Weekly BioML Digest [October 12, 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)

1359 matched filters -> 20 selected after LLM relevance + novelty ranking.

• 📰 Validation and analysis of 12,000 AI-driven CAR-T designs in the Bits to Binders competition
  Kosonocky, Clayton W, Abel, Alex M, ..., Wesley A, Marcotte, Edward M, Bits to Binders Competitors — Molecular Systems Biology, 2026-10-08
  Benchmarks 12,000 AI-designed CD20 binders in primary CAR-T cells, linking ML design strategies and sequence features to functional T-cell outcomes. Provides a large, functionally grounded resource and practical filters to improve generative protein binder design.

• 📰 Machine learning-assisted directed evolution of plant Rubisco.
  Julie L McDonald, Jiacheng Lin, Yunlong Zhao, Brian L Hie, ..., Bryan D Bryson, Spencer M Whitney, Matthew D Shoulders, Robert H Wilson — Science advances, 2026-10-09
  Uses a structure-informed protein language model (ESM-IF1) to prioritize mutation sites for ML-guided library design of Rubisco, yielding variants with up to 43% higher aerobic carboxylation efficiency. Demonstrates ML-assisted directed evolution to enhance CO2-fixation enzymes with agronomic potential.
  Affiliations: Department of Chemistry, Massachusetts Institute of Technology; Department of Biology, Massachusetts Institute of Technology; ...

• 🛠️ Accurate and well-powered case–control analysis of spatial molecular data
  Reshef, Yakir A., Sood, Lakshay, ..., Korsunsky, Ilya, Raychaudhuri, Soumya — Nature Methods, 2026-10-06
  Introduces VIMA, an ensemble variational autoencoder framework that learns patch-level embeddings from spatial omics and identifies case–control associated ‘microniches’ with calibrated statistics. Recovers known pathology and discovers new disease-linked spatial features across modalities.

• 🖥️ Advancing chemical reaction prediction in data-scarce drug discovery with active and geometric deep learning
  Minot, Mason, Stenzhorn, Yannick, ..., Atz, Kenneth, Schneider, Gisbert — Nature Computational Science, 2026-10-09
  Combines active learning with geometric graph neural networks and self-supervised tasks to efficiently acquire reaction data and predict C–H borylation outcomes and regioselectivity under data scarcity. Prospective tests pinpointed correct borylation sites on challenging substrates, accelerating medicinal chemistry.

• 📰 Protein large language model-assisted one-to-one gene homology mapping in cross-species single-cell transcriptome integration.
  Ze-Yu Kuang, Yuan-Chen Sun, Na-Na Wei, Yu-Juan Wang, Hua-Jun Wu — Genome research, 2026-10-09
  Leverages protein LLM embeddings fused with sequence similarity under a one-to-one constraint to expand and refine cross-species gene homology mappings for single-cell transcriptome integration. Improves alignment accuracy and enables discovery of unannotated cross-species marker pairs across 11 species.
  Affiliations: Department of Biomedical Informatics, School of Basic Medical Sciences; Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education/Beijing), Peking University Cancer Hospital and Institute; ...

• 🏛️ A structure-aware generative AI framework for revealing functional relationships in protein families.
  Divyanshu Shukla, Jonathan Martin, Faruck Morcos, Davit A Potoyan — Proceedings of the National Academy of Sciences of the United States of America, 2026-10-06
  Builds parallel sequence- and structure-alphabet variational autoencoders coupled by direct-coupling energy to create generative landscapes that disentangle sequence, structure, and function in protein families. Enables information-theoretic distances and targeted sampling to guide protein design and evolutionary insight.
  Affiliations: Bioinformatics and Computational Biology Program, Iowa State University; Department of Biological Sciences, The University of Texas at Dallas; ...

• 📰 Machine learning-assisted discovery of AdMysD for enhanced porphyra-334 biosynthesis.
  Longfei Yuan, Shuting Feng, Sheng Xie, Qian Liu, ..., Liqun Li, Jieru Xie, Yuexuan Li, Nan Wang — Bioorganic chemistry, 2026-10-05
  Uses UniRep embeddings and positive–unlabeled bagging within a sequence-similarity network to prioritize rare MysD enzymes, discovering AdMysD with 3× improved catalytic efficiency for porphyra‑334 biosynthesis. Establishes a PU-learning pipeline for enzyme discovery under sparse annotations.
  Affiliations: State Key Laboratory of Ocean Sensing & Ocean College, Zhejiang University; Hainan Institute, Zhejiang University; ...

