Weekly BioML Digest [September 07, 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)

1756 matched filters -> 20 selected after LLM relevance + novelty ranking.

• 🛠️ Aligning protein-generative models to experimental fitness with ProteinDPO
  Widatalla, Talal, Borah, Ashir A., ..., Rafailov, Rafael, Hie, Brian L. — Nature Methods, 2026-09-01
  Aligns a structure-conditioned protein language model with experimental stability via direct preference optimization (ProteinDPO), enabling generation and scoring of thermostable proteins and stabilizing multichain complexes and influenza hemagglutinin.

• 🔬 Predicting cellular responses to perturbation across diverse contexts with State.
  Abhinav K Adduri, Dhruv Gautam, Beatrice Bevilacqua, Mohsen Naghipourfar, ..., Alexander Dobin, Dave P Burke, Hani Goodarzi, Yusuf H Roohani — Cell, 2026-08-31
  Introduces State, a single-cell ML model trained across 167M cells to predict genetic, signaling, and chemical perturbation effects across diverse contexts, improving DE gene identification and generalization to unseen cell states.
  Affiliations: Arc Institute, Palo Alto; Arc Institute, Palo Alto; ...

• 🧫 AI-enhanced adaptive virtual screening of large libraries for ligand discovery
  Cecchini, Domiziana, Nigam, AkshatKumar, ..., Arthanari, Haribabu, Gorgulla, Christoph — Nature Biotechnology, 2026-09-01
  Presents AdaptiveFlow, an open-source ULVS platform combining property-grid prioritization and active learning to screen 69B molecules at cloud scale, integrating >1,500 docking/ML protocols and yielding nanomolar inhibitors with co-crystal validation.

• 🔭 Epitope-conditioned generation of T-cell receptor β-chain CDR3 candidates using a pre-trained transformer model
  Yang, Jiannan, He, Bing, ..., Li, Ting, Yao, Jianhua — Genome Biology, 2026-09-02
  ERTransformer uses pretrained epitope and TCR transformers to conditionally generate TCR β-chain CDR3s for given peptide–HLA targets, with wet-lab validation showing comparable or superior T cell activation to natural TCRs.

• 📡 Unsupervised discovery of functional sequence patterns from protein language model with MotifAE
  Hou, Chao, Liu, Di, Shen, Yufeng — Nature Communications, 2026-09-02
  MotifAE applies a sparse autoencoder with smoothness regularization to protein language model embeddings, revealing interpretable sequence motifs and domains and enabling stability-specific fitness landscape prediction.

• 🤖 NucleicBERT interprets RNA sequence space through self-supervised language modelling
  Upadhyay, Utkarsh, Herold, Julian, Götz, Markus, Schug, Alexander — Nature Machine Intelligence, 2026-09-03
  NucleicBERT is a self-supervised RNA language model that learns structural constraints from single sequences, matching or exceeding RNA structure/function predictors without MSAs and supporting explainable latent organization.

• 📰 Hierarchical discrete representations for coarse-to-fine protein conformation generation.
  SeokJun On, Yujin Jeong, Kang-Hyeon Kim, Kyungheon Kang, Eun-Sol Kim — Bioinformatics (Oxford, England), 2026-08-31
  Proposes a VQ-based hierarchical discrete representation to generate protein conformational ensembles coarse-to-fine, outperforming prior models while balancing diversity with structural validity for flexible biomolecular modeling.
  Affiliations: Department of Artificial Intelligence, Hanyang University; Department of Computer Science, Hanyang University

• 📰 Condition controllable generation of 3D molecules using textual prompts by multimodal equivariant diffusion model
  Fang, Yi, Liu, Yuan, ..., Pan, Xiaoyong, Shen, Hong-Bin — npj Digital Medicine, 2026-09-01
  TDmol aligns text with molecular 2D/3D modalities in an equivariant diffusion framework to generate valid conditioned 3D molecules from textual prompts, advancing controllable de novo design.

• 📰 Geo-Hete-HyperGNN as a prior-free molecular hypergraph representation learning strategy
  Zhou, Wei, Peng, Yingzi, ..., Liu, Zirui, Liu, Xiong — Communications Chemistry, 2026-09-01
  Geo-Hete-HyperGNN is an equivariant molecular hypergraph network fusing geometry with spectroscopy-informed hyperedges to improve transferable representations across MoleculeNet and spectroscopy prediction tasks.

