
ALT Editorial BioAI Frontier cover for the bioRxiv preprint of AbPoolRank. Typography reads MIT × Sanofi; 15 Structures; From 5 Models; Outperform 100; From Any Single Model; Antibody–Antigen Structure Selection; 58.0% vs. 51.5% Success@1; and AbPoolRank · bioRxiv Preprint, over a first-party navy abstract illustration of stylized antibody-like shapes, with a few highlighted complexes in the foreground and a denser, dimmer field behind them. This is an original editorial cover. The illustration is an abstract schematic, not an experimentally determined structure and not an official journal figure.

ALT Figure 1 from Varun Ullanat, Abhinav Gupta, Ruijiang Li, Pep Amengual-Rigo, Norbert Furtmann, Bryan Bryson and Bonnie Berger, “Breadth over Depth: Predictor-Agnostic Consensus for Improved Antibody–Antigen Structure Prediction,” bioRxiv 10.64898/2026.11.17.767212 v1 (posted 19 November 2026). Official v1 PDF extract at 400 dpi from page 9, including the complete on-page caption; artwork not redrawn; bioRxiv running header cropped. Source: https://www.biorxiv.org/content/10.64898/2026.11.17.767212v1 . Licensed under CC BY 4.0: https://creativecommons.org/licenses/by/4.0/ . Panel A is the method schematic: five co-folders contribute three structures each (15 total), ranked by AbPoolRank. Panels B–D report Success@1 versus sampling depth, CAPRI class at the 5×3=15-structure point versus each predictor sampled 100 times, and paired differences; the printed right-margin medium-or-better labels include 58% for AbPoolRank and 51% for OpenDDE (integers as shown on the figure).

ALT Figure 2 from Varun Ullanat, Abhinav Gupta, Ruijiang Li, Pep Amengual-Rigo, Norbert Furtmann, Bryan Bryson and Bonnie Berger, “Breadth over Depth: Predictor-Agnostic Consensus for Improved Antibody–Antigen Structure Prediction,” bioRxiv 10.64898/2026.11.17.767212 v1 (posted 19 November 2026). Official v1 PDF extract at 400 dpi from page 10, including the complete on-page caption; artwork not redrawn; bioRxiv running header cropped. Source: https://www.biorxiv.org/content/10.64898/2026.11.17.767212v1 . Licensed under CC BY 4.0: https://creativecommons.org/licenses/by/4.0/ . Success@1 of ranking methods on 15-structure mixed pools versus 100 structures from one co-folder. AbPoolRank on the mixed pool is labeled 58.0. A dashed line marks the strongest deep-sampling result, OpenDDE ranked by native confidence (51.5%). Stars denote oracle performance for the mixed pool and individual co-folders—an upper bound on the candidate set, not a selector result.

ALT Editorial BioAI Frontier cover for the Nature Biotechnology report on RhoAgent. Typography reads Nature Biotechnology; RNA Design Agent; Learns from Experiments; Up to 6× Cleavage Efficiency; Hovlinc ribozyme vs. expert-truncated construct; and RhoAgent, over a first-party navy abstract illustration of a stylized RNA molecule. This is an original editorial cover. The RNA graphic is an abstract schematic, not an experimentally determined structure and not an official journal figure.

ALT Official Nature Biotechnology Figure 4 (Huang et al., 2026, doi:10.1038/s41587-026-03362-3): iterative RhoAgent design and wet-lab optimization of the structurally uncharacterized HOV/hovlinc ribozyme versus the 83-nt human-expert-truncated construct (HOV expert), not wild type. Round 1–2 cleaved-fraction plots include a dashed “5× HOV expert” reference line; that line is not 6×. The abstract reports cleavage efficiency of up to sixfold relative to the expert-truncated construct. Official publisher figure from https://www.nature.com/articles/s41587-026-03362-3. Springer Nature or its licensor holds exclusive rights; not a Creative Commons licensed image.

ALT Official Nature Biotechnology Figure 1 (Huang et al., 2026, doi:10.1038/s41587-026-03362-3): overview of RhoAgent, an experimental-feedback RNA design agent, and of RDiff, its diffusion-based generative tool. In the workflow schematic, a design-task panel is labeled “Six-fold improvement” versus a human expert; that label is a schematic annotation, not a quantitative results plot. Official publisher figure from https://www.nature.com/articles/s41587-026-03362-3. Springer Nature or its licensor holds exclusive rights; not a Creative Commons licensed image.

