Weekly BioML Digest [August 31, 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)

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

• 🤖 A knowledge-driven framework for predicting single-cell responses for unprofiled drugs
  Feng, Jinghao, Zhao, Ziheng, ..., Zhang, Ya, Xie, Weidi — Nature Machine Intelligence, 2026-08-26
  Introduces MAP, a knowledge-driven framework that fuses a large biological mechanism graph with pretrained single-cell models to predict gene-expression responses for unprofiled drugs, enabling zero-shot, mechanism-aware perturbation modeling and pathway-level screening.

• 🤖 Large language models as uncertainty-calibrated optimizers for experimental discovery
  Ranković, Bojana, Griffiths, Ryan-Rhys, Schwaller, Philippe — Nature Machine Intelligence, 2026-08-28
  Presents GOLLuM, a Bayesian-objective training paradigm that calibrates LLM uncertainty to act as a reliable optimizer for experimental discovery, outperforming traditional Bayesian optimization across synthesis and molecular design tasks.

• 📰 Dockformer: A Transformer-Based Molecular Docking Paradigm for Large-Scale Virtual Screening.
  Zhangfan Yang, Junkai Ji, Shan He, Jianqiang Li, Tiantian He, Ruibin Bai, Zexuan Zhu, Yew-Soon Ong — IEEE transactions on neural networks and learning systems, 2026-08-28
  Develops Dockformer, a transformer that jointly predicts protein–ligand binding poses and confidence, achieving >100× speedups and higher accuracy than classical docking on PDBbind/PoseBusters for large-scale virtual screening.

• 📰 Grouper: symmetry-aware functional-group graph representations for generative exploration of chemical space
  Nehil-Puleo, Kieran, Craven, Nicholas C., McCabe, Clare, Cummings, Peter T. — npj Computational Materials, 2026-08-27
  Proposes Grouper, a symmetry-aware functional-group graph representation and combinatorial framework that efficiently generates and analyzes chemically valid molecular spaces, enabling tractable end-to-end discovery workflows.

• 📰 Symphony-Bind: Prediction of Protein Binding Sites for 11 Representative Small Molecules and Ions via Fine-Tuning Protein Language Models and Grouped Multi-Task Learning.
  Yizhe Li — ACS omega, 2026-08-25
  LoRA-enhanced ESM2 fine-tuned under grouped multi-task learning improves prediction of protein binding sites for diverse ligands/ions, balancing efficiency and accuracy.
  Affiliations: AIEN Institute, Shanghai Ocean University

• 📰 3D Geometric Equivariant Graph Neural Networks with Contrastive Learning for Protein–Ligand Binding Affinity Prediction
  Li, Gaili, Yuan, Yongna, Chen, Xiuping, Wang, Ping — Interdisciplinary Sciences: Computational Life Sciences, 2026-08-26
  Introduces a 3D geometrically equivariant GNN with contrastive learning to predict protein–ligand binding affinity, improving robustness to conformational variability on PDBbind and CSAR-HiQ benchmarks.

• 📰 Hierarchical Graph Representation Learning for Protein–Protein Interaction Site Prediction
  Xu, Wenjun, Wang, Xiaosong, ..., Sun, Yunyun, Gu, Lichuan — Interdisciplinary Sciences: Computational Life Sciences, 2026-08-26
  Presents HGRL-PPIS, a hierarchical protein graph (atom–residue–protein) with transformer-based cross-hierarchy messaging and prototypical training to predict protein–protein interaction sites, outperforming prior methods.

• 📰 METI-FS: a multi-evidence temporal integration framework for biomarker discovery in small-sample time-series transcriptomics
  Zhang, Zhen, Ma, Tengjiao, ..., Qin, Meirong, Wang, Ping — BMC Bioinformatics, 2026-08-25
  Introduces METI-FS, a leakage-controlled, multi-evidence time-series feature selection pipeline (maSigPro+WGCNA+effect-size testing) that stabilizes downstream ML biomarker discovery in small-sample multi-omics.

• 📰 DMFF: a deep learning-based multi-omics fusion framework for survival prediction and subtype classification in breast cancer
  Zhang, Shumei, Zhang, Yue, ..., Wang, Qiutong, Yang, Wen — Scientific Reports, 2026-08-27
  Develops DMFF, a deep multi-omics fusion framework (denoising autoencoders + PPI-GCN + attention fusion + Cox/ranking losses) for breast cancer survival prediction and de novo subtype discovery with leakage-controlled validation.