• 📰 PKP-Diffmol: A Physicochemical Knowledge-Prompt Encoding and Latent Diffusion Framework for Molecular Property Prediction.
  Ruizi Liu, Tongtong Yuan, Molin Guo — Journal of chemical information and modeling, 2026-09-01
  PKP-DiffMol augments fragment-level SMILES encoders with numerical and semantic physicochemical prompts plus latent diffusion–based augmentation to boost scaffold-split molecular property prediction across bioactivity and safety tasks.
  Affiliations: College of Intelligent Science and Engineering, Northeast Agricultural University; College of Computer Science, Beijing University of Technology; ...

• 📰 ABAG-Rank: Improving Model Selection of AlphaFold Antibody-Antigen Complexes by Learning to Rank.
  Matteo Tadiello, Marko Ludaic, Vsevolod Viliuga, Arne Elofsson — Bioinformatics (Oxford, England), 2026-09-04
  ABAG-Rank learns to rank AlphaFold antibody–antigen decoys using deepsets over simple geometric descriptors and AF confidence scores, substantially improving selection of accurate Ab–Ag complexes.
  Affiliations: Department of biochemistry and biophysics (DBB) and SciLifeLab, Stockholm University; Max Planck Institute for Polymer Research (MPIP), Ackermannweg 10

• 🧬 Squidly harnesses enzyme functional hierarchy and contrastive learning to efficiently predict catalytic residues from sequence.
  William J F Rieger, Mikael Bodén, Frances Arnold, Ariane Mora — eLife, 2026-09-04
  Squidly combines PLM token embeddings with contrastive learning and functional-hierarchy pairing to predict enzyme catalytic residues from sequence, surpassing structure-dependent baselines and enabling large-scale annotation.
  Affiliations: School of Chemistry and Molecular Biology, University of Queensland; Division of Chemistry and Chemical Engineering, California Institute of Technology

• 📰 DyProL: Dynamic Ensemble Representation Learning for Protein-Nucleic Acid Binding Site Prediction.
  Pengpai Li, Yiman Liu, Liya Liang, Rongming Liu — Advanced science (Weinheim, Baden-Wurttemberg, Germany), 2026-09-03
  DyProL models protein conformational ensembles rather than single structures to enhance nucleic acid binding-site prediction, particularly under realistic apo/predicted inputs.
  Affiliations: MOE Key Laboratory of Bio-Intelligent Manufacturing, School of Bioengineering; Ningbo Institute of Dalian University of Technology, Ningbo

• 📰 Retrieval-Augmented Residual Graph Neural Network for Protein-Protein Interaction Site Prediction.
  Jia Mi, Yawen Liu, Chong Chu, Chang Li, Jing Wan, Kunfeng Wang — IEEE transactions on neural networks and learning systems, 2026-09-03
  RGLLA-PPIS integrates AF3 structures, equivariant/residual GNNs, and retrieval-augmented priors from PLMs/LLMs to predict protein–protein interaction sites with improved accuracy and robustness validated against wet-lab data.

• 📰 A deep learning multi-attention Bi-GRU framework for kcat prediction with segmentation-based insights.
  Priyanka, Ramesh Chandra, Md Shah Fahad, Raushan Oraon, Ashish Ranjan — Enzyme and microbial technology, 2026-09-01
  KcatNeuroCortex uses a segmentation-aware Bi-GRU with multi-attention to learn local motifs and long-range dependencies from enzyme sequences, yielding interpretable and improved kcat predictions for engineering and kinetic modeling.
  Affiliations: BIT Mesra, India. Electronic address: reachpriyanka20@gmail.com.; BIT Mesra, India. Electronic address: rameshchandra@bitmesra.ac.in.; ...

• 📡 Metabolic engineering and deep learning-driven protein engineering for N-Acetylneuraminic acid biosynthesis in Escherichia coli
  Wang, Nan-Kai, Yue, Song, ..., Xu, Zheng-Hong, Shi, Jin-Song — Nature Communications, 2026-09-05
  Combines deep learning kinetic prediction (DLCatalysis) with pathway/metabolic engineering to identify and redesign rate-limiting enzymes, achieving 85.5 g/L NeuAc production in E. coli.

