Weekly BioML Digest [July 06, 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)

1921 matched filters -> 20 selected after LLM relevance + novelty ranking.

• 📡 Probe-based identification of metal-binding sites using deep learning representations
  Xu, Shijie, Onoda, Akira — Nature Communications, 2026-06-30
  Introduces PRIME, a hybrid deep learning framework that fuses protein language and structure models with a probe-generation algorithm to accurately predict diverse metal-binding sites (Zn, Ca, K, Na) on proteins, enabling high-throughput metalloproteomics.

• 🛠️ RNAbpFlow: base pair-augmented SE(3) flow matching for conditional RNA 3D structure generation
  Tarafder, Sumit, Bhattacharya, Debswapna — Nature Methods, 2026-06-30
  Presents RNAbpFlow, a base-pair–conditioned SE(3)-equivariant flow-matching model that generates all-atom RNA 3D ensembles without MSAs or templates, improving topology sampling and predictive modeling across large benchmarks.

• ⚗️ Membrane protein solubilization and structure determination using de novo-designed proteins.
  Ljubica Mihaljević, David E. Kim, Pooja Bandawane, Helen E. Eisenach, ..., Melissa J. Caimano, Kelly L. Hawley, Neil P. King, David Baker — Science, 2026-07-02
  Uses deep learning–based RFdiffusion to design de novo amphipathic ‘WRAP’ proteins that solubilize native membrane proteins for detergent-free structural determination, preserving fold and ligand binding.

• 🌿 Design of one-component quasisymmetric protein nanocages
  Lee, Sangmin, Chmielewski, David, ..., Veesler, David, Baker, David — Nature, 2026-07-02
  Combines a parametric cage representation with RoseTTAFold diffusion generative modeling to design one-component quasisymmetric nanocages (T=3–36) validated by EM, advancing programmable biologics delivery.

• 💻 OTalign: Optimal Transport Alignment for Remote Protein Homologs Using Protein Language Model Embeddings.
  Minsoo Kim, Hanjin Bae, Gyeongpil Jo, Kunwoo Kim, Jejoong Yoo, K. Joo — Bioinformatics, 2026-07-01
  OTalign formulates protein sequence alignment as unbalanced optimal transport over PLM embeddings with position-specific gap penalties, outperforming PLMAlign/HHalign on remote-homolog benchmarks and enabling differentiable fine-tuning.

• 📰 Knowledge Distillation of a Protein Language Model Yields a Foundational Implicit Solvent Model.
  J. Airas, Bin Zhang — Journal of chemical theory and computation, 2026-07-03
  Distills evolutionary information from ESM3 into a fast GNN implicit solvent model that drives stable MD and accurately captures folding free energy landscapes and IDP ensembles, unifying folded/disordered states.

• 📰 FoldDoF: Utilizing the Primary Degrees of Freedom of Protein Backbone for Geometric Modeling and Generation.
  Zefeng Zhu, Chen Song — Journal of chemical information and modeling, 2026-07-03
  FoldDoF introduces a peptide-unit rotation manifold representation for protein backbones with differentiable Cartesian conversion, improving generative protein design diversity and length generalization in FrameFlow variants.

• 📰 Dissecting and steering cell dynamics using spatially-informed RNA velocity with veloAgent
  Raghavan, Vishvak, Yoon, Brent, ..., Li, Yue, Ding, Jun — Molecular Systems Biology, 2026-07-02
  veloAgent integrates a deep generative model with agent-based simulations to infer RNA velocity with spatial coherence and gene/cell-specific kinetics, enabling scalable in silico perturbations for trajectory steering.

• 📰 Interpretable deep generative ensemble learning for single-cell omics with Hydra
  Wagle, Manoj M, Liu, Chunlei, ..., Patrick, Ellis, Yang, Pengyi — Molecular Systems Biology, 2026-07-02
  Hydra is an interpretable deep generative ensemble (VAEs + augmentation) for single-cell uni/multimodal data that jointly performs feature selection and cell-type prediction, highlighting biologically meaningful programs.

• 📰 BulkFormer: A large-scale foundation model for bulk transcriptomes.
  Boming Kang, Rui Fan, M. Yi, Chunmei Cui, Qinghua Cui — Cell systems, 2026-07-01
  BulkFormer is a 150M-parameter foundation model for bulk RNA-seq that couples a GNN over gene networks with a Performer encoder, outperforming single-cell–pretrained models across five tasks at lower cost.

