Turning scattered evidence into discovery decisions for life sciences

TL;DR: OpenAI’s life-science agent inside Codex autonomously turns scattered internal and public data into a ranked list of asthma targets (IL-33, TSLP, IL-1RA1) by spinning up sub-agents for genetics, translational biology and regulatory context, then merging their outputs into an evidence-backed decision. ## Turning scattered evidence into discovery decisions for life sciences ### The challenge: from raw inputs to ranked targets Scientists traditionally juggle wet-lab notebooks, external databases and literature, then manually knit the pieces into a go/no-go call. OpenAI’s new life-science model, code-named “GPT-Rosalind,” wraps this process into a single, reproducible Codex workflow. The demo task: compare and prioritise three validated asthma targets—IL-33, TSLP and IL-1RA1—using both an internal evidence package and live public data. ### Starting point: an internal evidence bundle The user drops a local folder into Codex. It contains: - Internal cellular assays - Biomarker strategy slides - Developability and safety read-outs - Target product profile (TPP) draft The model ingests the folder and returns a concise top-line verdict: “Rank order: 1) IL-33, 2) TSLP, 3) IL-1RA1, driven by superior internal potency and biomarker linkage.” Inline citations point to the exact Excel rows and PDF pages that support each statement. ### First expansion: calling the life-science plugin Codex flashes a prompt: “Human genetics or target-disease evidence can be extended. Run?” One click launches the life-science research plugin. The model has been fine-tuned to know which skill to invoke—GWAS catalog, GTeX, UK-Biobank, DisGeNET, OpenTargets—and how to weight their outputs. ### Parallel sub-agents to keep standards unbiased Rather than merge data streams prematurely, Codex spawns six specialist sub-agents: 1. Pascal – human genetics 2. Marie – transcriptomics & cell-type specificity 3. Rosalind – pathway & network topology 4. Darwin – animal-model phenotypes 5. Frida – regulatory precedent & safety flags 6. Watson – competitive landscape & patent expiry Each agent works in isolation, adheres to pre-set scoring rubrics, and deposits a JSON summary in a shared evidence bus. #### Example: Pascal’s genetics deep-dive Pascal is instructed: “Collect genome-wide significant SNPs (p < 5 × 10⁻⁸) for asthma, eQTL colocalisation probability > 0.8, and assign each to IL-33, TSLP or IL-1RA1.” It returns: - IL-33: 3 independent loci, 2 with credible set variants in enhancer regions; Mendelian randomisation OR = 1.42 (95 % CI 1.19–1.68). - TSLP: 1 locus, eQTL in airway epithelium, MR OR = 1.18. - IL-1RA1: no genome-wide significant SNPs; eQTL in whole blood but not lung. #### Other agents at a glance Marie finds IL-33 most selectively up-regulated in airway epithelium during viral exacerbations (log₂FC 2.1, FDR 3 × 10⁻⁵). Darwin reports Il33-knockout mice show 60 % reduction in house-dust-mite-driven airway hyper-responsiveness, versus 25 % for TSLP knockouts; Il1ra1⁻/⁻ lethal inflammatory phenotype raises safety flags. Frida surfaces an FDA “clinical hold” memo for an earlier IL-1RA1 agonist due to neutropenia. Waton notes IL-33 monoclonal antibody patents expire 2038, TSLP 2036, IL-1RA1 small-molecule cluster 2031. ### Synthesis: merging six orthogonal evidence layers When the last agent marks its task “COMPLETE,” the life-science model runs a Bayesian evidence-integration script. Each layer is weighted by the user-configurable “trust slider” (default: human genetics 30 %, translational data 25 %, safety 20 %, IP 15 %, competition 10 %). The posterior score confirms the original ranking: IL-33 (0.87) > TSLP (0.71) > IL-1RA1 (0.34). ### Deliverables back to the scientist A one-page memo is auto-generated: - Ranked target list with composite scores - Key drivers sentence: “IL-33 leads on genetic linkage, epithelial specificity and robust KO phenotype.” - Red-flag box: “IL-1RA1 dropped because of neutropenia signal and weak genetic support.” - Appendices: links to full agent logs, database versions, and reproducibility notebooks. ### Reproducibility & audit Every query, API call and prompt is version-stamped. A second scientist can re-run the identical workflow and obtain the same ranking, or swap in a new weighting scheme to test sensitivity. ### Road-map: deeper biological reasoning OpenAI states the life-science model is “learning to think longer and biologically smarter.” Next steps include protein-structure-aware scoring, automated CRISPR guide design, and integration with robotic lab automation so the ranked target list can flow directly into high-throughput validation. Source: https://www.youtube.com/watch?v=a-YJ6h7EJv8