Weekly BioML Digest [August 03, 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)

1802 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, 2026-07-29
  Introduces Raygun, a generative protein design framework that encodes sequences as length-agnostic probability distributions from language model embeddings, enabling controllable miniaturization/augmentation while preserving structure and function.

• 🧫 Property guidance for protein sequence generative models with ProteinGuide
  Xiong, Junhao, Gaur, Ishan, ..., Savage, David F., Listgarten, Jennifer — Nature Biotechnology, 2026-07-29
  ProteinGuide provides on‑the‑fly conditioning for pretrained protein generative models (ESM3, ProteinMPNN, diffusion/flow) to steer sequence design toward user‑specified properties, improving protein engineering and base editor optimization.

• 🧪 De novo design of potent CRISPR–Cas13 inhibitors
  Taveneau, Cyntia, Chai, Her Xiang, ..., Grinter, Rhys, Knott, Gavin J. — Nature Chemical Biology, 2026-08-01
  AI‑driven de novo protein design yields potent, specific inhibitors of Cas13a in cells, establishing a workflow for bespoke off‑switches to control RNA‑editing CRISPR systems.

• 🧪 Atomic-level protein–ligand recognition with PBCNet2.0 for probe discovery
  Yu, Jie, Sheng, Xia, ..., Zhang, Sulin, Zheng, Mingyue — Nature Chemical Biology, 2026-08-01
  PBCNet2.0 is a Cartesian tensor‑based Siamese neural network trained on 8.6M protein–ligand pairs for relative affinity prediction, matching physics-based accuracy while capturing subtle interaction effects and mutation sensitivity.

• 📡 Unify learns cellular evolution with universal multimodal embeddings
  Zhong, Huawen, Han, Wenkai, ..., Gao, Xin, Aranda, Manuel — Nature Communications, 2026-07-31
  Unify learns universal, cross‑species cell embeddings by integrating scRNA-seq with protein and general language model embeddings, enabling batch correction and transfer of perturbation responses across distant taxa.

• 🖥️ SpatialFormer: universal spatial representation learning from subcellular molecular to multicellular landscapes
  Wang, Jun, Huang, Yuanhua, Winther, Ole — Nature Computational Science, 2026-07-30
  SpatialFormer combines CNNs and Transformers trained on 700M cell pairs to learn multiscale spatial transcriptomics, improving cell‑type/niche annotation, co‑localization, and communication gene‑pair discovery.

• 📰 TcrDesign: de novo design of epitope-specific full-length T cell receptors
  Diao, Kaixuan, Chen, Jing, ..., Wang, Haopeng, Liu, Xue-Song — Science China Life Sciences, 2026-08-01
  TcrDesign uses transformer-based pretraining to jointly predict TCR–pMHC binding and generate full-length, epitope‑specific TCRs, with experimental validation of binding and functional activation.

• 📰 A Generative Neuro-Symbolic AI for Protein Sequence Design.
  Marianne Defresne, Delphine Dessaux, Samuel Buchet, Lucie Barthe, ..., Alain Roussel, Luis F Garcia-Alles, Thomas Schiex, Sophie Barbe — Advanced science (Weinheim, Baden-Wurttemberg, Germany), 2026-07-30
  EffieDes’s neuro‑symbolic protein design couples learned fitness with automated reasoning to satisfy hard constraints and optimize sequences, validated by selective assembly and a potent nanobody.
  Affiliations: TBI, Université de Toulouse; MIAT, Université de Toulouse; ...

• 📰 Generation of antifungals to combat drug resistance using language models and diffusion models.
  Yeji Wang, Yuemei Dong, Yi Zheng, Lintao Xu, ..., Tao Shen, Wei Zhao, Hongxiang Lou, Wenqiang Chang — Journal of advanced research, 2026-08-01
  MolDiffusion, a diffusion‑based generative platform, designs active single‑ and dual‑target antifungal compounds, with ~50% in vitro hit rate and two candidates showing in vivo efficacy in a mouse candidiasis model.
  Affiliations: Department of Natural Product Chemistry, Key Laboratory of Chemical Biology (Ministry of Education); Department of Clinical Pharmacy, Institute of Clinical Pharmacology; ...