• 📰 A deep learning framework for oligopeptide candidate discovery across metabolism-related contexts.
  Baichuan Xiao, Hao Zhu, Chao Ma, Xiaoran Wang, ..., Liguo Wang, Martin Cheung, Zhanzhan Li, Yong-Biao Zhang — Cell reports methods, 2026-09-04
  Deepeptide discovers oligopeptide candidates linked to metabolism-related processes across angiogenesis, lipid/glucose metabolism, and osteogenesis, with 62% prospective bioactivity validation.
  Affiliations: School of Engineering Medicine, Beihang University; Department of Chemistry, Tsinghua University; ...

• 📰 Mapping the combinatorial coding between olfactory receptors and perception with deep learning.
  Seyone Chithrananda, Judith Amores, Kevin K Yang — Cell systems, 2026-09-02
  MolOR links molecules to predicted olfactory receptor activation profiles via cross-attention between GNNs and receptor PLM embeddings, improving odor percept prediction and enabling orphan OR ligand discovery.
  Affiliations: Microsoft Research, Cambridge; Stanford University, Stanford; ...

• 📰 Decoding enzymatic landscapes: a knowledge graph-enhanced large language model framework for microbial enzyme production and catalysis systems.
  Qichang Tong, Lincong Zhou, Xu Liu, Xiaoqing Liu, ..., Jian Tian, Dongfei Han, Xianghua Yan, Feifei Guan — aBIOTECH, 2026-09-01
  MEPAM builds an ontology-driven enzyme knowledge graph with LLM-based extraction and develops an RAG QA system that accurately answers microbial enzyme production/catalysis queries, reducing hallucinations vs general LLMs.
  Affiliations: College of Animal Sciences and Technology, Huazhong Agricultural University; State Key Laboratory of Animal Nutrition and Feeding, Institute of Animal Science; ...

• 📰 MetaTIS: a tool to predict cognate and near-cognate translation initiation sites in human.
  Aram Papazian, Volkhard Helms — NAR genomics and bioinformatics, 2026-09-01
  MetaTIS ensembles fine-tuned genomic and protein language models to predict human translation initiation sites (AUG and near-cognate) using sequence context, outperforming on multiple riboprofiling-derived test sets.
  Affiliations: Center for Bioinformatics, Saarland Informatics Campus

🧬 Preprints (arXiv + bioRxiv)

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

• 🧬 Toward De Novo Protein Design from Natural Language
  Dai, F.; You, S.; Zhu, Y.; Gao, Y.; ...; Si, T.; Liu, J.; Lu, H.; Yuan, F. — bioRxiv, 2026-08-31
  Presents Pinal, a 16B generative model trained on 1.7B protein–text pairs that translates natural-language functional descriptions into protein sequences; designs multiple active proteins and enzymes de novo with experimental validation, including an enzyme surpassing its natural counterpart.
  Affiliations: Westlake University

• 🧬 Designing antimicrobials with programmable mechanism and safety
  Szymczak, P.; Torres, M. D. T.; Soares, D.; Hetzel, L.; ...; Günnemann, S.; Theis, F. J.; de la Fuente-Nunez, C.; Szczurek, E. — bioRxiv, 2026-09-01
  Introduces OmegAMP, a conditional diffusion model controlling charge, hydrophobicity, and length to program antimicrobial peptide activity and safety; wet-lab and in vivo studies show broad efficacy and mechanism-preserving motif-guided designs.
  Affiliations: Institute of AI for Health, Helmholtz Zentrum Munich

• 🧬 OmniSyn unifies target-aware molecular generation and optimization within a synthesis-native LLM framework across the human proteome
  QIN, Z.; Li, Y.; Zhang, Y.; Zhao, Y.; ...; Min, Y.; Yang, J.; He, X.; Cao, D. — bioRxiv, 2026-09-06
  OmniSyn is a protein-sequence-conditioned MoE LLM with a synthesis-action decoder that unifies target-aware ligand generation, synthesizability projection, and hit-to-lead optimization; achieves SOTA on MolGenBench and generates a 2.7B-compound proteome-scale library with high retrosynthetic success.
  Affiliations: Tongji University

