Weekly BioML Digest [October 05, 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)

1852 matched filters -> 20 selected after LLM relevance + novelty ranking.

• 🌿 Function-preserving watermarking of AI-generated proteins
  Stutz, David, Cowen-Rivers, Alexander I., ..., Gowal, Sven, Kohli, Pushmeet — Nature, 2026-09-30
  Introduces SynthIDBio, a dual watermarking framework that embeds function-preserving signatures into AI-generated protein sequences and AlphaFold3 structures, enabling provenance tracking without degrading binding performance.

• 🛠️ CytoVI: deep generative modeling of antibody-based single cell data
  Ingelfinger, Florian, Levy, Nathan, ..., Amit, Ido, Yosef, Nir — Nature Methods, 2026-09-30
  Presents CytoVI, a probabilistic deep generative model that integrates flow/cytof/CITE-seq to denoise, impute, and annotate antibody-based single-cell data, revealing disease-associated immune states.

• 📰 Fast and accurate construction of multiple sequence alignments from protein language embeddings with ARIES.
  Minh Hoang, Isabel Armour-Garb, Mona Singh — Genome research, 2026-10-01
  Develops ARIES, an MSA builder that aligns via PLM-derived reciprocal embedding similarity and dynamic time warping, outperforming traditional methods especially in low-identity regimes.
  Affiliations: Princeton University.; Princeton University mona@cs.princeton.edu.

• 📰 FLIGHTED: Inferring fitness landscapes from noisy high-throughput experimental data.
  Vikram Sundar, Boqiang Tu, Lindsey Guan, Kevin Esvelt — Cell systems, 2026-10-02
  Introduces FLIGHTED, a Bayesian approach that infers protein fitness with experimental-noise-aware uncertainty, boosting downstream ML performance and clarifying data vs. model scale limits.
  Affiliations: Computational and Systems Biology Program, Massachusetts Institute of Technology; Department of Biological Engineering, Massachusetts Institute of Technology; ...

• 📰 GenAI-Net: A generative AI framework for automated biomolecular network design.
  Maurice Filo, Nicolò Rossi, Zhou Fang, Mustafa Khammash — Science advances, 2026-10-02
  Proposes GenAI-Net, a generative AI framework coupling reaction proposal with simulation-based objectives to automatically design biomolecular reaction networks across diverse dynamical tasks.
  Affiliations: Department of Biosystems Science and Engineering, ETH Zürich; State Key Laboratory of Mathematical Sciences, Academy of Mathematics and Systems Science

• 📰 The trRosettaRNA server for RNA structure prediction
  Wang, Wenkai, Liu, Xiaocheng, Peng, Zhenling, Yang, Jianyi — Nature Protocols, 2026-10-01
  Describes trRosettaRNA, a deep learning server for RNA 3D structure prediction with end-to-end modeling and energy refinement, validated on Rfam and available as open-source.

• 📰 DreamFold: A World Model to efficiently generate protein folding pathways in the latent space
  Alan Ianeselli, J. Im, Eddie Cavallin, Mark B. Gerstein — Protein Science : A Publication of the Protein Society, 2026-09-30
  Presents DreamFold, a world-model generative framework that learns latent protein folding dynamics and generates folding pathways orders of magnitude faster than MD.

• 📰 scLDM: a conditional diffusion framework for single-cell perturbation prediction.
  Boyang Wu, Yuhang Liu, Yue Cheng, Xiangrong Liu, Quan Zou, Leyi Wei, Lei Xu — Bioinformatics (Oxford, England), 2026-09-28
  Introduces scLDM, a conditional latent diffusion model that predicts single-cell perturbation responses by conditioning on pre-perturbation state, cell type, and stimulus, surpassing SOTA.
  Affiliations: School of Information, Xiamen University; Faculty of Applied Sciences, Macao Polytechnic University; ...