• 📰 Tranquillyzer: A Neural Network Framework for Long-read Annotation and Demultiplexing.
  Ayush Semwal, Jacob Morrison, Ian Beddows, Theron Palmer, Mary F. Majewski, H. Josh Jang, Benjamin K. Johnson, Hui Shen — Genomics, proteomics & bioinformatics, 2026-07-01
  Tranquillyzer performs global, context-aware structural inference on noisy long-read single-cell libraries, delivering >99.7% structural filtering accuracy and >91% demultiplexing efficiency across protocols and custom architectures.

• 💻 Enhancing Cross-Context Generalization in Drug Perturbation Prediction with a Multimodal Conditional Diffusion Framework.
  Yanji Ma, Kang Du, Yan Li, Pengyong Li, Liang Yu — Bioinformatics, 2026-07-01
  PertDiff is a multimodal conditional diffusion model that predicts drug-induced transcriptome-wide perturbations using control expression, LLM-derived cell semantics, and pretrained molecular graphs, generalizing to unseen drugs/cell lines with clinical concordance.

• 📰 Generalizable Protein Folding Pathway Exploration with DA2-GRASP: Extending Beyond Miniproteins.
  Yanbing Wen, Hao Dong — Journal of chemical theory and computation, 2026-07-03
  DA2-GRASP learns low-dimensional folding pathways with a VAE and multidirectional gradient-guided sampling, reconstructing atomistic folding of medium proteins with sublinear scaling and accurate mutation thermodynamics.

• 📰 Hybrid Approach to Protein–Protein Complex Affinity Prediction Based on Language Models and Molecular Dynamics
  E. A. Bogdanova, A. Chernukhin, Alexey K. Shaytan — International Journal of Molecular Sciences, 2026-07-01
  HyBind-NN fuses ESM-2 sequence embeddings with 3D Voronoi geometry and MD-derived RMSF via multitask learning to predict protein–protein/peptide binding affinity, improving over sequence-only and rigid-structure models.

• 📰 DeepSSInter: Protein-protein contact prediction with a structure-aware protein language model.
  Derek Huang, Jiamin Lv, Xuan Yao, Peicong Lin, Sheng-You Huang — Protein science : a publication of the Protein Society, 2026-07-01
  DeepSSInter predicts inter-protein residue contacts using structure-aware PLM features (ESM2/SaProt) and triangle-aware modules, surpassing MSA-dependent methods and improving docking when integrated.
  Affiliations: School of Physics, Huazhong University of Science and Technology

• 📰 Structural mining and engineering of metagenome-derived Cas12a orthologs expands the CRISPR genome editing and multiplex diagnostics toolkit.
  Dagang Tao, Bingrong Xu, Sheng Li, Hailong Liu, ..., Xingxu Huang, Xinyun Li, Shuhong Zhao, Shengsong Xie — Molecular therapy : the journal of the American Society of Gene Therapy, 2026-07-01
  AI-guided structural discovery (AlphaFold2) identifies 1,261 Cas12a orthologs and engineers PcuCas12a MAX with high-fidelity editing across species; distinct orthologs enable multiplex trans-cleavage sensors for diagnostics.
  Affiliations: Key Laboratory of Agricultural Animal Genetics, Breeding and Reproduction; The Cooperative Innovation Center for Sustainable Pig Production, Huazhong Agricultural University; ...

• 📰 Machine learning-guided discovery of poly(ethylene terephthalate)-binding modules to enhance durable whole-cell degradation.
  Rui Long, Yaxin Tang, Chengyong Wang, Tangli Yang, Wei Liu, Ling Jiang — Bioresource technology, 2026-07-01
  ML-guided mining of carbohydrate-binding modules discovers a PET-binding module (tCBM13-1) that, when co-displayed with FAST-PETase on E. coli, boosts PET depolymerization by ~43% and maintains >64% activity after 10 cycles.