• 🧬 Forecasting viral evolution from phylogenetic trees
  Specht, I.; Park, S.; Chithrananda, S.; Driscoll, C. L.; Brixi, G.; Palacios, J. A.; Hie, B. L. — bioRxiv, 2026-09-05
  antiGen learns from phylogenetic-tree structure to forecast future viral mutations (e.g., SARS-CoV-2 spike), outperforming baselines across multiple viruses; predicted mutations retain infectivity in vitro.
  Affiliations: Stanford University

• 📄 Latent unified smooth Hamiltonians for excited state chemistry
  David Juergens, Martin Stöhr, Andreas E. Hillers-Bendtsen, O. Jonathan Fajen, Todd J. Martínez — arXiv, 2026-09-01
  Proposes a transformer-based architecture that learns a latent electronic-state Hamiltonian to jointly model ground and excited states with nonadiabatic couplings; accurately reproduces photochemical landscapes (e.g., conical intersections, Berry phase) for thymine and azobenzene.

• 🧬 RECON infers regions of interest from H&E images and reconstructs whole-slide molecular profiles at single-cell resolution
  Yang, X.; Hao, N.; Zhao, R.; Angel, S.; ...; Olson, D.; Yu, K.-H.; Ruiz de Luzuriaga, A.; Wan, G. — bioRxiv, 2026-09-01
  RECON selects ROIs from H&E at single-cell resolution and trains deep models to reconstruct whole-slide transcriptomic/proteomic profiles for all cells; outperforms superpixel-based methods and achieves high per-cell accuracy across markers.
  Affiliations: The University of Chicago

• 📄 Language-Informed Flow Matching for Trend-Guided Structure-Based 3D Molecular Generation
  Tianyu Gao, Zhikai Su, Jiashu Li, Wenjun Gao, Zichuan Ying, Zhe Zhao, Fei Zhang, Ye Wei — arXiv, 2026-08-31
  LiFT couples language-derived chemical priors with flow-matching 3D generation via cross-modal conditioning and a self-conditioned decoupled router; improves medicinal chemistry metrics while maintaining structural validity for SBDD and scaffold hopping.

• 📄 SurfSpec: Enhancing Off-Target-Agnostic Specificity by Bounding Pocket-Ligand Geometric Mismatch
  Minyeong Hwang, Yoorim Gang, Ziseok Lee, Wooyeol Lee, Young Bin Park, Jae-Mun Choi, Kyungsu Kim, Eunho Yang — arXiv, 2026-09-02
  SurfSpec formalizes pocket–ligand geometric mismatch to derive an off-target-agnostic specificity lower bound and iteratively grows ligands toward under-occupied pocket regions; reduces mismatch and improves empirical specificity on CrossDocked2020 while preserving affinity gains.

• 📄 SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign
  Jiarui Lu, Yuyang Wang, Yizhe Zhang, Jiatao Gu, Navdeep Jaitly, Joshua M. Susskind, Miguel Ángel Bautista — arXiv, 2026-09-03
  SimpleDesign is a single-stage Mixture-of-Transformer that co-generates protein sequences and 3D structures with modality-specific processing and joint objectives; achieves strong results on co-design and unconditional generation benchmarks.

• 📄 S3C-LLM: Skill-Code Guided Agentic Language Models for Spectrum-to-Structure Elucidation
  Xuanle Zhao, Xinyuan Cai, Xiang Cheng, Bo Xu — arXiv, 2026-08-31
  S3C-LLM is a skill- and code-grounded agentic LLM that performs diagnostic peak analysis, fragment reasoning, and constraint checking before structure generation; trained with step-level RL, it surpasses general LLMs and spectrum-specific baselines on spectrum-to-structure tasks.

• 📄 NEAT-POCKET: Pocket-Conditioned Autoregressive 3D Molecular Generation with a Neighborhood-Guided Set Transformer
  Roxane Axel Jacob, Daniel Rose, Thierry Langer, Johannes Kirchmair — arXiv, 2026-09-04
  NEAT-POCKET extends autoregressive 3D molecular generation to be pocket-conditioned with a neighborhood-guided set transformer modeling explicit hydrogens; delivers fast, competitive structure-based ligand generation and pocket-conditioned fragment completion.