• 📰 SMILES-Based Deep Learning Enables Discovery of Chemically Modified Ultrashort Antimicrobial Peptides
  Gang Hu, Piao-Ru Wu, Meng-Xi Li, Shuo Che, M. Seif, Mahmoud Emam, Hong Wang, Bin Wei — Journal of Medicinal Chemistry, 2026-10-07
  Develops SMAMP, a SMILES/graph deep learning framework that models chemically modified and noncanonical peptides for MIC prediction and design. Guided designs yielded 13/19 active ultrashort antimicrobial peptides, including in vivo–validated leads with low cytotoxicity.

• 💻 DeepSAP: An Integrated Generative and Predictive Deep-Learning Framework for Controllable Design of Self-Assembling Peptides.
  Yuan-Ming Song, Yu-Han Wu, Yu-Qiu Jiang, Li He, Shou-Bo Zhao, Xiao-Chen Deng — Bioinformatics, 2026-10-07
  Presents DeepSAP, an end-to-end conditional generative and predictive framework that pretrains on large peptide corpora to generate self-assembling peptides with desired physicochemical properties and screens them with a lightweight predictor. Generated peptides exhibit stable self-assembly validated by simulations.

• 📡 ProFormer: generalizable classification of single-cell and plasma proteomes using deep learning
  Krull, Karl K., Kühn, Arlene, ..., Brinker, Titus J., Krijgsveld, Jeroen — Nature Communications, 2026-10-05
  Proposes ProFormer, a transformer that directly ingests MS1-level peptide-ion features to classify samples without peptide/protein identification. Outperforms baselines across diverse LC‑MS datasets for single-cell and plasma proteome classification with explainable peptide-level attributions.

• 📰 Pretrained spectrum representations for diverse tandem mass spectrometry proteomics tasks.
  Justin A. Sanders, Melih Yilmaz, Jacob H. Russell, Wout Bittremieux, William E. Fondrie, N. Riley, Sewoong Oh, W. Noble — Molecular & cellular proteomics : MCP, 2026-10-07
  Introduces CasanovoTL, transferring pretrained de novo sequencing spectrum encoders to diverse tandem MS tasks. Pretrained representations and finetuning improve spectrum quality, chimericity, phosphorylation, and glycosylation predictions under limited data.

• 📰 Bioactivity-driven prediction of antibacterial synergy using machine learning models.
  Hannie Yousefabadi, Mahya Mehrmohamadi — PLoS computational biology, 2026-10-01
  HALO encodes drug pairs via multi-level Chemical Checker bioactivity signatures and uses rigorous drug-pair held-out validation to predict antibacterial synergy. Generalizes to unseen combinations and transfers across datasets, clarifying realistic performance under leakage-free evaluation.
  Affiliations: Department of Molecular Biotechnology, School of Biotechnology

• 📰 Robust prediction of drug combination side effects in realistic settings.
  Ruben Jimenez, Alberto Paccanaro — PLoS computational biology, 2026-10-05
  DCSE learns latent signatures for drugs, pairs, and side effects to predict polypharmacy adverse events, evaluating in realistic prospective warm- and cold-start settings. Consistently outperforms prior methods, improving robustness for clinical safety prediction.
  Affiliations: Escola de Matemática Aplicada, Fundação Getúlio Vargas; Department of Computer Science, Centre for Systems and Synthetic Biology

• 💻 Katalyst: knowledge-guided semantic alignment via contrastive learning for enzyme turnover prediction
  Rui-Lin Li, Sen-Yu Tang, Jia-Qi Deng, Tai-Xing Qiu, Fei Guo, Ji-Jun Tang, Xiao-Yi Liu — Bioinformatics, 2026-10-08
  Katalyst aligns enzyme sequences, textual descriptions, and reaction substrates/products into a shared latent space via contrastive learning with interaction-aware fusion to predict kcat. Outperforms state-of-the-art baselines and generalizes across sequence scales for enzyme kinetics.

• 📰 PNBind: Prediction of Nucleic Acid-Binding Sites Using Protein Structure and Protein Language Models.
  Yun-Hai Li, Guang-Hong Dang, Bo-Wen Shao, Peng-Pai Li, Zhi-Ping Liu — Proteins, 2026-10-08
  PNBind integrates ESM protein language embeddings, evolutionary signals, and 3D geometric features via a GVP backbone with invariant point attention to predict nucleic acid–binding residues. Achieves strong performance across DNA/RNA datasets with ablations highlighting PLM contributions.