• 📰 Deep learning unlocks sequence-divergent synthetic promoters to empower Streptomyces natural product engineering.
  Qun Zhou, Ye Wang, Jixin Dong, Xiaomin Zhao, Alan J X Guo, Xiaowo Wang, Yunzi Luo — Metabolic engineering, 2026-07-01
  A deep generative promoter design yields highly active, sequence-divergent Streptomyces promoters validated across hosts, delivering up to 28.6×, 25.7×, and 6.1× titer improvements for PTMs, daptomycin, and rapamycin.
  Affiliations: Frontiers Science Center for Synthetic Biology and Key Laboratory of Systems Bioengineering (Ministry of Education), School of Synthetic Biology and Biomanufacturing; State Key Laboratory of Synthetic Biology, Tianjin University; ...

• 📰 AI-Powered Deep Visual Proteomics Reveals Critical Molecular Transitions in Pancreatic Cancer Precursors.
  Jimin Min, Lisa Schweizer, Gijs Zonderland, Benson Chellakkan Selvanesan, ..., Ishani Ummat, Maximilian T Strauss, Andreas Mund, Anirban Maitra — Cancer discovery, 2026-07-01
  AI-guided Deep Visual Proteomics maps proteomes of pancreatic precursors and normal ducts at ~100 cells/region, revealing early stress, immune, and metabolic programs and detecting KRAS mutant peptides in incidental lesions.
  Affiliations: Laura and Isaac Perlmutter Cancer Center, Department of Medicine; Resolute Bio, Copenhagen; ...

• 📡 Accurate trajectory inference in time-series spatial transcriptomics with structurally-constrained optimal transport
  Bryan, John P., Farhi, Samouil L., Cleary, Brian — Nature Communications, 2026-06-29
  SOCS uses structure-constrained optimal transport to infer trajectories in time-series spatial transcriptomics, preserving contiguous biological units across time and yielding more plausible developmental dynamics.

🧬 Preprints (arXiv + bioRxiv)

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

• 🧬 A generalizable interface-seeded framework for de novo design of functional oligomers
  Chim, H. Y.; Idris, M. O.; Rieger, D.; Schlegel, P.; ...; Tinnefeld, P.; Schoeder, C. T.; Correia, B. E.; Khmelinskaia, A. — bioRxiv, 2026-07-03
  Introduces an AI-based, interface-seeded generative framework to design responsive protein homo-oligomers from modular interaction seeds, enabling chemically triggered and phosphorylation-controlled assemblies. Demonstrates functionalization into ligand-dependent membrane binders and phosphorylation-gated gene regulatory switches.
  Affiliations: Ludwig Maximilian University of Munich

• 🧬 De novo design of RNA pseudoknots with deep learning
  Townley, J.; Kladwang, W.; Baker, D.; Blair, H. M.; ...; Wu, V.; Yu, Z.; Eterna players,; Das, R. — bioRxiv, 2026-07-03
  Demonstrates de novo deep-learning design of RNA pseudoknots using an RNet foundation model and generative AI, matching expert human designs. Designs are experimentally validated by single-nucleotide chemical mapping, compensatory mutagenesis, and cryo-EM.
  Affiliations: Department of Biochemistry, Stanford University School of Medicine

• 🧬 Frequent, context-dependent effects of human genetic variation on Cas9 activity revealed by population-scale GUIDE-seq-2 and deep combinatorial CHANCE-seq profiling
  Tsai, S.; Flory, A. R.; Lazzarotto, C.; Li, Y.; ...; Matsubara, A.; Rashkin, S. R.; Ma, J.; Cheng, Y. — bioRxiv, 2026-07-02
  Presents GUIDE-seq-2 and CHANCE-seq with CHANCE-net, a deep model predicting context-dependent Cas9 off-target effects across human genetic variants. Population-scale profiling reveals variant-modulated editing risks, guiding safer genome-editing designs.
  Affiliations: St. Jude Children's Research Hospital

• 📄 SynLaD: Latent Diffusion for Generating Synthesizable Molecules Conditioned on 3D Pharmacophore Profiles
  Miruna Cretu, John Bradshaw, Patricia Suriana, Saeed Saremi, ..., Kirill Shmilovich, Kangway Chuang, Vishnu Sresht, Colin Grambow — arXiv, 2026-07-01
  SynLaD is a latent diffusion model conditioned on 3D pharmacophore profiles that jointly generates molecules and their synthesis routes. It improves synthesizable, shape-aligned ligand generation by unifying ligand design with reaction-constrained planning.