• 📰 AI-driven discovery and validation of a GIPC1 PDZ domain inhibitor for pancreatic ductal adenocarcinoma.
  H. K. Rachamala, Naga Malleswara Rao Nakka, R. Angom, Sai Manasa Varanasi, ..., Sourav Pr Mukherjee, H. Babiker, Krishnendu Pal, Debabrata Mukhopadhyay — Cell reports, 2026-07-31
  An AI-driven pipeline discovers a selective small‑molecule inhibitor of GIPC1’s PDZ domain, validated by HDX‑MS and in vivo synergy with gemcitabine to suppress PDAC growth and improve survival.

• 📰 Integrating chemical structures as treatments improves representations of microscopy images for morphological profiling.
  Yemin Yu, Emre Hayir, Neil Tenenholtz, Lester Mackey, Ying Wei, David Alvarez-Melis, Ava P Amini, Alex X Lu — PLoS computational biology, 2026-07-29
  MICON models compound‑induced cellular phenotypes by contrastive learning between chemical structures and high‑content images, improving reproducibility and cross‑site generalization in morphological profiling.
  Affiliations: Department of Computer Science, City University of Hong Kong; Microsoft Research, Cambridge; ...

• 📡 An embedding-based framework enables statistical testing of gene-set function hypotheses inferred by large language models
  Tan, Yanhao, Wang, Li-Ju, ..., Tseng, George C., Chiu, Yu-Chiao — Nature Communications, 2026-07-30
  LLM embeddings enable hypothesis testing for LLM‑proposed gene‑function links, rescuing cases where enrichment tests fail and validating on experimentally informed gene sets.

• 📰 Persistent sheaf Laplacian analysis of protein stability and solubility changes upon mutation.
  Yiming Ren, Junjie Wee, Xi Chen, Grace Qian, Guo-Wei Wei — Protein science : a publication of the Protein Society, 2026-08-01
  SheafLapNet leverages persistent sheaf Laplacian–based topological deep learning with protein transformers to predict mutation‑induced changes in protein stability/solubility, improving interpretability and accuracy over conventional TDA.
  Affiliations: Department of Mathematics, Michigan State University; The Frazer School, Gainesville; ...

• 📰 AF-CALVADOS: AlphaFold-guided simulations of multi-domain proteins at the proteome level.
  Sören von Bülow, Kristoffer E Johansson, Kresten Lindorff-Larsen — Protein science : a publication of the Protein Society, 2026-08-01
  AF‑CALVADOS integrates AlphaFold‑informed folded domains with coarse‑grained CALVADOS to simulate proteome‑scale ensembles of multi‑domain proteins, validated on >400 proteins and released for large‑scale analysis.
  Affiliations: Structural Biology and NMR Laboratory, Linderstrøm-Lang Centre for Protein Science

• 📰 PETIMOT: a novel framework for inferring protein motions from sparse data using SE(3)-equivariant graph neural networks.
  Valentin Lombard, Julien Nguyen Van, Sergei Grudinin, Elodie Laine — Acta crystallographica. Section D, Structural biology, 2026-08-01
  PETIMOT employs SE(3)-equivariant GNNs with PLM transfer learning and symmetry‑aware losses to infer continuous protein motions from sparse data, outperforming diffusion/flow models on dynamic ensembles.
  Affiliations: Department of Computational, Quantitative; Université Grenoble Alpes, CNRS; ...

