, 2026-11-09
    HyInv is a lightweight graph-based inverse-folding model with a fast sampler that designs diverse, well-folded sequences for given backbones, outperforming ProteinMPNN-class baselines at >50× speed.

  • 📄 MutAtlas: A PDB-Wide Energy-Guided Atlas of Protein Mutation Effects
    Ruihan Guo, Chaoran Cheng, Zhanghan Ni, Neil He, Bangji Yang, Ge Liu — arXiv, 2026-11-12
    MutAtlas compiles a PDB-wide atlas of single-site mutation effects by aligning signals from physics-based, language, and inverse-folding models and proposes unsupervised multi-source distillation that improves zero-shot variant prediction.

  • 🧬 DeePEn - A Depth sensitive benchmark for Protein Engineering
    Schmirler, R.; Brenner, M.; Franz, S.; Heinzinger, M.; Rost, B. — bioRxiv, 2026-11-12
    DeePEn introduces a depth-sensitive benchmark evaluating variant-effect prediction as mutational distance increases, revealing generalization gaps across protein engineering models and providing multi-metric evaluation.
    Affiliations: Technical University of Munich

Read more