• 📰 An Uncertainty-Guided Evidential Deep Learning Framework for Reliability-Aware Multimodal Fusion in Cancer Prognosis
  Yalu Huang, Yu-Shuai Yuan, Wenbin Ye, Wenlong Ming — Mathematics, 2026-08-25
  Evidential deep learning fuses multimodal cancer data with uncertainty-aware source weighting, improving survival prediction and enabling reliability-aware decision support.

• 📰 Finetuning Foundation Models for Temporal Clinical Transcriptomics Data.
  Sachin Mathur, Alexander Kagan, Peyman Passban, Hamid Mattoo, Euxhen Hasanaj, Ziv Bar-Joseph — Bioinformatics (Oxford, England), 2026-08-27
  Fine-tunes foundation gene embeddings and integrates them into temporal GNNs to model treatment response trajectories across diseases, recovering known mechanisms and uncovering responder/non-responder pathway differences.
  Affiliations: R&D Data and Computational Sciences, Sanofi; Department of Statistics, University of Michigan; ...

• 📰 Development and validation of machine learning model for selecting the optimal population pharmacokinetic model for vancomycin.
  Hayato Akamatsu, Yukinobu Kodama, Ayaka Terai, Masanobu Imamura, Kazutaka Oda, Hirofumi Jono, Kaname Ohyama — Journal of infection and chemotherapy : official journal of the Japan Society of Chemotherapy, 2026-08-26
  Builds and validates an ML model to select the optimal population PK model for vancomycin AUC-based dosing, using LASSO feature selection, nested CV, and SHAP to improve initial regimen planning within PAT software.
  Affiliations: Department of Hospital Pharmacy, Nagasaki University Hospital; Department of Molecular Pathochemistry, Graduate School of Biomedical Sciences; ...

• 📰 Development and temporal validation of AI-enhanced HEp-2 indirect immunofluorescence image analysis for specific autoantibody prediction in systemic autoimmune rheumatic diseases.
  Patrick Vanderboom, Surendra Dasari, Marlon J Sandino-Bermúdez, Anne Tebo, Melissa Snyder, Alí Duarte-García — Arthritis & rheumatology (Hoboken, N.J.), 2026-08-28
  AI models trained on ANA HEp-2 images predict disease-specific autoantibodies (e.g., dsDNA, Sm), improving negative-case classification over pattern-based heuristics and remaining stable post-assay change.
  Affiliations: Department of Laboratory Medicine and Pathology, Mayo Clinic; Department of Quantitative Health Sciences, Mayo Clinic; ...

• 📰 AI-derived tumor-infiltrating lymphocytes enhance the prediction of pathologic complete response in early-stage triple-negative breast cancer
  Pan, Xiaoxi, Ercan, Caner, ..., Huo, Lei, Yuan, Yinyin — npj Breast Cancer, 2026-08-29
  Automates stromal TIL quantification on TNBC biopsies via a context-aware histology pipeline, modestly improving pCR prediction over manual scoring in ARTEMIS and external cohorts.

• 📰 Deep learning prediction of intratumoral Fusobacterium from digital colorectal cancer H&E slides
  Xiang, Huairong, Hu, Lingling, ..., Sun, Liping, Tu, Huakang — BMC Medical Imaging, 2026-08-28
  Uses pathology foundation models to predict intratumoral Fusobacterium and species-level labels from H&E slides, revealing spatial microbial niches linked to mucinous/MSI phenotypes and tumor–stroma interfaces.

• 📰 Transcriptome-based machine learning classification of pediatric germ cell tumors using nanopore RNA sequencing
  Bezerra, Ana Flavia Souza Peres, Bhakta, Nickhill, ..., Wang, Jeremy R., Pinto, Mariana Tomazini — BMC Cancer, 2026-08-26
  Nanopore transcriptomics with ML accurately classify pediatric GCTs and subtypes from FFPE, supporting scalable, on-site molecular pathology in LMIC settings.

• 🖥️ A patient’s longitudinal history reconstructs their unmeasured molecular profile
  Tanis, Stephanie, Lopez Alvarez, Samantha, Davidson, Natalie R. — Nature Computational Science, 2026-08-26
  Introduces PULSE, an AI framework that reconstructs a patient’s unmeasured molecular profile from longitudinal clinical history, enabling virtual multi-omic inference for personalized care.