• 📰 RINAMI: Residue-attributed interpretable neural network for predicting absolute folding free energy by merging structure and sequence information.
  Naoki Tomita, George Chikenji — Protein science : a publication of the Protein Society, 2026-08-01
  RINAMI fuses ProteinMPNN structural signals with ESM2 sequence embeddings via cross‑attention to predict absolute folding free energy (ΔG), improving accuracy and offering residue‑level interpretability for design triage.
  Affiliations: Department of Applied Physics, Graduate School of Engineering

• 📰 Pocket restraints guided by B-cell epitope prediction improve Chai-1 antibody-antigen structure modeling.
  Joakim Nøddeskov Clifford, Morten Nielsen — Protein science : a publication of the Protein Society, 2026-08-01
  BepiPocket/DiscoPocket inject epitope‑prediction restraints into Chai‑1 Ab–Ag docking, substantially boosting accuracy and binding‑mode diversity across 1,628 complexes.
  Affiliations: Department of Health Technology, Technical University of Denmark

• 🏛️ Discovery of a phenazine-thiol conjugase from sparse data using genome-informed machine learning.
  Xiaoyu Shan, Inês B Trindade, Nathaniel R Glasser, Korbinian O Thalhammer, Matthew Scurria, Ariane Mora, Stuart J Conway, Dianne K Newman — Proceedings of the National Academy of Sciences of the United States of America, 2026-08-04
  ML‑CITO couples genome context with contrastive learning in protein language space to discover phenazine‑interacting enzymes from just 14 known sequences, identifying a phenazine‑thiol conjugase with validated activity.
  Affiliations: Division of Biology and Biological Engineering, California Institute of Technology; Resnick Sustainability Institute, California Institute of Technology; ...

• 📰 BrainShuttle-ESM: A Multi-Stage Transformer Architecture for Predicting Blood-Brain Barrier-Penetrating Short Peptides
  Sathiyajith J.N., G. C, Pratiti Bhadra — Computational biology and chemistry, 2026-08-01
  BrainShuttle‑ESM fine‑tunes ESM‑2 with a multi‑stage strategy to predict BBB penetration of short peptides, achieving AUC ~0.88 and providing residue‑level attention insights into amphipathicity and hydrophobicity.

• 📰 SaintGSE: Transformer-based efficient and explainable gene set enrichment analysis.
  Min-Seung Jeon, Jiho Nam, Minseok Lee, Chanmi Cho, Siyoung Yang, Seong-Il Eyun — Osteoarthritis and cartilage, 2026-08-01
  SaintGSE trains a transformer on large DEG–pathway labels for efficient, explainable gene‑set inference, with Integrated Gradients‑prioritized drivers and experimental validation in osteoarthritis models.
  Affiliations: Department of Life Science, Chung-Ang University; Department of Biological Sciences, Sungkyunkwan University; ...

🧬 Preprints (arXiv + bioRxiv)

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

• 📄 Accurate structural modeling of chemically diverse molecular interfaces with Vilya-2
  Vilya Research, :, Pascal Sturmfels, Naozumi Hiranuma, ..., Jeffrey K. Holden, Adam P. Moyer, Patrick J. Salveson, Ivan Anishchanka — arXiv, 2026-07-28
  Introduces Vilya-2, an all-atom diffusion transformer for peptide–protein and small-molecule–protein complex modeling, achieving SOTA peptide interface recovery and strong docking generalization to power de novo peptide therapeutics.

• 🧬 Modeling the structure-conditioned sequence landscape for large-scale protein design with TriFlow
  Srinivasan, H.; Yuan, R.; Zhang, J.; Cong, Q.; Zhou, J. — bioRxiv, 2026-08-01
  TriFlow combines a three-track RoseTTAFold-like architecture with discrete flow matching for efficient, globally contexted protein sequence design, markedly improving de novo binder success and specificity at scale.
  Affiliations: University of Chicago

• 📄 SE(3)-MeanFlow: Few-Step Protein Backbone Generation on Lie Groups
  Yikun Bai, Binghang Lu, Yikai Liu, Elaheh Akbari, ..., Ping He, Shuchan Wang, Ruqi Zhang, Guang Lin — arXiv, 2026-07-29
  SE(3)-MeanFlow extends few-step MeanFlow to Lie groups for protein frames, enabling high-quality backbone generation with far fewer evaluations via an SE(3) alpha-Flow objective and rectification.