• 📰 Protein–Protein Interface Residue Prediction by Integrating Protein Language Representations and Structural Graph Attention
  Zhen Hou, Wei He, Wei-Nan Cao, Jia-Rui Li, Hong-Quan Li, Shu-Chen Sun, Hua Yang — ACS Omega, 2026-10-08
  EvoStruct-GAT represents proteins as residue-level graphs combining ESM-2 embeddings, DSSP features, and distance-aware graph attention to predict protein–protein interface residues without a known partner. Delivers robust AUROC/AUPRC and outperforms baselines under homology-reduced splits.

• 📰 PlantPTM: A deep-learning framework integrating protein language models with multi-view features for predicting diverse post-translational modification sites in plants.
  Gensheng Dou, Kaiyuan Wang, Zhaohui Qin, Min Zhang, Wentian Zhang, Chunbo Miao, Junzhou Li, Zhen Chen — Molecular plant, 2026-10-05
  PlantPTM fuses protein language models with evolutionary and multi-view features to predict nine plant PTM types, maintaining accuracy on low-data and unseen-species settings. Delivers state-of-the-art AUROCs and is provided as a web service for plant proteomics.
  Affiliations: State Key Laboratory of High-Efficiency Production of Wheat-Maize Double Cropping, College of Agronomy; State Key Laboratory of High-Efficiency Production of Wheat-Maize Double Cropping, College of Agronomy; ...

• 📰 Structure-aware prediction of functional non-synonymous variants in a maize EMS mutant library.
  Li Qin, Xiaolong Guo, Fan Xia, Fugui Xie, ..., Yuxuan Lou, Xiangfeng Wang, Xuerong Yang, Xiaoduo Lu — Molecular plant, 2026-10-05
  MaizeMEP integrates protein language model–based, structure-aware mutational effect predictions (SaProt, GPSite) to score over 1M missense variants from 8,556 EMS maize lines. Identifies variants that fine-tune plant height with minimal yield penalty, enabling trait optimization.
  Affiliations: Institute of Molecular Breeding for Maize, Qilu Normal University; State Key Laboratory of Maize Bio-breeding, Frontiers Science Center for Molecular Design Breeding; ...

• 📰 ProCoDiff-Edit: Protein-Conditioned Counterfactual Latent Diffusion for Structure-Constrained Drug Editing and Cold-Start Affinity Prediction
  Muhammad Hussain, Maisam Abbas, Viet Thanh Duy Nguyen, Dinh-Quy Nguyen, Quang-Thai Ho — Computers, 2026-10-09
  ProCoDiff-Edit couples a ChemBERTa–ESM-2 affinity predictor with a protein-conditioned latent diffusion model to edit molecules in SELFIES space under scaffold and property constraints. Supports cold-start affinity prediction and constrained generation with structure-aware evaluation.

• 📰 Predictive β-Sheet Protein Stabilization via Computationally Guided Metal Coordination with Simulations and AlphaFold3
  Zhuo-Jian Lu, Sheng-Bang Xu, Bin Zheng, Ao-Qi Wan, ..., Yi-Feng Tian, Hong-Jiao Liu, Yan-Wei Wang, Peng Zheng — ACS Nano, 2026-10-08
  Combines all-atom MD to locate mechanical soft spots with AlphaFold3 to assess engineered metal-chelation feasibility, guiding site-specific coordination clamps that more than double unfolding forces in β-sheet proteins. Demonstrates an AI-assisted workflow to program protein mechanical stability.

🧬 Preprints (arXiv + bioRxiv)

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

• 📄 De novo design of monoclonal and bispecific antibodies with OFAntibody
  Valhalla Team — arXiv, 2026-10-07
  Presents OFAntibody, an all-atom generative framework that jointly designs monoclonal and bispecific antibodies with multi-interface structural supervision and epitope conditioning. It achieves high hotspot/energy pass rates across bispecific formats, enabling programmable multi-target recognition.

• 🧬 Predicting viral host range for reverse zoonosis surveillance
  Irvine, E. B.; Bikias, T.; Schifferli, T.; Plath, L.; ...; Sahin, B.; Frei, L.; Taft, J. M.; Reddy, S. T. — bioRxiv, 2026-10-09
  Trains a transformer on 8.8M RBD:ACE2 interaction measurements to predict cross-species SARS-CoV-2 host compatibility from sequence with calibrated uncertainty. The model generalizes zero-shot to unseen ACE2 orthologs and prioritizes species for reverse-zoonosis surveillance.
  Affiliations: ETH Zurich