• 🚀 Programmable protein degraders enable selective knockdown of pathogenic β-catenin subpopulations in vitro and in vivo
  Tianzheng Ye, Azmain Alamgir, C. Robertus, Darianna Colina, ..., David Putnam, Christopher A. Alabi, Pranam Chatterjee, M. DeLisa — Science Advances, 2026-08-28
  Combines a protein language model–guided peptide design algorithm (SaLT&PepPr) with uAb degraders to selectively target pathogenic β-catenin subpopulations, validating in vitro and in vivo with LNP-mRNA delivery.

• 📰 DyAb: sequence-based antibody design and property prediction in a low-data regime.
  Joshua Yao-Yu Lin, Jennifer L Hofmann, Andrew Leaver-Fay, Wei-Ching Liang, ..., Vladimir Gligorijevic, Andrew Watkins, Kyunghyun Cho, Nathan Frey — mAbs, 2026-12-31
  Builds DyAb on a pretrained protein language model to predict pairwise antibody affinity differences and guide low-data sequence design, yielding variants with >10× improved binding.
  Affiliations: Prescient Design, Genentech; Department of Antibody Engineering, Genentech; ...

• 📰 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
  Combines genomic and protein language models into a meta-learner to predict cognate and near-cognate human translation initiation sites, improving AUG/non-AUG TIS detection across datasets.
  Affiliations: Center for Bioinformatics, Saarland Informatics Campus

🧬 Preprints (arXiv + bioRxiv)

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

• 📄 RIBOSPAN: A Long-Context RNA Foundation Model for Versatile RNA Modeling
  Ziyuan Wang, Bohao Tang, Fei Zhang, Shuo Han, Pengfei Liu — arXiv, 2026-08-24
  Introduces RIBOSPAN, a 1.61B-parameter long-context RNA foundation model (10k nt) that sets state-of-the-art performance on RNA property prediction and zero-shot mutation-fitness modeling. Also presents a conditioned discrete-diffusion framework for full-length mRNA generation and redesign, enabling protein-preserving CDS optimization.

• 🧬 CodonMamba: a foundation model for programmable mRNA coding sequence design
  Lang, M.; Fang, X.; Wang, Z.; Chen, M.; ...; Zhu, X.; Tam, K. Y.; Zhang, J.; Li, X. — bioRxiv, 2026-08-24
  Presents CodonMamba, a codon language model that achieves state-of-the-art across 12 mRNA prediction tasks and enables programmable CDS design via inference-time codon usage priors. This steerable generation framework supports cross-host retargeting and multi-criteria optimization without retraining.
  Affiliations: Hangzhou Institute of Medicine, Chinese Academy of Sciences

• 📄 RegimeFormer: A Large Protein Model of Global Perturbation Regimes
  Siyuan Ma, Yi Chai, Yi Wu, Qixin Zhang, ..., Yun Liu, Yang Liu, Tingting Zhu, Dacheng Tao — arXiv, 2026-08-27
  RegimeFormer is a large protein perturbation model trained on a global sequence atlas that learns reproducible protein-level perturbation regimes. It improves substitution-specific predictions under low-homology settings and provides priors that enhance downstream transcriptomic and drug-response modeling.

• 🧬 MultiFlow: coupled flow matching for predicting single-cell multiomic perturbation responses in unseen cellular contexts
  Wang, H.; Zhang, C.; Zhang, M.; Nie, X.; Liu, Q. — bioRxiv, 2026-08-25
  MultiFlow couples flow-matching models for paired RNA and ATAC to jointly generate and predict perturbation responses in unseen cellular contexts. It outperforms modality-specific methods while preserving cross-modal coordination such as concordant peak–gene effects and neighborhood structure.
  Affiliations: Yale University

• 🧬 Gene expression inference from cell-free DNA using uncertainty-aware deep learning
  Patton, R. D.; McDeed, A. P.; Netzley, A.; Pawar, A.; ...; MacPherson, D.; Haffner, M. C.; Nelson, P. S.; Ha, G. — bioRxiv, 2026-08-28
  Develops Triton+Proteus, an uncertainty-aware deep learning framework that infers gene expression from standard-depth cfDNA WGS. It reconstructs tumor transcriptional programs and therapeutic target activity across cancers, enabling minimally invasive functional genomics in precision oncology.
  Affiliations: Fred Hutchinson Cancer Center