• 🧬 Elucidating enzyme-substrate specificity through co-folding foundation model
  Cheng, X.; Seo, S.; Huh, C.; Chen, J.; ...; Guo, P.; Weng, J.-K.; Kim, W. Y.; Jin, W. — bioRxiv, 2026-08-02
  Boltz2ESI is a co-folding foundation model that predicts enzyme–substrate interactions by natively capturing active-site plasticity, outperforming sequence and rigid-docking baselines and aiding pathway de-orphaning.
  Affiliations: Northeastern University

• 🧬 Illuminating the Ligandable Proteome with AI Protein Profiling
  Dayhoff, G. W.; Kortzak, D.; Liu, R.; Shen, M.; Lin, J.; Zhang, Z.-Y.; Shen, J. — bioRxiv, 2026-07-30
  AiPP, a sequence-based multitask platform built on ESMC with LatentLift label harmonization, maps proteome-wide ligandable residues and context, guiding covalent probe/inhibitor discovery (e.g., PTPN6).
  Affiliations: University of Maryland School of Pharmacy

• 🧬 The Human Bindome: A Proteome-scale Atlas of Designed Binder Candidates
  Wenckstern, J.; Diaz-Rovira, A. M.; Kuhn, J.; Ban, A.; ...; Picotti, P.; Winter, G.; Taipale, M.; Correia, B. E. — bioRxiv, 2026-07-30
  The Human Bindome scales DL-based binder design to >300k candidates covering 8,296 human proteins, providing sequences, structures, and confidence scores as a proteome-wide resource for perturbation and target discovery.
  Affiliations: EPFL

• 🧬 ConfDock: Atom-specific Uncertainty Quantification for Molecular Docking via Conformal Prediction
  Hao, H.; Elhendawy, N.; Wang, Y.; Lu, C. — bioRxiv, 2026-08-01
  ConfDock pairs GNN-based quantile estimation with split conformal prediction to produce atom-level uncertainty intervals for docking poses, achieving rigorous coverage with substantially tighter bounds across diverse complexes.
  Affiliations: University of Illinois Chicago

• 📄 Implicit Machine Learning Force Fields Accelerate Molecular Dynamics Simulations
  Johannes Maeß, Leon Werner, J. Thorben Frank, Winfried Ripken, Martin Michajlow, Joshua Futterer, Klaus-Robert Müller, Stefan Chmiela — arXiv, 2026-07-31
  Implicit ML force fields cast deep models as fixed-point solvers to reuse representations across MD timesteps, delivering 2–5x compute/memory savings while preserving accuracy for invariant/equivariant GNN force fields.

• 📄 MEGA-CL: A Molecular Foundation Model for Generalizable ADMET Prediction through Graph External Attention and Contrastive Learning
  Tinghui Jin, Kedu Jin, Ying Li, Guanghui Ren, ..., Li-bin Wei, Xijing Chen, Di Zhao, Jinfeng Liu — arXiv, 2026-07-27
  MEGA-CL is a foundation GNN with external attention and contrastive learning for generalizable ADMET prediction, showing strong accuracy across 21 tasks and prospective validation for microsomal clearance and CYP inhibition.

• 📄 Persistent Manifold Learning of Protein Properties
  Xingjian Xu, Zhe Su, Guo-Wei Wei, Chunmei Wang — arXiv, 2026-07-27
  Persistent Manifold Learning represents binding interfaces as multiscale manifolds and extracts topological/spectral invariants fused with protein/molecular LMs, outperforming SOTA on metalloprotein–ligand and PPI benchmarks.

• 📄 Catalyst Diffusion Transformer: Generative Inverse Design of Heterogeneous Catalysts
  Hayoung Doo, Dong Hyeon Mok, Seoin Back, Jonggeol Na — arXiv, 2026-07-27
  CatDiT is a latent diffusion transformer for inverse design of heterogeneous catalysts conditioned on adsorbate type, binding energy, and class, generating valid alloy/oxide surfaces and enriching NRR-active candidates.

