I am happy to share that our work on Metrn and Metrnl structure prediction, domain identification and SNP analysis got recently published in the #jbsd thanks to @KulkarniMj @reemabanarjee and all my co-authors @csir_ncl @AcSIR_India #meteorin #metrnl doi.org/10.1080/07391102.202…

TopoFlow: Evolutionarily Conditioned Flow Matching for Protein Conformational Ensemble Generation 大域的な構造多様性と局所的な柔軟性を捉え、タンパク質の構造アンサンブルを生成するフローマッチング手法 分子動力学ベンチマークで高性能 github.com/iobio-zjut/Topofl… biorxiv.org/content/10.64898…

UCSF’s Kevan Shokat, PhD, just won the $1M Stephenson Global Prize for research in pancreatic cancer. He cracked KRAS, a cancer driver long considered “undruggable.” Drugs built on his work now treat lung and pancreatic cancer. ow.ly/xVlw50ZRSXh

Feeling proud of where SpliceCraft is at right now. Still got more details to add and UI jank to solve, but powerful af and open source. Made, with love, for this community. Enjoy. SpliceCraft.bio

Good morning mito-folks, don't miss this rich selection of papers related to #mitochondrialmedicine presorted by @Bims_BiomedNews biomed.news/bims-mitmed/2026… The predicted interactome of the human mitochondrial proteome

サイトカインを腫瘍の中でだけONにする「プロカイン」 サイトカインは抗腫瘍免疫を活性化するが標的組織以外にも毒性がある。今回人工タンパク質を用いることで、腫瘍で多いプロテアーゼによって切断されたときだけ活性化するプロドラッグを作成。ANDゲートなども可能 biorxiv.org/content/10.64898…

🎉 Out today in Cell: the paper I've been looking forward to sharing for a long time 🎉 We used cryo-ET to find a molecular machine nobody knew existed on the surface of a minimal bacterium, and worked out what it does. 🧵(1/6) #TeamTomo #cryoET

High-throughput physics-based enzyme engineering biorxiv.org/content/10.64898…

De novo design of cysteine proteases doi.org/10.1101/2025.11.21.6… Code: Design pipeline and computational analysis - workflow and analysis scripts; model weights are not bundled. github.com/hojaec/RFD2_MI_cy… Kinetics and mammalian-cell image analysis - notebooks for the experimental assays. github.com/hojaec/Cysteine-p…

AI-designed proteases make the cut ✂️ RFD2-MI scaffolds catalytic atoms + peptide. ProteinMPNN co-designs enzyme + substrate seqs. AlphaFold3 checks bound + unbound structures. 13/69 selected designs cut paired substrates. #ProteinDesign

1/7 We are pleased to report the 2026 VDJdb release, just published in Nucleic Acids Research: 203,230 TCR:pMHC records over 2,090 epitopes and 236 MHC alleles, 37-fold the first release, 4-fold our 2022 COVID-19 compendium. #immunology #bioinformatics academic.oup.com/nar/advance…

Thank you, Nil and @m_madan_babu, for your cool commentary on our recent Nature paper. Very well done! The paper: nature.com/articles/s41586-0… The commentary: ‘Superdark’ protein looks like a G-protein-coupled receptor — but shows unconventional behaviour nature.com/articles/d41586-0…

AI and ML Papers Explained (HUGE list): github.com/dair-ai/ML-Papers… Compiled by @omarsar0 @dair_ai 💯🏆🚀💥 #ML #MachineLearning #DataScience #Algorithms #NeuralNetworks #DeepLearning #Mathematics

Excited to share T-REX 🦖: Target-adaptive Rescue-Explore-eXploit, an agentic campaign controller for high-throughput de novo protein binder design! T-REX treats protein design as an online agentic allocation problem to leverage diverse tools with efficient GPU usage 🧵