• 🧬 FlexiTAC enables controllable PROTAC linker generation across diverse structural settings using a Bayesian flow network with posterior guidance
  Li, Y.; Zhao, Y.; Zhou, L.; Huang, C.; ...; Qin, Z.; Fan, K.; Yang, J.; Cao, D. — bioRxiv, 2026-08-30
  Introduces FlexiTAC, a Bayesian flow network that generates 3D PROTAC linkers with controllable rigidity directly from warhead/E3 contexts, supported by PROTAC-3D and PROTAC-Bench. It delivers higher validity and pose quality than 3D baselines and allows differentiable flexibility steering without retraining.
  Affiliations: Tongji University

• 🧬 Vipsania: Unsupervised Deep Gene Finding
  Krieg, R.; Becker, F.; Saenko, S.; Diehl, J.; Stanke, M. — bioRxiv, 2026-08-30
  Vipsania is the first unsupervised deep gene finder, embedding a differentiable HMM in a sequence model to learn gene structures from unannotated eukaryotic genomes. It surpasses supervised methods across clades and adapts to non-standard genetic codes, enabling scalable pan-eukaryotic annotation.
  Affiliations: University of Greifswald

• 📄 PHASE: encoding global protein ensembles with local Hamiltonians and all-atom backmapping
  Daniele Angioletti, Marco Nobile, Matteo Carli, Vittorio Limongelli — arXiv, 2026-08-24
  PHASE converts atomistic protein conformational ensembles into explicit local-Hamiltonian statistical models that reproduce microstate statistics and organize activation landscapes. A backmapping module reconstructs all-atom structures, providing an interpretable, generative route to ensemble modeling with QUBO encodings.

• 🧬 Tree-aware conditional language modeling recovers mutational patterns of viral evolution
  Polunina, P. V.; Maier, W.; Rubin, A. F. — bioRxiv, 2026-08-26
  evoPLM-Tree is a phylogeny-aware conditional autoregressive protein language model that generates descendant sequences from ancestral context. Applied to SARS-CoV-2 spike, it recovers lineage-specific mutational patterns aligned with deep mutational scanning constraints.
  Affiliations: University of Melbourne

• 🧬 A multimodal perturbation atlas defines the phenotypic resolution of cellular morphology.
  Liu, C.; Hillsley, A.; Sekhar, M.; Jones, C. A.; ...; Mehta, S. B.; Royer, L. A.; Gomez-Sjoberg, R.; Leonetti, M. D. — bioRxiv, 2026-08-25
  Builds a multimodal perturbation atlas (~65M cells) combining live/fixed microscopy, quantitative phase imaging, and scRNA-seq for 1,000 CRISPR KOs, with deep learning analyses. Phase imaging yields higher phenotypic resolution than matched-cost fluorescence or scRNA-seq, advancing live-cell morphology as a high-precision readout.
  Affiliations: Biohub, San Francisco

• 🧬 Chemi-Proteome Language Attention Network Empowers Fragment-Based Ligand Interactome and Binding Sites Discovery with Evidence
  Liao, B.; He, J.; zhao, M.; Cui, X.; ...; Dong, C.; Sun, H.; Zhang, L.; Zhang, J. — bioRxiv, 2026-08-30
  C-PLANK trains a bilinear attention network on cellular chemoproteomics to predict fragment–protein interactions with interpretable residue–atom fingerprints and a systems-level evidential score (CISI). It outperforms current frameworks and guided discovery of a cellular SIRT3 agonist.
  Affiliations: Shanghai Jiao Tong University

• 📄 LUCAID: Agentic Multimodal AI for Lung Cancer Precision Pathology
  Marie-Lisa Eich, Kai Standvoss, Timo Milbich, Alexander Möllers, ..., Lukas Ruff, Maximilian Alber, Frederick Klauschen, Simon Schallenberg — arXiv, 2026-08-24
  LUCAID is an agentic multimodal AI system for precision lung cancer pathology that integrates nine specialized modules from QC to biomarker scoring and structured reporting. It achieves 93% concordance with expert-adjudicated decisions in prospective validation, surpassing experienced pathologists.