• 🧬 ConformFlow: scalable normalizing flow for protein conformational ensemble generation
  Liu, Y.; Lin, G.; Chen, M. — bioRxiv, 2026-07-02
  ConformFlow is a sequence-conditioned normalizing flow that learns protein conformational ensemble distributions with exact likelihood training and single-step sampling. It reproduces MD-referenced ensembles and generalizes across proteins at far lower computational cost than diffusion models.
  Affiliations: Purdue University

• 🧬 SpatialFuser: a unified framework for integrative analysis of unpaired spatial multi-omics data
  Cai, W.; Li, W. — bioRxiv, 2026-07-02
  SpatialFuser integrates unpaired spatial multi-omics via a multi-head graph attention autoencoder, geometry-aware optimal transport matching, and contrastive fusion. It resolves fine-grained tissue heterogeneity, aligns cross-slice data, and fuses epigenome–transcriptome–proteome–metabolome signals.
  Affiliations: School of Medicine, Sun Yat-Sen University

• 🧬 Function-guided design of active enzymes
  Hu, M.; Wu, L.; Yang, Y.; Li, F.; Zhu, L. — bioRxiv, 2026-06-29
  EnzymeArt couples a function-conditioned generative sequence model with structure-guided refinement and substrate-aware prioritization to design active enzymes. Across ADH, MDH, and lipase, most designs are functional, with top variants showing strong steady-state kinetics.
  Affiliations: Zhejiang University

• 🧬 Unbalanced Perturbation Dynamics For Cell Fate Design
  Peng, Q.; Wang, Y.; Li, J.; Wang, X.; Xiao, Y.; Zhou, P. — bioRxiv, 2026-07-04
  U-Pert is an unbalanced generative model that jointly learns transcriptomic state transitions and cell-number dynamics from unpaired single-cell perturbation snapshots. It enables accurate forward prediction of unseen perturbations and inverse design of interventions for specified molecular and population outcomes.
  Affiliations: Center for Machine Learning Research, Peking University

• 📄 Towards Generalizable and Evidential Nuclear Magnetic Resonance-Based Molecular Structure Elucidation via Large Language Model Agent
  Zheng Fang, Chen Yang, Yusen Tan, Yunpeng Zhao, ..., Xiaojian Wang, Wenjie Du, Yuqiang Li, Jun Xia — arXiv, 2026-06-29
  NMRAgent is an LLM-driven evidential reasoning agent that plans, proposes, and verifies molecular structures from NMR spectra against chemical knowledge graphs and tools. It markedly improves scaffold-split accuracy and elucidates previously unknown natural products with transparent, evidence-based reasoning.

• 🧬 MintCNA: A Unified Framework for Integrative Copy Number Profiling with Single-Cell Multi-Omics Data
  Bao, W.; Qin, F.; Xiao, F. — bioRxiv, 2026-07-01
  MintCNA unifies scDNA-seq and scRNA-seq for copy-number profiling via an attention-guided convolutional autoencoder and multivariate change-point detection with missingness-aware CUSUM. It improves sensitivity and specificity of single-cell CNA calls and breakpoint detection in tumors.
  Affiliations: Department of Biostatistics, College of Public Health and Health Professions and College of Medicine

• 📄 Structure-Regularized Interpretable TCR-Epitope Prediction
  Jiarui Li, Zixiang Yin, Yunbei Zhang, Janet Wang, Samuel J. Landry, Zhengming Ding, Ramgopal R. Mettu — arXiv, 2026-06-29
  TCR-SRIM combines protein language embeddings with interpretable contact prototypes and structure regularization for TCR–epitope binding prediction. It achieves SOTA generalization to unseen epitopes with residue-level interaction explanations and assesses the impact of predicted vs. experimental structures.

• 📄 Enerzyme: A Framework for Efficient Training of Reactive Neural Network Potentials for Enzyme Catalysis with Application to Methyltransferases
  Weiliang Luo, Heather J. Kulik — arXiv, 2026-07-01
  Enerzyme provides electrostatics-aware neural network potentials and automated reactive dataset generation for enzyme QM-cluster models. With <1,000 datapoints it reproduces reaction energetics and TS structures for methyltransferases, capturing polarization and charge transfer with multitask-learned charges.