De Novo Design of Protease-Activatable Cytokine Prodrugs 🚀 New preprint from David Baker!🚀 1. The authors designed cytokine prodrugs that remain largely inactive until tumor-associated proteases remove a custom protein mask. In mouse tumor models, an optimized IL-21 prodrug controlled tumors while reducing treatment-associated weight loss. 2. Their central idea is to design each mask against the cytokine’s receptor-binding surface. This lets the team tune how strongly the mask blocks signaling, rather than relying on a naturally occurring receptor fragment with a fixed shape and affinity. 3. The team used RFdiffusion2 to generate masks, ProteinMPNN to design their sequences, and AlphaFold2 predictions to select candidates. They then improved the mask interfaces experimentally and linked the masks to IL-21-like and IL-2-like cytokine mimics through MMP2/MMP9-cleavable peptides. 4. The optimized constructs, Pro21 and Pro2, showed approximately 197-fold and 150-fold differences in signaling activity between their intact and protease-cleaved states. Protease treatment restored activity close to that of the corresponding unmasked cytokine mimics. 5. Masking also sharply reduced receptor binding. For the IL-21 mimic, measured IL-21R binding changed from a KD of 21.8 pM to 15.1 µM after incorporation into Pro21. 6. The authors extended the approach to a split IL-2 mimic. Masks prevented its two fragments from assembling prematurely, suppressing activity by at least 800-fold at the highest concentration tested; MMP2 cleavage restored signaling. 7. By attaching targeting domains to the split fragments, the team created a two-input system: protease cleavage and localization on antigen-expressing cells were both needed for strong activation. A version targeting PD-L1 and HER2 showed its highest reporter activity on double-positive cells. 8. Linker design proved as consequential as mask design. More readily cleaved Pro21 variants controlled tumors but also caused weight loss. An intermediate-sensitivity linker, L4, gave a better balance of antitumor activity and tolerability in the tested mice. 9. In MC38 colon and KPC.1 pancreatic tumor models, Pro21-L4 improved tumor control and survival while maintaining body weight better than the unmasked IL-21 mimic. A non-cleavable control had weaker antitumor activity, supporting a role for protease-dependent activation. 10. Adding PD-L1 targeting allowed Pro21-L4 to retain tumor control in the KPC.1 model at a dose equimolar to the unmasked mimic. It also produced less splenic STAT3 activation, consistent with reduced peripheral signaling. These are preclinical findings from small mouse groups, rather than evidence of clinical safety or efficacy. 📜Paper: doi.org/10.64898/2026.09.19.… #ProteinDesign #CancerImmunotherapy #Cytokines #ComputationalBiology

De novo design of protease-activatable cytokine prodrugs biorxiv.org/content/10.64898…

Next week: a major update to our open-source monomer database + explorer! Now featuring ~4,000 peptide monomers, launching alongside our talk at @BoulderPeptide, where Julie Owen will take us on a guided tour through chemical space. Another surprise too, so watch this space!👀