• 📰 Deep generative framework for modeling single-cell drug perturbation response.
  Yongqing Zhang, Chenpeng Wu, Tianhao Li, Zhigan Zhou, Zhengxiao Huang, Aochen Zhang, Zixuan Wang — Neural networks : the official journal of the International Neural Network Society, 2026-10-01
  Proposes scDPR, which combines latent attribute adaptation with optimal-transport causal graph learning to predict single-cell drug-perturbed transcriptomes and disentangle direct drug effects.
  Affiliations: School of Computer Science, Chengdu University of Information Technology; School of Computer Science, Chengdu University of Information Technology; ...

• 📰 BindRNAgen: Protein-binding RNA Sequence Generation Using Latent Diffusion Models.
  Yan Zhou, Xiaojian Liu, Shengfan Wang, Lin Zhu, Biao Zhang, Yan Huang, Hong-Bin Shen, Xiaoyong Pan — Journal of molecular biology, 2026-10-01
  Presents BindRNAgen, a VAE + conditional latent diffusion framework that generates RNA sequences binding given RBP sequences using PLM conditioning, validated via docking/MD.
  Affiliations: Institute of Image Processing and Pattern Recognition, Shanghai Jiao Tong University; School of Health Science and Engineering, University of Shanghai for Science and Technology; ...

• 📰 Katalyst: Knowledge-Guided Semantic Alignment via Contrastive Learning for Enzyme Turnover Prediction.
  Ruilin Li, Senyu Tang, Jiaqi Deng, Taixing Qiu, Fei Guo, Jijun Tang, Xiaoyi Liu — Bioinformatics (Oxford, England), 2026-10-03
  Introduces Katalyst, a multimodal contrastive framework aligning enzyme sequences, text, substrates, and products to predict kcat with improved accuracy and generalization.
  Affiliations: College of Computer Science and Software Engineering, Shenzhen University; School of Computer Science and Control Engineering, Shenzhen University of Advanced Technology; ...

• 📰 ConTP reshapes transporter functional space to resolve substrate specificity beyond evolutionary proximity
  Feng, Chenjie, He, Wenjia, ..., Tester, Mark, Gao, Xin — Communications Biology, 2026-09-28
  ConTP realigns protein-language-model embeddings around substrate semantics via evolution-informed contrastive learning, enabling substrate-level transporter annotation beyond phylogeny.

• 📰 Genome-wide-scale prediction of compound-protein interactions using foundation and language models based on three-dimensional structures of compounds and proteins.
  Yuga Moriyama, Yuki Matsukiyo, Tsuyoshi Kimura, Hiyori Yamamoto, Yuko Sakajiri, Tomokazu Shibata, Ryusuke Sawada, Yoshihiro Yamanishi — Bioinformatics (Oxford, England), 2026-09-28
  DESTIN integrates 3D-aware compound and protein features (foundation model + PLMs) to predict compound-protein interactions at genome scale, excelling on complex membrane targets.
  Affiliations: Department of Complex Systems Science, Graduate School of Informatics; Integrated Glyco-Big Data Center (iGDATA), Institute for Glyco-core Research (iGCORE); ...

• 📰 MacroTox: A Macroscopic Graph Topology-Based Multimodal Learning Framework for Robust Molecular Toxicity Prediction.
  He Huang, Qinyi Wang, Manzhan Zhang, Pan Dou, Xiaobo Yang, Honglin Li, Shiliang Li — JACS Au, 2026-09-28
  MacroTox introduces a topology-driven multimodal framework that dynamically builds intrabatch drug–drug similarity graphs to improve toxicity prediction and interpretability.
  Affiliations: Innovation Center for AI and Drug Discovery, School of Pharmacy; School of Pharmacy, East China University of Science and Technology; ...

• 📡 Evaluating generative models for structure-based drug discovery from de novo design to lead optimization
  Cao, Duanhua, Fan, Zhehuan, ..., Teng, Dan, Zheng, Mingyue — Nature Communications, 2026-10-01
  MolGenBench benchmarks 17 structure-based generative models, exposing gaps in target awareness, actives recovery, and generalization, and providing standardized datasets/metrics.

