
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.





















