Towards Accurate Prediction of Mutation-Induced Changes in Protein Structure 1. The paper builds a quantitative framework to measure how single amino-acid mutations deform protein structures, using matched wild-type and mutant X-ray crystal structures from the PDB and explicitly controlling for “native” structural variability. 2. Key dataset contribution: 6,967 wild-type–mutant sequence pairs, leveraging 27,200 wild-type and 16,176 mutant X-ray structures, with each wild-type sequence required to have at least five experimental “duplicate” structures to estimate intrinsic fluctuations. 3. Core metric: for each residue i, local deformation Di is defined from RMS changes in Cα–Cα distances to its local Voronoi neighbors (rather than global RMSD), reducing sensitivity to rigid-body motion and focusing on local rearrangements around mutations. 4. Innovation that removes thermal/experimental noise: normalized mutation-induced deformation eDi = (average deformation between wild-type duplicates vs mutant structures) divided by (average deformation among wild-type duplicates). This isolates mutation-specific effects from baseline structural variability. 5. Main structural finding: mutation-induced changes are strongly localized. Averaged across proteins, deformation peaks at the mutation site and decays rapidly with distance, approaching a near-baseline plateau for spatial distances r > ~12–15 Å (and similarly within ~±10 residues along sequence). 6. Distributional result: eDi at mutation sites is roughly exponential—many mutations look “neutral” structurally (eDi ≈ 1), while a smaller fraction induce substantially larger local deformations. 7. AlphaFold3 evaluation: predicted mutant structures were generated (AF3 v3.0.1; up to 20 seeds per mutant, selecting top-ranked per seed). Correlation between predicted and experimental normalized deformation at the mutation site is moderate overall (Pearson ρ ≈ 0.55) but collapses for strongly perturbative mutations (down to ρ ≈ 0.2 at high deformation Z-score cutoffs). 8. Important nuance: unnormalized comparisons can look deceptively good because mutant deformation correlates with wild-type duplicate fluctuations. After normalization, AF3 appears biased toward wild-type-like ensembles and struggles specifically where mutations cause large structural responses. 9. Distance dependence is weaker than magnitude dependence: the correlation between predicted and experimental normalized deformation decreases only modestly with distance (from ~0.55 at r = 0 to ~0.35 far away), while the dominant failure mode is large mutation-induced deformation. 10. Physical interpretability: a single feature—normalized local change in relative solvent accessibility (^ΔrSASA) computed from experimental structures—correlates strongly with deformation (ρ ≈ 0.6) and, unlike AlphaFold3, the correlation does not degrade for strongly perturbative mutations; however, it currently cannot be used directly for prediction because it requires the mutant structure. 💻Code: github.com/lzyttxs/ 📜Paper: arxiv.org/abs/2609.24842 #ProteinStructure #Mutations #AlphaFold3 #ComputationalBiology #StructuralBiology #Bioinformatics #PDB #SASA #ProteinEngineering

MolExplain: An Interactive Tool for Explainable Molecular Property Prediction 1 MolExplain is an interactive web tool that pairs molecular property prediction with substructure-level visual explanations, so users see not only a probability score but also which regions of the molecule push the prediction for or against the property. 2 The core pipeline: SMILES → 2048-bit Morgan fingerprint (radius 2, RDKit) → XGBoost classifier → SHAP TreeExplainer per-bit attributions → RDKit bitInfo maps bits back to atom-centered environments → scores are smoothed into a continuous heatmap overlay on the 2D structure. 3 The demonstration task is cyclic peptide membrane permeability (from CycPeptMPDB). Labels are defined by PAMPA thresholding: permeable if PAMPA ≥ −6, otherwise non-permeable, framing permeability as a binary classification problem. 4 Model performance reported for the held-out test set: 79.8% accuracy and 0.876 ROC-AUC, emphasizing a practical tradeoff—using a strong, efficient fingerprint + tree model that supports faithful, deterministic SHAP explanations rather than maximizing accuracy with less transparent deep models. 5 The key design contribution is the “predict + interpret + modify + resubmit” loop: users can draw/edit molecules in a Ketcher sketcher or paste SMILES, run a prediction on demand, then iteratively redesign the structure guided by the heatmap attribution. 6 In a use-case walkthrough, MolExplain highlights unmethylated backbone nitrogens as strong anti-permeability contributors (blue). Iteratively N-methylating these sites shifts the prediction from 42% to 93.4% permeable, and the previously blue regions turn red—mirroring the known chemistry that backbone N-methylation can improve passive diffusion by reducing H-bond donors. 7 The paper is explicit about attribution limitations from fingerprint overlap: atoms adjacent to a true driver can inherit similar SHAP signal because Morgan environments overlap (e.g., a methyl group near an NH may incorrectly appear strongly anti-permeability). The smoothed regional heatmap helps, but interpretation still benefits from domain knowledge. 8 The authors also note an informal (not validated) generalization observation: applying the system to a small molecule outside the cyclic-peptide training distribution yields attribution patterns consistent with general permeability intuition (polar groups blue, aromatic ring red), suggesting fingerprints may carry transferable substructural signal. 9 System/engineering details: React (Vite) frontend + FastAPI backend; heatmaps rendered server-side as SVG via RDKit SimilarityMaps; SMILES validated with Pydantic; and a model-registry pattern allows adding new properties by dropping in a new saved model and config entry without changing the UI or API. 📜Paper: arxiv.org/abs/2609.25355 #ExplainableAI #XAI #Cheminformatics #DrugDiscovery #MolecularML #Visualization #SHAP #RDKit #XGBoost #CyclicPeptides