ALT Editorial BioAI Frontier cover for the Science report on dEVA, a de novo bi-zinc enzyme-design method. Typography reads Science; De Novo Enzyme Design; No Natural Template.; No Predefined Catalytic Motif.; Up to 3 × 10¹³ Rate Enhancement; Relative to the uncatalyzed reaction; and dEVA · Bi-Zinc Metalloenzyme, over a first-party navy abstract illustration of a stylized bi-zinc pocket. This is an original generated editorial cover. It is not an official figure from the paper and is not a true protein structure or experimental rendering.

ALT Figure 1 from the bioRxiv preprint: overview of dEVA, a multi-objective evolutionary protein-design algorithm, including defined objectives, a scaffold library, the Pareto-front knee point, and iterative refinement (NSGA-II). Official bioRxiv v2 PDF extract at 400 dpi; artwork not redrawn; bioRxiv running header cropped; not a figure from the Science version of record. Figures 1 and 4 reproduced from Gina El Nesr, Simon L. Dürr, Irimpan I. Mathews, Qi Wen, Kewei Zhao, Ritimukta Sarangi, Ursula Rothlisberger, Fanny Sunden and Po-Ssu Huang, “Zero-shot design of a de novo metalloenzyme,” bioRxiv 10.64898/2026.04.23.720277 (this version posted 18 June 2026). Source: https://www.biorxiv.org/content/10.64898/2026.04.23.720277v2 . Licensed under CC BY 4.0: https://creativecommons.org/licenses/by/4.0/ .

ALT Figure 4 from the bioRxiv preprint: zero-shot design of a de novo binuclear zinc metalloenzyme (desB), including a hydrolysis scheme, activity screens, the bi-zinc design model, alanine-mutant tests, EXAFS, kinetics on phosphomonoester and phosphodiester substrates, and a comparison with natural phosphomonoesterases. Official bioRxiv v2 PDF extract at 400 dpi; artwork not redrawn; bioRxiv running header cropped; the official caption continues on the following PDF page and was not stitched onto this extract; not a figure from the Science version of record. Figures 1 and 4 reproduced from Gina El Nesr, Simon L. Dürr, Irimpan I. Mathews, Qi Wen, Kewei Zhao, Ritimukta Sarangi, Ursula Rothlisberger, Fanny Sunden and Po-Ssu Huang, “Zero-shot design of a de novo metalloenzyme,” bioRxiv 10.64898/2026.04.23.720277 (this version posted 18 June 2026). Source: https://www.biorxiv.org/content/10.64898/2026.04.23.720277v2 . Licensed under CC BY 4.0: https://creativecommons.org/licenses/by/4.0/ .

ALT Editorial ranking graphic for the 2026 Virtual Cell Challenge final test, not an official Arc Institute image. Headline: “Still Closer to a Simple Baseline.” A large number, 0.4399, is labelled “Best official overall score.” A horizontal official scale runs from 0, “cell-context-mean baseline,” to 1, “split-half replicate,” with a marker at 0.4399. Caption: official scale, not a percentage; 1 is a landmark, not a maximum. Below, official winners-page overall scores: 1 BioInsight AI Lab, Illumina, CellularAI algorithm, 0.4399, $100,000; 2 FlowState, flowstate, 0.3348, $50,000; 3 X-Team, Xaira Therapeutics, X-Cell Transfer, 0.3263, $25,000. Bars are drawn on the same 0-to-1 official scale. Footer: winners-page names and scores; each prize is half cash and half NVIDIA Brev credits; source virtualcellchallenge.org; held-out cell contexts D, E and F. Artwork is an original editorial graphic from official 2026 final-test results, not a screenshot of the official site.

ALT Editorial BioAI Frontier cover for this gene-expression preprint. Typography reads AlphaGenome / Borzoi; Fine-Tuned.; Still Behind Elastic Net.; Predicting Gene Expression Across Individuals; and 5,120 Participants · Whole Blood · Preprint, over a first-party navy abstract background. This is an original generated editorial cover created for this post. It is not an official figure from the paper.