• 📄 ReactionAtlas: Ab origine exploration of chemical reaction networks with machine learning
  Stefan Gugler, Max Eissler, Khaled Kahouli, Klaus-Robert Müller — arXiv, 2026-06-29
  ReactionAtlas builds reaction networks ab origine by proposing reactions with a generative model and filtering to valid transition states with a DFT-trained ML force field. Starting from prebiotic seeds, it maps ~47k reactions among ~12k compounds in carbohydrate chemistry, including formose cycles.

• 🧬 ConfDock: Atom-specific Uncertainty Quantification for Molecular Docking via Conformal Prediction
  Hao, H.; Elhendawy, N.; Wang, Y.; Lu, C. — bioRxiv, 2026-07-01
  ConfDock combines GNN-based quantile estimation with split conformal calibration to deliver atom-specific uncertainty intervals for docking poses. It achieves substantially narrower, structure-aware intervals while maintaining finite-sample coverage across multiple docking engines.
  Affiliations: University of Illinois Chicago

• 🧬 OpenGerminal: an open-source implementation of the Germinal antibody design pipeline
  Han, B.; Li, S. — bioRxiv, 2026-06-29
  OpenGerminal open-sources the Germinal antibody design pipeline, replacing proprietary components and adopting AbLang for antibody LM guidance. It boosts cofolding pass rates on VHH targets while maintaining structural confidence, broadening accessibility for epitope-directed antibody design.
  Affiliations: University of Virginia

• 🧬 Benchmarking the translational potential of AI-based drug-resistance prediction from Mycobacterium tuberculosis whole-genome sequencing data
  Liu, C.; Zhu, H.; Zhou, P.; Thanh, N. T.; ...; Adisasmito, W.; Zheng, X.; Wang, H.; Yang, Y. — bioRxiv, 2026-07-03
  Constructs the largest unified benchmark for AI-based M. tuberculosis drug-resistance prediction from WGS, harmonizing 54k records across drugs and lineages. Finds classical ML (e.g., XGBoost) leads in AUPRC/F1 and highlights persistent bottlenecks for emerging drugs and XDR-TB.
  Affiliations: Center for Single-Cell Omics, School of Public Health; Hainan International Medical Center, Shanghai Jiao Tong

• 📄 MolSafeEval: A Benchmark for Uncovering Safety Risks in AI-Generated Molecules
  Tong Xu, Xinzhe Cao, Zhihui Zhu, Keyan Ding, Huajun Chen — arXiv, 2026-07-01
  MolSafeEval benchmarks safety risks in AI-generated molecules by integrating a molecular safety knowledge graph with LLM-based reasoning to detect and explain toxic/reactive hazards. It standardizes safety evaluation across unconditional, property, target-based, and text-conditioned generation tasks.

• 🧬 A High-Confidence Atlas of Protein Methylation Enables AI-Driven Detection of Methylated Peptides
  Wang, S.; Hartmaring, Y.; Schlaffner, C. N.; Bowler-Barnett, E.; ...; Sun, Z.; Renard, B. Y.; Jones, A. R.; Vizcaino, J. A. R. — bioRxiv, 2026-07-04
  Curates a high-confidence atlas of human protein methylation via stringent FLR control and fine-tunes a deep spectrum model (AHLF-Methylation) by transfer learning. The approach improves detection of methylated peptides and integrates curated sites into UniProt and PeptideAtlas.
  Affiliations: University of Liverpool

• 🧬 VirProtRAG: Literature-grounded viral protein function annotation with retrieval-augmented generation
  Guan, J.; Shang, J.; Peng, C.; Sun, Y. — bioRxiv, 2026-07-04
  VirProtRAG fuses hybrid literature retrieval with rank-aware re-ranking and LLM generation to annotate viral protein functions with linked evidence. It expands Swiss-Prot viral annotations and supports sequence- and text-based querying with improved interpretability.
  Affiliations: City university of Hong Kong

• 📄 Diffeomorphic Optimization
  Ludwig Winkler, Andrew Leaver-Fay, Joseph Kleinhenz, Pan Kessel — arXiv, 2026-07-01
  Introduces diffeomorphic optimization that performs manifold-aware gradient descent via generative model mappings, keeping optimization on the data manifold. Applied to protein design, it improves secondary-structure targeting and binding optimization versus guidance baselines at lower cost.