• 🧬 Autonomous AI agents discover reverse transcriptases with tandem repeat arrays
  Yoon, P. H.; Athukoralage, J. S.; Ameisen, E.; Kauderer-Abrams, E.; Perry, N. T.; Durrant, M. G. — bioRxiv, 2026-10-11
  Shows that autonomous LLM agents can mine metagenomes at scale to discover novel molecular systems, identifying a new family of array-associated reverse transcriptases by recognizing atypical repeats directly in DNA. Demonstrates LLM-driven anomaly detection as a path to biological discovery.
  Affiliations: Anthropic, PBC

• 🧬 Synthesis-aware generative design in trillion-scale chemical spaces for automated drug discovery
  Seo, S.; Sung, Y.; Hwang, S.-Y.; Kang, M.; ...; Lim, J.; Wanner, B. M.; Han, J. M.; Kim, W. Y. — bioRxiv, 2026-10-09
  Introduces a synthesis-aware generative flow-network that encodes reaction logic to navigate an 11-trillion make-on-demand space while optimizing properties and synthesis compatibility. Prospective validation with automated synthesis yields first-in-class SLC1A5 variants at high hit and synthesis success rates.
  Affiliations: Department of Chemistry, KAIST; HITS Inc., Seoul

• 📄 Linear Fitness Subspace in Protein Language Models Enables Sample-Efficient Directed Evolution
  SiYuan Ma, Canran Xiao, Zikai Xiao, Albert Gao, Liang He, Xuan-Yu Wang, Shuying Cao, Xiaojun Jia — arXiv, 2026-10-06
  Proposes the Linear Fitness Subspace hypothesis and a Subspace-Guided Evolutionary Search that learns assay-specific, low-dimensional directions in PLM representation deltas. Achieves more sample-efficient directed evolution across diverse assays by performing surrogate modeling and acquisition in the learned subspace.

• 📄 MD-LLM-2: A Transferable Language Model of Molecular Dynamics with Physical Conditioning and Explicit Path Probabilities
  Mhd Hussein Murtada, Z. Faidon Brotzakis, Michele Vendruscolo — arXiv, 2026-10-07
  Develops MD-LLM-2, a transferable sequence- and temperature-conditioned language model that generates protein conformational trajectories with explicit path likelihoods. It reproduces thermodynamic trends and predicts equilibrium/first-passage statistics, enabling likelihood-based analysis of simulated or generated paths.

• 📄 NEMORA: Neural Equivariant Multipole Operators for Long-Range Atomistic Learning
  Jay L. Kaplan, Samuel Varner, Rebecca Willett, Juan J. de Pablo — arXiv, 2026-10-07
  Introduces NEMORA, a neural equivariant extension of the Fast Multipole Method that learns long-range tensorial interactions with linear scaling. It substantially reduces force/energy errors when augmenting short-range atomistic GNNs, enabling accurate, scalable nonlocal physics in molecular simulations.

• 📄 Learning consistent molecular mechanics force fields from first principles
  Berkay Günes, Leif Seute, Jigyasa Nigam, Frauke Gräter — arXiv, 2026-10-06
  Learns bonded and nonbonded force-field parameters consistently from ab initio data with electrostatic-potential supervision and charge-equilibration architecture. Achieves state-of-the-art geometry and sampling accuracy, unifying first-principles-consistent classical force fields without external nonbonded parameters.

• 📄 RAGenome: Scaling Retrieval-Based Genomic Language Models to Long Contexts
  Frederikke Isa Marin, Panagiotis Antoniadis, Dionysia Danai Brilli, Andreas Bjerregaard, Rachael DeVries, Yan Li, Ole Winther, Wouter Boomsma — arXiv, 2026-10-08
  Presents RAGenome, a retrieval-augmented genomic language model that scales MSA-based pretraining to hundred-fold longer contexts. It improves long-range genomic tasks such as gene finding while remaining competitive on evolutionary variant prioritization.

• 📄 Atom-JEPA: Joint-Embedding Predictive Architecture for 3D Atomistic Systems
  Kasper Helverskov Petersen, Rasmus Hannibal Tirsgaard, François R J Cornet, Mikkel Jordahn, Mikkel N. Schmidt — arXiv, 2026-10-06
  Introduces Atom-JEPA, a self-supervised joint-embedding predictive architecture on 3D atomistic data with atom/substructure objectives. Pretraining transfers broadly, achieving SOTA on molecular ADMET and strong performance on materials properties.

• 📄 Neural Petri flows for chemical reactions
  Jose Eduardo Escrig Molina, Daniel Probst — arXiv, 2026-10-06
  Proposes Neural Petri Flow, which hard-wires Petri-net conservation/enabling and learns only local rate laws for reaction graphs. It attains strong atom-mapping, reaction prediction, EC classification, and electron-transfer step prediction while guaranteeing chemically valid transitions.