ALT Figure 3 from the paper: test-set Spearman correlations for pretrained and fine-tuned Borzoi and AlphaGenome versus elastic net across 150 genes. Figures 3, 1 and 5 reproduced from Xiaonan Liu, Yu-Jyun Huang, Junwei Lu, Jiecong Lin, Minoli Perera, Kari North, Anne E. Justice, Xihong Lin and Tamar Sofer, “Fine-tuned DNA sequence language models for predicting gene expression in admixed population: A comparison with variant-based methods,” bioRxiv 10.64898/2026.11.17.767377 (this version posted 18 November 2026). Source: https://www.biorxiv.org/content/10.64898/2026.11.17.767377v1 . Licensed under CC BY 4.0: https://creativecommons.org/licenses/by/4.0/ . Official PDF extracts; artwork not redrawn, running headers and page numbers removed. The caption for Figure 3 continues in the source PDF.

ALT Figure 1 from the paper: methods of predicting gene expression, comparing a variant-based elastic-net model with sequence-based models that take personal DNA sequences as input. Figures 3, 1 and 5 reproduced from Xiaonan Liu, Yu-Jyun Huang, Junwei Lu, Jiecong Lin, Minoli Perera, Kari North, Anne E. Justice, Xihong Lin and Tamar Sofer, “Fine-tuned DNA sequence language models for predicting gene expression in admixed population: A comparison with variant-based methods,” bioRxiv 10.64898/2026.11.17.767377 (this version posted 18 November 2026). Source: https://www.biorxiv.org/content/10.64898/2026.11.17.767377v1 . Licensed under CC BY 4.0: https://creativecommons.org/licenses/by/4.0/ . Official PDF extracts; artwork not redrawn, running headers and page numbers removed. The caption for Figure 3 continues in the source PDF.

ALT Figure 5 from the paper: cross-ancestry differences in performance of ancestry-specific models evaluated in European- and African-enriched groups. Figures 3, 1 and 5 reproduced from Xiaonan Liu, Yu-Jyun Huang, Junwei Lu, Jiecong Lin, Minoli Perera, Kari North, Anne E. Justice, Xihong Lin and Tamar Sofer, “Fine-tuned DNA sequence language models for predicting gene expression in admixed population: A comparison with variant-based methods,” bioRxiv 10.64898/2026.11.17.767377 (this version posted 18 November 2026). Source: https://www.biorxiv.org/content/10.64898/2026.11.17.767377v1 . Licensed under CC BY 4.0: https://creativecommons.org/licenses/by/4.0/ . Official PDF extracts; artwork not redrawn, running headers and page numbers removed. The caption for Figure 3 continues in the source PDF.

ALT Editorial BioAI Frontier cover for the tmol preprint. Typography reads Rosetta Scoring on GPUs; Up to 362× Scoring Throughput; Physics for AI Protein Design; tmol · Preprint; and Batched scoring + gradients vs single-threaded PyRosetta, over a first-party navy abstract background. This is an original generated editorial image, not a figure from the paper.

ALT Figure 1 from the paper: batched modeling and performance. The figure shows tmol’s PoseStack workflow, how pairwise interactions are assigned to GPU thread blocks, throughput versus protein size for score-plus-gradient and FastRelax (the 800-residue score-plus-gradient bracket is labeled 359×), and agreement with Rosetta energies. Author-source figure PDF rasterized without modification, not redrawn. Figures 1 and 2 reproduced without modification from Andrew Leaver-Fay, Kieran Didi, Jeff Flatten, Alex Ford, Guangfeng Zhou, Nathaniel Corley, Hahnbeom Park, Maxwell W. Kazman, Rafael I. Brent, Joseph Kleinhenz, Henry Solberg, Andrew M. Watkins, Brian Kuhlman, David Baker and Frank DiMaio, “tmol: a GPU-accelerated implementation of the Rosetta energy function for machine learning,” arXiv:2611.21857. Source: https://arxiv.org/abs/2611.21857 . Licensed under CC BY 4.0: https://creativecommons.org/licenses/by/4.0/ .

ALT Figure 2 from the paper: energetics and design. The figure reports energy-function terms, computational benchmarks versus PyRosetta, and tmol-guided protein–protein interface design results, including hydrogen-bond and water-bridge rewards. Author-source figure PDF rasterized without modification, not redrawn. Figures 1 and 2 reproduced without modification from Andrew Leaver-Fay, Kieran Didi, Jeff Flatten, Alex Ford, Guangfeng Zhou, Nathaniel Corley, Hahnbeom Park, Maxwell W. Kazman, Rafael I. Brent, Joseph Kleinhenz, Henry Solberg, Andrew M. Watkins, Brian Kuhlman, David Baker and Frank DiMaio, “tmol: a GPU-accelerated implementation of the Rosetta energy function for machine learning,” arXiv:2611.21857. Source: https://arxiv.org/abs/2611.21857 . Licensed under CC BY 4.0: https://creativecommons.org/licenses/by/4.0/ .