• 📰 stFormer integrates spatial ligand signaling into a foundation model for spatial transcriptomics.
  Shenghao Cao, Kaiyuan Yang, Jiabei Cheng, Jiachen Li, Hong-Bin Shen, Xiaoyong Pan, Ye Yuan — Cell reports methods, 2026-09-30
  stFormer (as above) injects spatial ligand signaling via cross-attention and biased SC learning to unify Visium datasets and improve ST foundation modeling.
  Affiliations: Department of Mathematics, College of Sciences; Newtouch Center for Mathematics of Shanghai University, Shanghai University; ...

• 📰 PRISM-Gen: A physics-informed multi-fidelity framework for broad-spectrum coronavirus mpro inhibitor discovery.
  Hongmei Wang, Xinjia Zeng, Lirong Zhu, Aiping Yang, Meiling Zhang — Computational biology and chemistry, 2026-10-01
  PRISM-Gen couples fragment-tree generation with a physics-informed, multi-fidelity electronic screening cascade and cross-coronavirus docking to prioritize broad-spectrum Mpro inhibitors.
  Affiliations: School of Basic Medical Sciences, Tianjin Medical University; Department of Computer and Information Technology, Purdue Polytechnic Institute; ...

• 💻 Multi-Objective-Conditioned-Return-Blocks Decision Transformer: A Novel Approach for Multi-Objective De Novo Drug Design.
  Eyal Mazuz, Or Gindes, L. Rokach, Bracha Shapira — Bioinformatics, 2026-10-02
  M-GoRT uses decision-transformer goal conditioning with return-to-go blocks for multi-objective de novo drug design, achieving competitive success while preserving diversity.

• 📰 Enhancing Conditional Molecular Generation With Pretrained SMILES Transformer and Contrastive Representation Learning.
  Dongcheng Hu, Xin Peng, Song He, Zhangpeng Wei, Weimin Zhong, Feng Qian — IEEE transactions on cybernetics, 2026-10-01
  Contras-STE pretrains a SMILES Transformer with contrastive learning and Gumbel-Softmax feedback to guide conditional generation, improving validity/novelty under synthesis constraints.

• 📰 Property-Conditioned Transformer Language Model for De Novo Molecular Generation Using Self-Referencing Embedded Strings.
  Ashish Patel, Rajesh A. Maheshwari, Atul Patel, Ankit Faldu, Anjali Mahavar — Molecular informatics, 2026-09-29
  Introduces a property-conditioned Transformer over SELFIES that injects descriptor embeddings to control molecule generation without RL, achieving 100% validity and matched property distributions.

🧬 Preprints (arXiv + bioRxiv)

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

• 📄 Fold'EM: Direct atomic structure inference from Cryo-EM particles
  Advaith Maddipatla, Märt-Erik Mäeots, Marco Pegoraro, Nikolaus Dräger, Roberto Covino, Sanketh Vedula, Martin Pacesa, Alex M. Bronstein — arXiv, 2026-10-01
  Combines protein generative priors directly with cryo-EM particle images to infer atomic structures without density reconstruction, recovering accurate models with few particles. Resolves heterogeneous conformational states by jointly inferring orientations and structure.

• 🧬 De novo design of flexible protein interactions with GuideFlip
  Yi, K.; Chen, Q.; Zhang, D.; Tian, P.; ...; McLaughlin, S. H.; Tate, C. G.; Jamali, K.; Scheres, S. H. W. — bioRxiv, 2026-09-30
  Co-designs sequence and structure via guided discrete flow matching, enabling de novo binders to flexible/disordered targets with reduced hydrophobic bias. Achieves high experimental hit rates, including a state-selective GPCR nanobody confirmed by cryo-EM.
  Affiliations: MRC Laboratory of Molecular Biology

• 📄 Predictor-Guided Latent Space Codon Optimization for Maximizing Protein Expression
  Alberto Caron, Tianyu Cui, Dmytro S. Lituiev, Mangal Prakash, ..., Nikolay V. Manyakov, Tommaso Mansi, Scott Oloff, Rui Liao — arXiv, 2026-10-02
  Recasts codon optimization as gradient-based search in the latent space of an mRNA language model with expression, stability, and naturalness objectives. Improves wet-lab antibody expression while conserving structure and synonymous fidelity.

