Cached at:
04/21/26, 04:31 PM
# GPT-Rosalind speeds up follow-up experiments in life-science research
TL;DR: OpenAI’s GPT-Rosalind and the Life-Sciences Research Plugin translate a high-priority target (TSLP) into a concrete perturbation experiment, outputting exact reagents, concentrations, time-points and read-outs so wet-lab teams can run the plate the same afternoon.
## From target ranking to an executable protocol
After the target-prioritisation step ranked TSLP highest, the model is asked to “design a perturbation experiment with specific experimental parameters.”
Within seconds it returns a complete 96-well layout that includes:
- Cell line: NHBE (normal human bronchial epithelial), passage ≤ 3
- Stimulus: 10 ng mL⁻¹ recombinant TSLP, 100 µL per well
- Inhibitor arms: 0.1, 1, 10 µM of anti-TSLP nanobody (clone 3E12)
- Vehicle: 0.1 % DMSO, matched volume
- Time-points: 0, 2, 6, 24 h
- Read-outs: qPCR (IL-33, CCL26, OCSTAMP), secreted-protein ELISA (IL-5, IL-13), high-content imaging for goblet-cell hyperplasia
- Controls: scrambled siRNA, non-targeting nanobody, cell-free background, inter-plate calibrators
Each condition is replicated in triplicate; the plate map is exported as a JSON that can be dropped directly into robotic liquid-handlers (Hamilton, Tecan, Opentrons).
## Evidence layer: grounding every reagent in public data
The Life-Sciences Research Plugin queries > 40 open databases (NCBI Gene, UniProt, ChEMBL, ENCODE, DepMap, PubChem, ClinicalTrials.gov) to justify the choices:
- TSLP–IL-33 axis: 17 CRISPR-knockout and 4 RNA-seq studies show ≥ 2.1-fold increase in IL-33 transcript when TSLP is over-expressed.
- CCL26 (eotaxin-3): cited in 42 papers as the most responsive chemokine in human airway epithelia after TSLP stimulation.
- Nanobody 3E12: KD 3.2 nM, no off-target hits in a 228-kinase panel (data from the 2022 paper by Liu et al., Science Translational Medicine).
- NHBE cells: maintain physiological TSLPR expression up to passage 4 ( LungMAP consortium scRNA-seq, 2023 release).
The evidence snippets are appended to the protocol so reviewers can audit every number.
## Hypothesis generation beyond the obvious
With biosafety-level constraints relaxed, the model can propose orthogonal perturbations:
1. CRISPRa tiling of the TSLP distal enhancer (chr5: 134.2 Mb) to find the minimal enhancer that still gives a 3-fold luciferase signal.
2. co-culture with M2 macrophages to test epithelial–immune feedback; model suggests adding 1:5 ratio macrophages, 20 ng mL⁻¹ IL-4 to polarise.
3. PROTAC-mediated TSLPR degradation (dTAG-13 system) to compare functional versus genetic knock-down kinetics.
Each hypothesis is scored for novelty (PubMed overlap < 15 %), tractability (existing reagents available from Addgene or ChemBridge), and disease relevance (GWAS p-value < 5 × 10⁻⁸ for asthma).
## Optimising the protocol in silico before spending reagents
The plugin plugs into a kinetic model of NF-κB / STAT5 signalling parameterised with 312 published phospho-proteomics curves.
It simulates the expected IL-5 secretion under three inhibitor doses and predicts that 1 µM nanobody will give 72 % suppression at 6 h with an EC90 confidence interval of 0.7–1.4 µM—matching the wet-lab result within 8 %.
When the user specifies a budget cap of $500 per plate, the model prunes the 24 h time-point and reduces qPCR replicates to duplicates, still keeping statistical power > 0.8 (α = 0.05, expected effect size 1.5-fold).
## Closing the loop: data returns to the model
After the run, scientists upload raw Ct values and ELISA OD450.
GPT-Rosalind auto-normalises against the housekeeping gene RPL32, fits a 4-parameter logistic curve for the ELISA, and updates an internal Bayesian posterior over the TSLP → IL-5 edge weight.
The new posterior is stored in a secure workspace so the next round of prioritisation already “knows” how strong the TSLP blockade was, shortening the design cycle from weeks to hours.
## Current access and roadmap
The Life-Sciences Research Plugin is available in closed beta through ChatGPT Enterprise and the Codex API.
OpenAI plans to expand coverage to single-cell multi-omics, high-throughput CRISPR, and in-vivo dosing regimens by Q4 2024.
Source: https://www.youtube.com/watch?v=k9xZuTMCR1Q