GaiaLab turns a gene panel and disease context into an evidence-traceable research hypothesis—with the mechanisms, contradictions, sources, and next experiment visible for inspection.
Every output is hypothesis-generating and must be independently evaluated by qualified experts before any clinical, therapeutic, or diagnostic use.
The Global Ledger of Biological Evidence
From a gene panel to a cited, testable drug-repurposing hypothesis — in about a minute.
Every hypothesis is ranked by mechanism and traced to its evidence across 76+ live databases — PubMed literature and ClinicalTrials.gov trials — then GaiaLab shows you the contradictions and the next experiment, so you can challenge it. Research use only; every prediction is timestamped and pre-registered.
We publish our accuracy openly — including where we fall short. Confirmed exact-name forward predictions so far: 0 — and we say so, in public, on the live ledger.
GaiaLab identifies gene pairs that independently dysregulate the same disease pathway via distinct molecular mechanisms — with no direct protein-protein interaction. Like convergent evolution in nature (sharks and dolphins arriving at the same body plan), these pairs suggest the pathway is attacked from two independent angles, making pathway-level therapeutics more powerful than single-gene targeting.
Olaparib
→
BRCA1/2 · PARP inhibition
✦ FDA-approved · PARP inhibitor · concordance tracked
Midostaurin
→
FLT3-ITD AML · FLT3 inhibition
✦ FDA-approved · concordance tracked
Osimertinib
→
EGFR-mutant NSCLC · 3rd-gen TKI
✦ FDA-approved · concordance tracked
Ivosidenib
→
IDH1-mutant AML · IDH1 inhibition
✦ FDA-approved · concordance tracked
Nivolumab
→
Pan-cancer · PD-1 checkpoint
✦ FDA-approved · PD-1 checkpoint · concordance tracked
GaiaLab is not a report generator. It is a persistent translational intelligence layer where claims carry provenance (PMID, confidence, contradiction score), drug candidates transition through a lifecycle (proposed → grounded → validated → deprecated), and disease boards accumulate evidence across every analysis your team runs — with exponential decay for stale signals and contradiction alerts when new data conflicts with prior conclusions.
Six-factor weighted score (target overlap, clinical evidence, mechanism alignment, pathway relevance, safety profile, disease context) ranks FDA-approved and investigational agents into Tier I–III. AlphaFold pLDDT provides an orthogonal structural druggability signal. CIViC and OncoKB evidence levels calibrate confidence assignments. Scores are computational estimates. Two benchmarks: Temporal-holdout calibration — AUROC 0.90 on 22 known drug approvals held out by approval year, 22/22 temporal recalls, 8/8 negative controls passed, mean rank 3 — not a prospective result — methodology →. Retrospective benchmark (deduplicated, sealed-lockbox-excluded methodology) — see live value and CI at /calibration vs 0.50 random baseline — full validation data →. These measure different things; neither is a clinical predictor.
Identifies gene pairs that independently dysregulate the same disease pathway via distinct molecular mechanisms — with no direct protein-protein interaction between them. Inspired by convergent evolution in biology (sharks and dolphins independently arriving at the same body plan). When two panel genes converge on the same pathway without interacting, pathway-level therapeutic targeting is stronger than targeting either gene alone. Detected pairs are flagged in gene cards, generate novel hypotheses (novelty 0.80), and boost drug scores for agents targeting the shared pathway. ⟳ See gene cards after analysis.
Each candidate is timestamped at output time and matched against ClinicalTrials.gov. Matches where the prediction timestamp precedes trial registration date are classified as prospective; all others are retrospective. Prospective concordance indicates the system surfaced a hypothesis before investigators registered a corresponding trial — it does not constitute efficacy evidence. Completed trial status does not imply positive outcome. Full ledger: /validation.
Each gene is queried against CIViC (community-curated clinical variant evidence, Levels A–E; Washington University) and OncoKB (FDA-recognised precision oncology knowledge base; Memorial Sloan Kettering). Returns: drug associations per variant, AMP/ACMG tier, oncogene/TSG classification, and Level 1–R2 biomarker designations. CIViC: no token required. OncoKB: institutional token recommended.
Checkpoint and resistance genes are mapped to immunotherapy escape mechanisms via pathway enrichment, protein interaction topology, and cross-source agreement. Mechanistic assignments are derived from structured database outputs, not generative text synthesis. Assignments flagged as low-agreement carry an explicit uncertainty annotation.
Force-directed PPI graph with temporal overlay. Displays hub centrality, curated and computationally predicted edges, and topological context aggregated across STRING, BioGRID, and related interaction databases. Predicted edges are visually distinguished from experimentally validated interactions.
Each output claim is linked to one or more PubMed citations and assigned a polarity classification: supporting, contradicting, or mixed. Claims without citation support are flagged, not suppressed. Full PMID traceability is preserved in exported evidence packages.
Each run produces a tamper-evident snapshot encoding gene inputs, database versions, model configuration, scoring parameters, and complete outputs. Snapshots can be diffed against prior runs or replayed independently. Intended for methods-section documentation and internal audit trails.
Evidence Depth Score, Contention Index, and grounded citation ratio gate every output. Claims below configurable thresholds are flagged with explicit rationale. Nothing is silently discarded — suppressed claims are logged and accessible in the full evidence export.