• 🧬 RNASeek: A Cross-Phyla Generative Foundation Model for Multipurpose RNA Modeling and Reinforcement Learning-Based Design
  Chen, S.; Fernandes, N.; Li, W. V.; Wang, L.; Jia, Z.; Xiang, J. S. — bioRxiv, 2026-09-28
  Introduces a cross-phyla RNA foundation model coupled with GRPO to design functional ribozymes and 3′ UTRs under user constraints. Predictors reveal interpretable sequence features, and designs validate experimentally with wild-type level or better activity.
  Affiliations: University of California, Riverside

• 📄 pCoMole: Pareto-Constrained Molecule Editing with Discrete Flows
  Tong Chen, Maximilian Holsman, Lin Zhao, Pranam Chatterjee — arXiv, 2026-10-01
  Develops Pareto-constrained discrete flows for sequence editing that enforce hard biochemical feasibility while optimizing multiple properties. Wet-lab tests retain GFP fluorescence after aggressive deletions and preserve Cas9 PAM specificity.

• 🧬 Accurate ab initio gene prediction in eukaryotes with Tiberius in multiple clades
  Gabriel, L.; Bruna, T.; Kaur, A.; Krishnan, A.; ...; Wheat, C. W.; Grigoriev, I. V.; Stanke, M.; Hoff, K. J. — bioRxiv, 2026-10-01
  Extends end-to-end deep ab initio gene prediction across major eukaryotic clades with lineage-specific models, approaching evidence-based pipelines. Delivers high exon/transcript accuracy at a fraction of compute, enabling large-scale genome annotation.
  Affiliations: Institute of Mathematics and Computer Science, University of Greifswald

• 📄 Leveraging secondary-structure information for accurate nucleic acid structure prediction with OFoldNA
  Valhalla Team — arXiv, 2026-09-29
  All-atom diffusion model that incorporates nucleic-acid secondary structure for RNA/DNA folding and protein–nucleic acid co-folding. Achieves SOTA on FoldBench and gains further when base-pairing information is supplied.

• 📄 Generative modeling of intrinsically disordered protein regions by reinforcing sparse autoencoder features
  Jason X. Liu, Sebastian Ibarraran, Frank Hu, Soojung Yang, ..., Anagha Aneesh, Lacramioara Bintu, Alexander R. Dunn, Grant M. Rotskoff — arXiv, 2026-10-01
  Trains a protein LM on 54M predicted IDRs and introduces RL-SAE to steer generation via interpretable sparse-autoencoder features. Enables composable IDR designs with improved predicted localization and transcriptional activity.

• 📄 Continuous Variational Synthesis
  Alan N. Amin, Mattia G. Gollub, Andrei Slabodkin, Elizabeth B. Wood, Eli N. Weinstein — arXiv, 2026-09-28
  Trains continuous-space variational synthesis controllers for DNA and discretizes post hoc to satisfy hardware/wetware constraints. Achieves superior quality–diversity tradeoffs across enzymes, peptides, antibodies, and regulatory elements with in vitro retention.

• 🧬 Generative modeling reveals the connection between nuclear morphology and gene expression
  Wen, S.; Vinas, R.; Bues, J.; Lambert, C.; ...; Wan, Y.; Karthaus, W.; Deplancke, B.; Brbic, M. — bioRxiv, 2026-09-30
  A bidirectional generative framework linking nuclear morphology to single-cell transcriptomes using paired measurements. Quantifies shared variance, identifies morphology-associated genes, and maps spatial neighborhood influences during development.
  Affiliations: EPFL

• 📄 DRIFT: Disentangled Responsive-Invariant Flow Transport for Single-Cell Perturbation Prediction
  Mustapha Bounoua, Giulio Franzese, Pietro Michiardi — arXiv, 2026-09-28
  Disentangles invariant and responsive cellular states and transports only the responsive block via conditional flow matching. Improves single-cell perturbation predictions, including combinatorial and unseen treatments.