ALT Sidetrak binding to kinesin-1. Fluorescence polarization shows an apparent dissociation constant of 110 ± 13 nM for labeled Sidetrak binding to the truncated K560 construct. Structural predictions and co-immunoprecipitation experiments support binding at the TPR docking site; the interface-disrupting 8A mutant impairs binding.

ALT Sidetrak disrupts the MAP7–kinesin-1 interaction. In cell-lysate competition assays, Sidetrak reduces MAP7-associated kinesin-1 recovery. Cell images and quantification show reduced MAP7-dependent recruitment of kinesin-1 to microtubules with Sidetrak, but not with the binding-defective 8A control.

ALT Sidetrak expression alters mitochondrial distribution in cultured cells. Fluorescence images and quantification show mitochondria clustering near the nucleus in cells expressing GFP-Sidetrak, unlike the binding-defective 8A control. The main mitochondrial analysis includes 60 cells across three independent experiments. Additional panels show perinuclear accumulation of endogenous kinesin-1.

ALT Editorial BioAI Frontier cover for the FoldMapper preprint. Typography reads FoldMapper; Evolutionary Context. No Query Sequence Needed.; and Protein Inverse Folding, over a first-party navy abstract background with a stylized protein fold. This is an original editorial cover created for this post. It is not an official figure from the paper.

ALT Figure 2 from the paper, a two-panel recovery–diversity plot. (a) TEDMonoBench (TED and PDB). (b) PPIBindBench non-peptide and peptide binders. Best-of-100 sequence recovery versus mean pairwise diversity at T=0.1, 100 samples per structure, with the Pareto front marked. Official PDF extract with no content edits (rasterized from the author-submitted PDF; not a byte-identical raw source file). From Haelyn Kim and Juyong Lee, “FoldMapper: Protein Inverse Folding Model with Structure-Derived Evolutionary Profiles and a Multiscale Graph Encoder,” bioRxiv 10.64898/2026.11.15.766879, https://www.biorxiv.org/content/10.64898/2026.11.15.766879v1, licensed under CC BY-NC 4.0, https://creativecommons.org/licenses/by-nc/4.0/.

ALT Figure 3 from the paper. Structure self-consistency on RFdiffusion de novo backbones: success rate (scTM ≥ 0.8) and best/mean scTM for monomer, binder, and oligomer sets, plus representative designs. ProteinMPNN is excluded from this main figure because those backbones were screened with ProteinMPNN sequences. Official PDF extract with no content edits (rasterized from the author-submitted PDF; not a byte-identical raw source file). From Haelyn Kim and Juyong Lee, “FoldMapper: Protein Inverse Folding Model with Structure-Derived Evolutionary Profiles and a Multiscale Graph Encoder,” bioRxiv 10.64898/2026.11.15.766879, https://www.biorxiv.org/content/10.64898/2026.11.15.766879v1, licensed under CC BY-NC 4.0, https://creativecommons.org/licenses/by-nc/4.0/.

ALT Figure 1 from the paper, the FoldMapper architecture. A backbone enters at left; Foldseek retrieves structural homologues; their identities are aggregated into an SPSF profile; a multiscale graph encoder and autoregressive decoder output a designed sequence. Official PDF extract with no content edits (rasterized from the author-submitted PDF; not a byte-identical raw source file). From Haelyn Kim and Juyong Lee, “FoldMapper: Protein Inverse Folding Model with Structure-Derived Evolutionary Profiles and a Multiscale Graph Encoder,” bioRxiv 10.64898/2026.11.15.766879, https://www.biorxiv.org/content/10.64898/2026.11.15.766879v1, licensed under CC BY-NC 4.0, https://creativecommons.org/licenses/by-nc/4.0/.

ALT Editorial BioAI Frontier cover for the Anthropic × BigHat preprint “Reasoning in the loop.” Text reads Anthropic × BigHat; Claude Helps Turn Hypotheses into Experiments; Antibody Lead Optimization; and 7 Rounds · ~1,010 Variants · Preprint, over an abstract antibody-and-measurement illustration on a dark field. Claude is emphasized in yellow. This is an original editorial generated cover, not an official Anthropic or BigHat visual and not a figure from the paper.