Each conclusion is cross-validated across 16 independent data channels — genomics, protein structure, pathway enrichment, literature, drug bioactivity, clinical trials, disease association, interaction networks, expression, safety, and others — each channel aggregating one or more of the 76+ integrated databases. Agreement across channels elevates confidence score; divergence triggers a contradiction flag and downgrades the claim. No single source is treated as determinative.
Outputs export as JSON evidence packages or formatted briefs. Each export includes scoring context, PMID citations, contradiction annotations, and complete model configuration metadata. Format is designed for direct insertion into methods sections or internal research reports.
Prediction outcomes are periodically verified against ClinicalTrials.gov status updates. Hypothesis outputs are cross-checked against new PubMed entries. Calibration drift is detected and a recalibration multiplier is applied at each server cycle. Confidence score distributions are published at /validation.
Kaplan-Meier OS curves stratified by mutation status across 15 TCGA cohorts (BRCA, LUAD, GBM, PAAD, and 11 additional). Returns log-rank p-value, hazard ratio, and median OS via the cBioPortal public API. Survival data is observational; no causal inference is implied. No institutional subscription required.
Users can submit confirmed, refuted, partial, or inconclusive outcomes from the results page. Submissions are aggregated into per-decile calibration curves and applied as recalibration multipliers at the next server cycle. All submitted outcomes are treated as self-reported and are not independently verified.
Each analysis is compared against prior runs in the knowledge graph for the same disease context. A contradiction alert is raised when a therapeutic score shifts more than 20 points relative to prior runs for the same gene–drug pair. Weekly digests list the most materially changed conclusions. Alerts indicate score drift, not independent evidence of a clinical finding.
Each gene is queried against CIViC (Clinical Interpretation of Variants in Cancer; Washington University in St. Louis). Returns evidence levels A (validated association) through E (inferential), AMP/ACMG tier, and drug associations per variant. CIViC is community-curated; evidence quality varies by entry and requires independent verification.
Gene panels are enriched against MSigDB Hallmark (50 gene sets), KEGG 2021 Human, and WikiPathways via the Enrichr API. Statistical significance is assessed with Benjamini-Hochberg FDR correction. Results indicate overrepresentation in curated gene sets — not direct measurement of pathway activity in a specific tumour.
When an API token is configured, each gene is queried against OncoKB (Memorial Sloan Kettering; FDA-recognised). Returns Level 1 biomarkers (FDA-approved companion diagnostics), Level 2 (standard of care), Level 3B (investigational), and Level R1/R2 (resistance markers), plus oncogene vs. tumour suppressor classification. Requires institutional token for full access.
GaiaLab generates ranked repurposing candidates from a gene list in under 60 seconds, drawing from 76+ biological databases (CIViC, OncoKB, DGIdb, DrugCentral, OpenAlex, PharmGKB, MSigDB Hallmark, AlphaFold, DepMap, TCGA, ClinGen, COSMIC, JASPAR, STRING, gnomAD, OT Genetics, and 38 others). Each candidate is scored across six evidence dimensions by an eight-stage analysis pipeline (Planner, Retriever, Validator, Contradiction, BioMechanism, Translational, EnvironmentContext, and Narrator agents), cross-referenced against ClinicalTrials.gov; a six-role structured AI debate is available as a Pro layer. Performance data (AUROC, calibration curves, concordance breakdown) is published at /validation. All outputs are computational research hypotheses. Independent experimental validation is required before any therapeutic or clinical application.
Houston, TX · Disease areas: GBM, AML, Alzheimer's, breast cancer, NSCLC, pancreatic cancer, and others · Research use only · partnerships@gailabai.com
Developed in Houston, TX. Designed for translational research teams that need rapid hypothesis generation across gene panels without institutional informatics infrastructure. Primary use cases: target prioritisation, drug repurposing triage, mechanistic hypothesis scoping before wet-lab investment.
No login or subscription required for standard analyses. API access and team workspaces available on paid plans. All analyses return the same evidence — access tier affects export formats and rate limits, not scoring or data sources.
GaiaLab is model-agnostic: an MCP server with multi-provider failover (DeepSeek, OpenAI, Google, Anthropic) and no special access to any of them — the difference is what it does with them, not which model it runs. It is deliberately narrow and accountable: it timestamps every drug–disease prediction and scores it against real ClinicalTrials.gov outcomes independently of the model that generated it, publishes its accuracy openly with naive baselines (not marketing numbers), and goes deep on high-burden, neglected diseases such as Lassa fever. Verification is independent of the model, and our first prospective test is now publicly pre-registered (DOI 10.5281/zenodo.21420447, readout 2028) — not a model checking its own work. We are not yet prospectively validated, and we say so.
Download a complete audit snapshot containing evidence packages, scoring context, data sources, and model configuration metadata.
Includes reproducible gene inputs, data source versions, and full model configuration details.
| KRAS | BRAF | NRAS | BRCA1 | EGFR | PALB2 | BRCA2 | ALK | |
|---|---|---|---|---|---|---|---|---|
| 2201 | 2197 | 2191 | ||||||
| 354 | 22 | 344 | 344 | |||||
| 346 | 336 | |||||||
A gene panel in, a defensible research hypothesis out — with the evidence and the caveats attached.