Towards Accurate Prediction of Mutation-Induced Changes in Protein Structure 1. The paper builds a quantitative framework to measure how single amino-acid mutations deform protein structures, using matched wild-type and mutant X-ray crystal structures from the PDB and explicitly controlling for “native” structural variability. 2. Key dataset contribution: 6,967 wild-type–mutant sequence pairs, leveraging 27,200 wild-type and 16,176 mutant X-ray structures, with each wild-type sequence required to have at least five experimental “duplicate” structures to estimate intrinsic fluctuations. 3. Core metric: for each residue i, local deformation Di is defined from RMS changes in Cα–Cα distances to its local Voronoi neighbors (rather than global RMSD), reducing sensitivity to rigid-body motion and focusing on local rearrangements around mutations. 4. Innovation that removes thermal/experimental noise: normalized mutation-induced deformation eDi = (average deformation between wild-type duplicates vs mutant structures) divided by (average deformation among wild-type duplicates). This isolates mutation-specific effects from baseline structural variability. 5. Main structural finding: mutation-induced changes are strongly localized. Averaged across proteins, deformation peaks at the mutation site and decays rapidly with distance, approaching a near-baseline plateau for spatial distances r > ~12–15 Å (and similarly within ~±10 residues along sequence). 6. Distributional result: eDi at mutation sites is roughly exponential—many mutations look “neutral” structurally (eDi ≈ 1), while a smaller fraction induce substantially larger local deformations. 7. AlphaFold3 evaluation: predicted mutant structures were generated (AF3 v3.0.1; up to 20 seeds per mutant, selecting top-ranked per seed). Correlation between predicted and experimental normalized deformation at the mutation site is moderate overall (Pearson ρ ≈ 0.55) but collapses for strongly perturbative mutations (down to ρ ≈ 0.2 at high deformation Z-score cutoffs). 8. Important nuance: unnormalized comparisons can look deceptively good because mutant deformation correlates with wild-type duplicate fluctuations. After normalization, AF3 appears biased toward wild-type-like ensembles and struggles specifically where mutations cause large structural responses. 9. Distance dependence is weaker than magnitude dependence: the correlation between predicted and experimental normalized deformation decreases only modestly with distance (from ~0.55 at r = 0 to ~0.35 far away), while the dominant failure mode is large mutation-induced deformation. 10. Physical interpretability: a single feature—normalized local change in relative solvent accessibility (^ΔrSASA) computed from experimental structures—correlates strongly with deformation (ρ ≈ 0.6) and, unlike AlphaFold3, the correlation does not degrade for strongly perturbative mutations; however, it currently cannot be used directly for prediction because it requires the mutant structure. 💻Code: github.com/lzyttxs/ 📜Paper: arxiv.org/abs/2609.24842 #ProteinStructure #Mutations #AlphaFold3 #ComputationalBiology #StructuralBiology #Bioinformatics #PDB #SASA #ProteinEngineering

ReverseScreen.ai: Pharmacophore-Guided Reverse Screening Across the Growing Co-Complex Proteome 3次元ファーマコフォアでPDBの共結晶リガンドを索引化、類似リガンド結合構造を検索 リバーススクリーニングの前段用 reversescreen.ai biorxiv.org/content/10.64898…