ALT Figure 4 from the paper. Panel (a) shows a binding-assay curve for the VHH (nanobody) parent with a 1:1 kinetic fit. Panel (b) shows the same parent’s thermal assay: fluorescence and 266 nm scattered light versus temperature, with model fits. Scattering rises while fluorescence is still almost unchanged, illustrating aggregation beginning while the protein is still folded, so a reported melting or aggregation temperature can mislead. Unaltered official figure from Stanton, Solomon, Delaney, Montgomery, Humphries, Acker, Henrici, Eid, Abrams and Greenside, “Reasoning in the loop: antibody lead optimization by probing mechanism and measurement,” arXiv:2611.19812, https://arxiv.org/abs/2611.19812, licensed under CC BY 4.0, https://creativecommons.org/licenses/by/4.0/.

ALT Figure 5 from the paper. Averaged thermal-assay traces for parents and leads: fluorescence (top) and 266 nm scattered light (bottom) versus temperature for the VHH/nanobody (a) and the scFv (b). The VHH lead shows little scattering; the scFv lead remains in solution to a higher temperature than its parent. Additional panels test the physical model by changing protein concentration and heating rate against preregistered predictions. Unaltered official figure from Stanton, Solomon, Delaney, Montgomery, Humphries, Acker, Henrici, Eid, Abrams and Greenside, “Reasoning in the loop: antibody lead optimization by probing mechanism and measurement,” arXiv:2611.19812, https://arxiv.org/abs/2611.19812, licensed under CC BY 4.0, https://creativecommons.org/licenses/by/4.0/.

ALT Figure 6 from the paper. Final-round designs plotted against the revised specifications. (a) VHH/nanobody off-rate versus the aggregation metric p k_agg. (b) scFv off-rate versus the solution-collapse temperature T_collapse, the temperature at which the sample falls out of solution. Dashed lines mark constraints; the selected leads are indicated. Unaltered official figure from Stanton, Solomon, Delaney, Montgomery, Humphries, Acker, Henrici, Eid, Abrams and Greenside, “Reasoning in the loop: antibody lead optimization by probing mechanism and measurement,” arXiv:2611.19812, https://arxiv.org/abs/2611.19812, licensed under CC BY 4.0, https://creativecommons.org/licenses/by/4.0/.

ALT Editorial BioAI Frontier cover for Isomorphic Labs’ $5 billion Series C. Text reads ISOMORPHIC LABS, $5 BILLION, SERIES C, and Scaling AI Drug Discovery, over an abstract protein-and-ligand illustration on a dark field. This is an original editorial generated cover, not an official Isomorphic Labs or Alphabet visual.

ALT BioAI Frontier original cover for an external analysis of Claude Fable 5.2. Headline: Claude Fable 5.2. Hook: Outperforms Protein Language Models. Qualifier: In next-amino-acid prediction. Footer: External analysis · Tested open models. Abstract teal molecular ribbons on a deep navy field. Typography and background are an original BioAI Frontier cover; the background was first-party generated for this post and is not a figure from Xinming Tu and Leo Tianlai Chen's analysis or the LLM-bioseq-probe repository.

ALT Chart from Xinming Tu and Leo Tianlai Chen's 16 November 2026 analysis: next-amino-acid prediction accuracy of Claude models compared with open protein language models, including ProGen3, ProtGPT3 and ESM, on real protein sequences versus shuffled-sequence controls. Copyright (c) 2026 Xinming Tu. Available under the MIT License: permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, subject to the following condition: the above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. Full license: https://github.com/XinmingTu/LLM-bioseq-probe/blob/main/LICENSE

ALT Chart from Xinming Tu and Leo Tianlai Chen's 16 November 2026 analysis: next-amino-acid prediction results for 22 general-purpose language models, including Claude, GPT, Gemini and open-weight models, comparing accuracy on real sequences with shuffled-sequence controls. Copyright (c) 2026 Xinming Tu. Available under the MIT License: permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, subject to the following condition: the above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. Full license: https://github.com/XinmingTu/LLM-bioseq-probe/blob/main/LICENSE

ALT Chart from Xinming Tu and Leo Tianlai Chen's 16 November 2026 analysis: how each model's real-versus-shuffled next-amino-acid prediction advantage changes with protein-family size, a pattern also seen in dedicated protein language models. Copyright (c) 2026 Xinming Tu. Available under the MIT License: permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, subject to the following condition: the above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. Full license: https://github.com/XinmingTu/LLM-bioseq-probe/blob/main/LICENSE

ALT BioAI Frontier editorial cover for OpenScientist in Nature Medicine. Text highlights a six-blood-protein survival-risk score, c-index 0.826, tested on 176 unseen participants. This is one of five biomedical case studies; the hidden cohort included 17 deaths. Six conceptual beads connect to a score tile on a navy background. The illustration is decorative, not molecular or experimental data.