• 📄 Recovering molecules from coarse-grained beads: free-energy-conditioned generative backmapping across chemical space
  Luis Itza Vazquez-Salazar, Tristan Bereau — arXiv, 2026-09-03
  Juniper is a discrete denoising diffusion model that backmaps coarse-grained bead representations to atomic molecules conditioned on partition free energy; recovers valid, diverse candidates aligned with CG free-energy targets to close the screen-to-compound gap.

• 📄 Elite-Weighted Supervised Fine-tuning for Goal-Directed Molecular Optimization
  Shiyun Wa, Yifei Wang, Anna G. Green, Simone Sciabola, Ye Wang — arXiv, 2026-08-31
  Elite-Weighted Supervised Fine-tuning (EW-SFT) guides molecular generators by elite selection under the model’s native loss, sidestepping policy-gradient RL; consistently improves shape/similarity-constrained optimization across generators, tasks, and oracles.

• 📄 Training Large Language Models for Small-Molecule Design with Synthetic Task Scaling
  Frank Hu, Shriram Chennakesavalu, Zichen Wang, Patricia Suriana, Bodhi Vani, Kirill Shmilovich, Kangway Chuang, Colin Grambow — arXiv, 2026-09-04
  Shows that curriculum-based post-training on synthetic tasks enables LLMs to learn molecular design strategies that transfer to expensive structure-based lead optimization; surpasses larger frontier models without direct use of costly reward functions.

• 🧬 Allosteric pathways govern Gα protein coupling selectivity at promiscuous GPCRs
  Boora, T. S.; Chen, H.; Cho, A.; van der Velden, W. J. C.; ...; RODIN, A. S.; Branciamore, S.; Vaidehi, N.; Laporte, S. — bioRxiv, 2026-09-03
  Combining whole-receptor mutagenesis, MD, interpretable ML, and Bayesian networks, maps allosteric residue networks controlling G-protein selectivity at promiscuous GPCRs; reveals long-range communication as a principal determinant of coupling preference.
  Affiliations: Research Institute of the McGill University Health Centre

• 🧬 Towards Sparse Causal Features for Zero-shot Mutation Effect Prediction in a Protein Language Model
  Mohanty, S.; Phutela, M.; Green, A. G. — bioRxiv, 2026-09-03
  Introduces a sparse feature-circuit framework (sparse autoencoders, attributions, activation patching) to causally identify latent features mediating zero-shot mutation-effect predictions in ESM-2; circuits align with 3D contacts and efficiently recover model behavior.
  Affiliations: University of Massachusetts

• 📄 Learning Task-Specific Antibody Representations via Function-Aware Masking
  Ayan Goel, Thomas A. Walton, Amirali Aghazadeh — arXiv, 2026-09-01
  Function-aware masking injects biological priors (e.g., IMGT annotations, structural cues) into antibody LM pretraining via targeted corruption; yields substantial gains on CDR-, binding-, and structure-related downstream tasks.

• 📄 Towards AI-Driven Nanomedicine Discovery: A Benchmark and Multimodal Learning Framework for Nano Self-Assembly Prediction
  Quan Hao, Mengyue Fan, Zifan Dong, Jianduo Zhao, ..., Fei Xia, Jigang Wang, Liguo Zhang, Chong Qiu — arXiv, 2026-09-03
  NSA-Bench standardizes nano self-assembly prediction, and NSA-Net integrates graph, sequence, and physicochemical descriptors for pairwise self-assembly; achieves high ROC-AUC with interpretable features, supporting nanomedicine formulation design.

• 📄 WEECFP-SuRGE: Wide Embedded Extended Connectivity Fingerprint with Substructure Rotary Graph-distance Encoding
  Robert Epps — arXiv, 2026-09-04
  WEECFP provides a near-lossless continuous fingerprint and a SuRGE transformer with graph-distance rotary encodings; leads the TDC ADMET leaderboard without external pretraining and enables near-exact SMILES reconstruction from tokens.

• 📄 Schrödinger Bridges on Lie Group Manifolds for Probabilistic Intrinsic Generation
  Shizhe Zhang, Mingyang Zhao, Lei Ma — arXiv, 2026-09-02
  Develops Schrödinger bridges for kinetic dynamics on Lie-group manifolds with calibrated endpoint constraints and efficient controllers; validated on protein/RNA torsions and conformational pathway generation in compact representations.