• 📄 Data Fusion and Contrastive Alignment for Unconstrained IR Molecular Structure Elucidation
  Ethan J. Mick, Campbell A. Sweet, Matthias J. Young, Derek T. Anderson — arXiv, 2026-07-28
  A transformer with a Mixture-of-Experts decoder and non-additive Choquet aggregation plus contrastive alignment enables unconstrained IR-to-structure elucidation, boosting Top-K accuracy well beyond IR-only baselines.

• 🧬 A Preparation-Free Mixture-of-Experts Framework for Protein-Ligand Affinity Prediction
  Bao, H.; Dong, S. — bioRxiv, 2026-07-28
  HydrAffinity uses pre-trained encoders and a sparse Mixture-of-Experts to predict protein–ligand affinity without explicit interaction features, matching interaction-based SOTA on CASF-2016 and enabling fast screening.
  Affiliations: Lanzhou University

• 📄 Oracle-Budgeted Molecular Optimization with Short-Term Graph Memory
  Jiannan Yang, Veronika Thost, Xiang Ling, Tengfei Ma — arXiv, 2026-07-30
  Short-term graph memory adds a lightweight online surrogate to pre-screen generated molecules under a fixed oracle budget, consistently improving top-10 optimization scores across generators without extra oracle calls.

• 📄 Q-Steer: Action-Value Guidance for Molecular Policy Optimization
  Xinyu Wang, Jinbo Bi, Minghu Song — arXiv, 2026-07-29
  Q-Steer augments molecular language-model rollouts with an offline prefix action-value scorer to steer next-token sampling, improving optimization rewards across backbones/optimizers without increasing online oracle usage.

• 🧬 How Bias Shapes the Leaderboard: Scoring Function Performance Under Scrutiny
  Graber, D.; Kopko, J.; Stockinger, P.; Nakandalage, R.; Kuhn, B.; Mishra, S. — bioRxiv, 2026-07-30
  Reveals pocket-frame leakage and other biases in ML docking benchmarks; under artifact-free and OOD protocols, ML models’ gains collapse and Vina often generalizes better, underscoring the need for rigorous evaluation.
  Affiliations: Seminar for Applied Mathematics, Department of Mathematics and ETH AI Center

• 📄 Chem World: A Large-Scale Benchmark and Physics-Informed Framework for Trustworthy Chemical Property Prediction
  Tianyou Bai, Huan Wang, Mingchen Gao, Fangyue Lin, Pinze Ren, Zhenlin Zhao, Siming Dong — arXiv, 2026-07-30
  Chem World unifies 17 datasets into a rigorous chemical property benchmark and proposes a physics-informed Mixture-PINN, improving accuracy and robustness for trustworthy computational chemistry.

• 🧬 PG-LLM: Benchmarking General-Purpose Language Models for Protein Variant Ranking
  Arora, R. K.; Chen, L. T.; Du, M.; Marks, D.; Church, G. — bioRxiv, 2026-07-28
  PG-LLM benchmarks general LLMs for protein variant ranking on 217 assays, showing top models rival ESM2-650M and surpass many sequence-only predictors while quantifying remaining gaps to specialist methods.
  Affiliations: Harvard Medical School

• 🧬 Sequence determinants of pathogenicity in glucose-6-phosphatase linked to glycogen storage disease type 1a
  Stein, R. A.; Hawes, E. M.; Norphlet, C. M.; Rakonick, M. H.; ...; Lucerne, A. M.; Da Silva, V. R.; O'Brien, R. M.; Claxton, D. P. — bioRxiv, 2026-07-28
  Integrates AlphaMissense with biochemical/biophysical assays to map mechanisms of G6PC1 missense variants, linking stability to catalytic capacity and refining clinical classification in glycogen storage disease.
  Affiliations: Vanderbilt University

• 🧬 Transplanting enzyme active site geometry into antibody CDRs for catalytic antibody design
  Zhu, Y. — bioRxiv, 2026-07-28
  Computationally transplants enzyme active-site geometry into antibody CDRs using constrained diffusion-based loop reconstruction and sequence design, outlining a route to catalytic antibody engineering.
  Affiliations: Shaanxi University of Technology