ALT Original Figure 1: OpenScientist investigation workflow. A scientist specifies a task and optionally provides data. The agent searches PubMed, reads papers, writes and runs code, and analyzes files using scientific skills. Findings, hypotheses, plots and transcripts enter a persistent knowledge state; optional user feedback guides later iterations. A final report is compiled. This is computational analysis, not an automated wet-lab experiment. Source: Roberts et al., Nature Medicine (2026), Fig. 1, CC BY 4.0; rendered from accepted manuscript.

ALT Original Figure 4: plasma proteomic survival modeling. Panel a shows demographics and plasma proteomics from 500 participants supplied with a survival-modeling task. Panel b compares concordance indices; panel c compares with biological-clock models. The dagger note identifies AI Agent (OwnSplit) as internal 50/50 train-test validation, not the competition holdout. These evaluation settings must not be conflated with the separate 176-person hidden-cohort rerun reported in the text. Original axes and manuscript markings retained. Source: Roberts et al., Nature Medicine (2026), Fig. 4, CC BY 4.0.

ALT Original Figure 3: Alzheimer's biomarker analysis. Panel a shows the SEABIRD proposal, prespecified analysis plan and data from 325 participants. Panel b compares human ROC curves on the left with OpenScientist curves on the right, retaining biomarker legends, AUCs and confidence intervals. Both identify pT217/T217 as the strongest PET-status discriminator. This is the final optimized run after prompt refinement and data curation. Original manuscript markings retained. Source: Roberts et al., Nature Medicine (2026), Fig. 3, CC BY 4.0.

ALT BioAI Frontier editorial cover with the headline Baker Lab × Green Lab and text Choose the C–H Site. Design the Enzyme. Supporting text states One Molecule. Two Target Sites. Estradiol: C6 or C16. Preprint, wet-lab validated. A stylized folded-protein pocket holding an amber substrate represents the design concept; this is an illustration, not a measured molecular structure.

ALT Original Figure 4: site-selective estradiol hydroxylation. A shows C6 hydroxylation to 6alpha-hydroxyestradiol or C16 hydroxylation to estriol. B shows 96 C6 designs and 67 C16 designs; initial yields were below 1% and about 1%, with 93% and >99% diastereomeric excess. C shows Q20N, W70K, I121V and V140Q substitutions increasing C16 yield to 37% and turnovers to 37 while retaining >99% d.e. D overlays the substrate-free, heme-bound dOHes-2-NKVQ crystal structure with the design, C-alpha RMSD 0.65 angstrom. These percentages for d.e. describe stereoselectivity, not yield. Source: Kalvet et al., bioRxiv (2026), Fig. 4, CC BY 4.0.

ALT Original Figure 1: design principles and workflow for de novo C-H hydroxylases. A-D show hydroxylation applications, oxygen/peroxide activation, the Glu-His dyad promoting heterolytic peroxide cleavage, and Compound I hydrogen abstraction and radical rebound. E-F show the minimal heme-based catalytic motif and target substrate geometry, RFdiffusion2/3 backbone generation, iterative LigandMPNN/Rosetta Relax sequence design, and Chai-1 or AlphaFold3 filtering. Original panels and labels retained. Source: Kalvet et al., bioRxiv (2026), Fig. 1, CC BY 4.0.

ALT Fig. 1. MDClaw uses an execution graph to constrain tool calls, keep auditable records, and support branching workflows. a, A coding agent uses the harness to plan and run MD for a human or AI scientist’s question. b, 100 command-line tools and 14 domain skills cover preparation, simulation and analysis. c, DAG rules reject invalid stage transitions, inputs and settings, and let results be audited. d, Skills guide planning; tools execute the plan and support branching, such as sequence variants, force-field changes and replicas.

ALT Fig. 2. MDDataBench task systems, prompts and artifact-based scoring. a, 31 test systems, shown from deposited structures, are held out from 128 development tasks. b, An example 2PQL prompt with required artifacts. c, Submissions are scored on preparation and MD checks against reference data, not on the agent’s written report. d, The same prompt is run under three conditions; the primary metric is pass^3, the share of systems that pass all three repeats.
























