OpenScientist: Supplementary Case Study Data

Supplementary data for "OpenScientist: evaluating an open agentic AI co-scientist to accelerate biomedical discovery." This dataset contains research logs, knowledge states, provenance files, and generated figures for seven case studies spanning neurodegenerative disease transcriptomics, multiple myeloma genomics, survival proteomics, drug side effect analysis, and more. Each case study includes the full agent iteration log, configuration, and final report. Input datasets are included for three case studies where data sharing is permitted: case study 3 (neurofibrillary tangle transcriptomics), case study 4 (multiple myeloma RNA-seq), and case study 7 (drug side effect frequencies from clinical trials). Version 2 (August 2026) adds supplementary material for case study 3 (neurofibrillary tangle transcriptomics): 60 additional independent runs of the same question and dataset, comparing three agent/model configurations — Claude Code with Claude Opus 4.8, and the omp harness with Kimi K3 and GLM 5.2 — each run 10 times online and 10 times in a fully air-gapped configuration with no network access and literature served from a local 40-million-article MEDLINE mirror. Files: ..._supplemental_online_30runs_data.tar.gz and ..._supplemental_airgapped_30runs_data.tar.gz each contain 30 runs with reports (Markdown, PDF, HTML), all generated figures, per-iteration transcripts, the exact prompt, and full version/commit provenance. ..._supplemental_60runs_metrics.csv gives per-run runtime, tool-call counts, literature searches, findings recorded and token usage for all 60 runs. Version 3 (September 2026) adds two further supplementary sets for case study 3 (neurofibrillary tangle transcriptomics). ..._supplemental_reproducibility_10runs_data.tar.gz contains 10 independent runs of the same question on Claude Opus 4.6, all run on 2026-03-08 with byte-identical inputs (the research question and the input .h5ad were SHA-256 verified identical across all 10) and a 10-iteration budget. Each run includes the report (Markdown, PDF, HTML), knowledge_state.json with structured findings and literature list, per-iteration transcripts, and the job UUID. These runs quantify run-to-run variability under fixed inputs. ..._supplemental_genomic_skill_comparison_data.tar.gz contains 30 runs — 15 using the original genomics Agent Skill and 15 using a revised version — with the two skill definitions included alongside the jobs. The revised skill adds explicit guidance on donor-level pseudobulk aggregation and mixed-effect modelling; the paired arms allow the effect of that guidance on the analysis the agent chooses to be assessed directly. Each run includes reports (Markdown, PDF, HTML), per-iteration provenance transcripts, and container logs. Version 4 (October 2026) adds the run sets generated for the second revision of the manuscript, renumbers the case-study files to match the revised manuscript (the suvorexant proteomics analysis is now Case Study 5; neurofibromatosis is Case Study 6, drug side effects Case Study 7, and the CSF Aβ42/40 prediction model Case Study 8), and adds RUNS_MANIFEST.csv, which maps every job in this record to its experiment, supplementary table and row label. Run IDs quoted in the supplementary tables are the first 8 hexadecimal characters of the job ID. New archives, all for Case Study 3 (neurofibrillary tangle transcriptomics; same research question and input file as the existing archives, OpenScientist commit fac6971, Claude Opus 4.6, run 2026-10-18 with a 10-iteration budget): ..._supplemental_iteration_controls_10runs_data.tar.gz (10 replicate runs on unmodified code; Experiment 4 and the control arm of the two ablations), ..._supplemental_knowledge_state_ablation_10runs_data.tar.gz (10 runs in which the knowledge-state summary and outline were withheld from the agent; Experiment 5; the patched orchestrator file is included as ..._knowledge_state_ablation_iteration.py.patched), and ..._supplemental_formula_removed_10runs_data.tar.gz (10 runs with the hypothesis-prioritization scoring rules removed from the agent skills and system prompt; OpenScientist pull request 310). Each run directory contains the report (Markdown, PDF, HTML), knowledge_state.json, claude_iterations.log and per-iteration provenance transcripts; the input .h5ad is not duplicated. Development runs behind the per-case-study run tables: Case_study_1_SEABIRD_supplemental_development_11runs_data.tar.gz (the 11 SEABIRD runs of Supplementary Table 2; input data not included), Case_study_2_Survival_Proteomics_supplemental_development_3runs_data.tar.gz (3 runs; input data not included) and Case_study_3_NFT_Transcriptomics_supplemental_exploratory_2runs_data.tar.gz (the two exploratory runs that preceded the benchmark run a66ae54c). Version added 2026-11-05: analysis scripts behind the manuscript figures and tables (Case_study_3_ksab_score.py, Case_study_3_exp4_plot_EDFig2.py, Case_study_3_EDFig1_gene_inclusion_matrix.py, Case_study_1_Fig3_roc_recompute.py) with README_analysis_scripts.md. The gene tokeniser used for Supplementary Tables 5-6 and Extended Data Fig. 2 was corrected to split hyphen-joined gene names.