• 🧬 A kinetic-aware approach to infer metabolic variations and flux using transcriptomics and metabolomics data
  Zhu, H.; Wan, C.; Yang, M.; Fang, Y.; ...; Kelley, M. R.; Cao, S.; Fishel, M. L.; Zhang, C. — bioRxiv, 2026-10-02
  Estimates condition-specific metabolic flux changes by integrating transcriptomics and metabolomics through a Michaelis–Menten model with MCMC uncertainty. Recovers experimentally observed shifts and outperforms baselines in cancer perturbations.
  Affiliations: Department of Biomedical Engineering, Center for Biomedical Data Sciences

• 🧬 GraPPI: A Self-Supervised Graph Encoder for Transferable Protein--Protein Interaction Modeling
  Song, Z.; Shi, Z.; Sun, G.; Negi, S.; Fausther-Bovendo, H.; Zhao, H. — bioRxiv, 2026-09-29
  Pretrains a heterogeneous graph encoder on protein complexes via masked-edge prediction to learn interaction-aware representations. Transfers to binding-mode classification, affinity prediction, and ΔΔG with competitive or superior performance.
  Affiliations: Department of Biochemistry and Molecular Biology, University of Texas Medical Branch

• 📄 Structure-aware Reinforcement Learning for Protein Directed Evolution
  Zikun Nie, Suyuan Zhao, Yizhen Luo, Siqi Fan, Zaiqing Nie — arXiv, 2026-09-30
  Introduces a structure-aware RL framework for directed evolution using delta-structure fusion and a hierarchical action policy. Outperforms sequence-only MLDE baselines and reveals experimentally validated epistasis in GFP.

• 📄 StabilityArc: Decoding Protein Sequence Embeddings into Generalizable Stability Landscapes
  Aaron L. Feller, Andrew D. Ellington, Claus O. Wilke — arXiv, 2026-09-30
  Decodes frozen ESMC embeddings with a shared transformer to predict per-position substitution effects and pairwise epistasis. Improves cross-protein zero-shot stability prediction and augments supervised fitness models.

• 📄 LEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling
  Biswajit Banerjee, Claudia Alvarez Carreno, Anton S. Petrov — arXiv, 2026-09-29
  Presents evolution-informed tokenization from conserved MSA regions, enabling a compact 200M protein LM to handle long inputs. Outperforms much larger models on remote homology, showing tokenization can substitute for scale.

• 📄 SupraTITO: Transferable Generative Molecular Dynamics for Supramolecular Systems
  Weilong Chen, Nuno Costa, Julija Zavadlav — arXiv, 2026-10-01
  Learns sequence- and topology-conditioned transfer operators for generative MD of peptide self-assembly, enabling long physical rollouts. Generalizes to held-out sequences and concentrations while preserving molecular integrity.

• 📄 Multi-Scale Temporal Flows for Peptide Trajectory Generation
  Xichen Sun, Wentao Wei, Xiaoxi Zhang, Yuedong Yang, Jiahua Rao — arXiv, 2026-10-01
  Trains multi-scale temporal flows with explicit time embeddings to generate peptide trajectories across sampling strides. Captures rare transition frames and matches inter-state fluxes, achieving SOTA on linear and cyclic peptides.

• 📄 Mirror-Score: Calibrated, Inference-only Scoring Exposes the Limits of Sequence-compatibility Ranking in D-peptide Design
  Jiada Li — arXiv, 2026-09-28
  Provides a calibrated, inference-only scoring framework for D-peptide/L-protein complexes, correcting failures of raw sequence-compatibility ranking. Demonstrates that cofolding confidence tracks affinity on a viral-entry family and releases a benchmarking suite.

• 🧬 CHACAM: a cell-cell interaction-guided hierarchical attention model for high-precision cell identity annotation of scRNA-seq data in early C. elegans embryogenesis
  Chen, X.; Ju, X.; Murali, M.; Li, H.; Chen, M.; Zhang, M. Q. — bioRxiv, 2026-09-30
  Integrates scRNA-seq with curated ligand–receptor and contact maps via hierarchical attention and triangulation to resolve sister-cell identities. Achieves near-perfect annotation in C. elegans embryogenesis and identifies new discriminative markers.
  Affiliations: The University of Texas at Dallas