• 📄 TopoLIA: Topology-Inspired Latent Interaction Attention for Drug Discovery
  Xiang Liu, Yiming Ren, Mustafa Hajij, Pietro Liò, Guo-Wei Wei — arXiv, 2026-10-09
  TopoLIA introduces topology-inspired latent interaction attention that mediates source-target-dependent information exchange via a compact latent space and persistent spectral features. It achieves SOTA on protein–ligand docking/scoring, revealing compact interaction spaces as effective inductive biases.

• 🧬 REFinder: Mining new enzymes from metagenomes
  Prabakaran, R.; Bromberg, Y. — bioRxiv, 2026-10-08
  REFinder routes reads labeled by a DNA language model (REBEAN) to assemble only putative enzymatic sequences, boosting novel enzyme recovery with lower compute than full assemblies. It uncovers thousands of structurally credible, homology-invisible enzymes from metagenomes.
  Affiliations: Emory University

• 🧬 Single-Cell Behavioromics Enables Fate Prediction of Individual Hematopoietic Stem Cells
  Chankhunthod, A.; Sorenson-Graff, R.; Mercoli, J.; Ishida, T.; ...; Wellington, R.; Fong, Y.; Hadland, B.; Zhao, H. G. — bioRxiv, 2026-10-07
  Combines live-imaging ‘behavioromics’ with deep learning to predict single HSC long-term engraftment with near-single-cell precision. Links reduced motility and Activin signaling to LT-HSC fate, enabling actionable modulation and high-fidelity fate prediction.
  Affiliations: Fred Hutch Cancer Center

• 🧬 From Sparse Visits to Continuous Molecular Trajectories with Flow Matching
  Passban, P.; Gupta, S.; Guan, B.; Roosta, T. — bioRxiv, 2026-10-11
  CohortFM uses flow matching to learn continuous, patient-specific molecular trajectories from sparse longitudinal transcriptomics. It decouples biological progression from calendar time, improving interpolation/forecasting and revealing disease programs invisible to visit-level analyses.
  Affiliations: Sanofi

• 🧬 Tractable and Thermodynamically Consistent pKa Prediction via First-Dissociation Aggregation
  Liu, Z.; Zhu, Q.; Li, Q.; Zhu, Y.; ...; Dai, B.; Li, Y.; Liu, Z.; Yang, S. — bioRxiv, 2026-10-06
  Derives an exact identity for macroscopic first-dissociation pKa and couples it to a dual-head GNN that enforces thermodynamic consistency. DTi-pKa attains accurate aqueous pKa with site-resolved microstate outputs and principled macro–micro coupling.
  Affiliations: School of Pharmacy, Shihezi University

• 📄 Electron-Density-Informed Molecular Property Prediction Using Surrogate Functionals
  Manuel V. Klockow, Simon Wagner, Roman Remme, Dominic Plein, Fred A. Hamprecht — arXiv, 2026-10-09
  Uses surrogate functional constraints from orbital-free DFT to enforce that provided electron densities minimize an energy functional during property prediction. Improves generalization and SOTA accuracy for extensive properties on QM9/QM40 versus density-informed baselines without constraints.

• 📄 CircuitATLAS: Agentic reasoning over a systems neuroscience knowledge graph for target discovery in circuitopathies
  Gabriel Ocana-Santero, Marko Tvrdic — arXiv, 2026-10-07
  CircuitATLAS builds a provenance-grounded systems neuroscience KG and an agentic reasoning loop to find controllable molecular ‘control points’ for circuitopathies. It nominates and validates ATP1A3 as a cortical excitability control node, linking reasoning to in vivo tests and chemistry follow-up.

• 📄 Inference-Time Steering Across Scaffolds and Epitopes for Sample-Efficient De Novo Nanobody Design
  Mason Minot, Simon Friedensohn, Cédric R. Weber — arXiv, 2026-10-09
  Frames nanobody generation across scaffolds/epitopes as inference-time compute allocation with global SMC and BaSE-based budgeters. Dynamic steering raises yield per GPU-hour ~5× while maintaining target and configuration diversity.

• 🧬 Ab initio CryoEM maps minimizing preferred orientation problems in heterogeneous mixtures: ReconSIREN
  Herreros, D.; Sorzano, C. O. S.; Carazo, J. M. — bioRxiv, 2026-10-10
  ReconSIREN is a deep learning ab initio cryo-EM reconstruction framework that mitigates preferred orientation via a new optimization scheme and geometrical priors. Its MoE architecture yields accurate consensus maps and conformational landscapes without traditional refinement.
  Affiliations: National Centre for Biotechnology (CSIC)