• 📄 Optimizing RNA yield using deep neural networks coupled to massively parallel screening
  Dinghai Zheng, Justin Hong, Jun Wang, Adrien Villain, Mickaël Costallat, Fernando Ulloa Montoya, Vikram Agarwal — arXiv, 2026-08-24
  Combines massively parallel sequencing assays with deep CNNs to predict in vitro transcription RNA yield from promoter-adjacent DNA sequence. The model (r=0.94) enables pre-experimental ranking of designs to improve mRNA manufacturability for vaccines and therapeutics.

• 📄 Packora: Systematic Design for Generative Molecular Crystal Structure Prediction
  Nayoung Kim, Kiyoung Seong, Sungsoo Ahn — arXiv, 2026-08-27
  Packora is a flow-based generative model that jointly predicts atomic coordinates and lattices for molecular crystal structures, supporting multi-component and organometallic systems. It outperforms baselines on CSP generation and ranking benchmarks with systematic design choices for robust performance.

• 📄 MolEmb: Multimodal Large Language Models Can Be Strong Molecular Embedding Models
  Xinjian Zhao, Xiangru Jian, Yaoyao Xu, Xiaozhuang Song, Wei Pang, Lei Bai, Tianshu Yu — arXiv, 2026-08-24
  MolEmb adapts multimodal LLMs into general molecular embedding models by contrastively aligning molecular profiles with text, enabling context-conditioned representations. It rivals specialist encoders on property prediction and supports cross-modal molecule–text retrieval and diagnostic context-aware retrieval.

• 📄 CIR-DDG: backbone-agnostic residual correction of antibody-antigen affinity changes with explicit cross-chain geometry
  Weilun Yu, Zhiheng Zou, Yonggui Huang, Honggang Qi, ..., Yu Xiao, Gang Xu, Jiangtao Wang, Xi Chen — arXiv, 2026-08-27
  CIR-DDG is a lightweight residual adapter that augments fixed backbones with explicit cross-chain geometric descriptors to predict mutation-induced antibody–antigen ΔΔG. It consistently improves correlations and transfers to deep mutational scanning benchmarks without retraining.

• 📄 GRAS: Guided Reduced-Variance Proposals and Adaptive Selection for Training-Free Reward Alignment in Discrete Diffusion
  Kwanyoung Kim — arXiv, 2026-08-27
  GRAS provides training-free reward alignment for discrete diffusion via reduced-variance gradient estimates and adaptive resampling temperature. It achieves best-in-class training-free optimization on regulatory DNA and protein design tasks, matching or surpassing fine-tuned baselines.

• 🧬 Molecular Determinants of Functional Bacterial sRNA-mRNA Interactions Revealed by Integrating RNA Interactomes and Interpretable Machine Learning
  Safari, F.; Mediati, D. G.; Alquethamy, S.; Tree, J. J.; Vafaee, F. — bioRxiv, 2026-08-27
  Integrates Hfq-CLASH interactomes with multi-omics and interpretable ML to predict functional sRNA–mRNA regulatory outcomes. Feature attributions reveal that RNA structure, accessibility, and protein-binding context—especially target-side Hfq occupancy—govern regulatory efficacy beyond base pairing.
  Affiliations: School of Biotechnology and Biomolecular Sciences, University of New South Wales (UNSW Sydney)

• 📄 Interpreting Latent Protein Language Model Features with Geometric Annotations
  Siddharth Setlur, Djordje Mihajlovic, Darrick Lee — arXiv, 2026-08-26
  Uses sparse autoencoders on ESM-2 with geometric annotations of the Cα backbone to interpret latent pLM features at residue level. Identified geometric neurons generalize beyond database labels and causally modulate predicted contact maps upon ablation.

• 🧬 Interpretable Forecasting of Kidney Cancer Progression via Generative AI and Symbolic Reasoning
  Prol-Castelo, G.; Syrri, E.; Manginas, N.; Manginas, V.; ...; Katzouris, N.; Paliouras, G.; Valencia, A.; Cirillo, D. — bioRxiv, 2026-08-26
  Combines a VAE to synthesize pseudo-longitudinal RNA-seq trajectories with symbolic rule induction to learn interpretable finite-state models of kidney cancer stage progression. The symbolic forecaster approaches LSTM performance while yielding human-readable rules and probabilistic transition timing.
  Affiliations: Barcelona Supercomputing Center (BSC)