Validation case study · CNS
We predicted a CNS blockbuster's clinical profile before any clinical data.
A blinded test on esketamine (Spravato®, J&J). We removed everything clinical from our knowledge graph, kept only molecular pharmacology, and asked the model to reconstruct the drug's real indications and side effects. It did.
The $2.9B problem
In CNS, the costly mistake is strategic, not molecular.
Most late-stage CNS failures don't come from bad molecules. They come from the wrong indication, the wrong patient population, or endpoints that can't capture the drug's signal. By the time that shows up in Phase II, hundreds of millions and several years are already gone.
<7%
of CNS candidates entering Phase I reach approval
16 yrs
average timeline per approved CNS drug
$2.9B
average cost per approved drug
40%
of Phase II failures are due to lack of efficacy
The validation protocol
A blinded, mechanism-only test against a known answer
Esketamine is one of the most exhaustively documented CNS drugs, which makes it an ideal positive control. To keep the test honest, the model is blinded to everything we already know about it.
Remove clinical knowledge
Every documented esketamine link, including indications, adverse effects and off-label uses, is deleted from the graph. The model starts with no access to its clinical reality.
Reintegrate molecules only
Esketamine is reconnected through its molecular targets alone: the NMDA receptor subunits (GRIN1, GRIN2A-D) and its CYP-metabolizing enzymes. Downstream pathway effects propagate from there.
Predict vs. documented reality
The model reconstructs esketamine's indications and side effects from mechanism alone, and we compare against the record built by regulators, clinicians and decades of trials.
Evaluated under a strict zero-leakage protocol: ~500,000 known relationships are hidden before training, and the model must recover them. The graph itself encodes ~8M biological relations across 9,000 drugs and 26,000 diseases, curated from 20+ pharmacological databases.
The results
From molecular data alone, the model reconstructed esketamine's clinical reality.
100%
of esketamine's documented indications recovered in the top 13 predicted, out of ~26,000 diseases.
#1
Treatment-resistant depression, esketamine's defining use, ranked first at the phenotype level.
18 / 20
predicted side effects clinically supported, including 12 newly surfaced beyond our graph.
0.85
AUPRC for indication prediction (0.84 macro), the best of every benchmarked CNS baseline.
Explore the predictions
What the model returned
Top-20 predicted diseases. 100% of esketamine's documented indications appear in the top 13 of ~26,000. Beyond depression, the model surfaces repositioning signals: some already in esketamine trials, some still research-stage.
- #1Anxiety disorderDocumented indication
- #2Bipolar disorderInvestigated in trials
- #3SchizophreniaPredicted signal
- #4Neurotic disorderResearch-stage signal
- #5Endogenous depressionDocumented indication
- #6Major affective disorderPredicted signal
- #7Dysthymic disorderInvestigated in trials
- #8Parkinson diseaseResearch-stage signal
- #9Substance abuse / dependencePredicted signal
- #10EpilepsyResearch-stage signal
- #11Unipolar depressionDocumented indication
- #12Manic bipolar affective disorderPredicted signal
- #13Major depressive disorderDocumented indication
- #14Hypertensive disorderPredicted signal
- #15Monogenic obesityPredicted signal
- #16ArrhythmiaPredicted signal
- #17Poor appetitePredicted signal
- #18Mood changesPredicted signal
- #19PainInvestigated in trials
- #20AnxietyPredicted signal
Highlighted = part of esketamine's documented clinical record, recovered from mechanism alone.
Ranking accuracy (AUPRC)
AUPRC asks: when ranking drug candidates for an indication, how well are the real treatments concentrated at the top? It runs 0 (wrong) to 1 (perfect). Baselines are external reference points from a peer-reviewed benchmark.
Scale 0–1 · higher is better · Theremia indication AUPRC 0.85 (0.84 macro).
From molecular pharmacology to clinical strategy, before a single patient.
What this means for your asset
Four capabilities, in weeks, before Phase I
The esketamine test shows the model captures real biology even in complex CNS cases. For a first-in-class compound, that delivers what traditionally takes years of clinical data:
Indication & phenotype prioritization
A ranked, mechanistically grounded view of which diseases, and which patient subgroups within them, your asset is most likely to address. It sharpens go/no-go before Phase I.
Side-effect anticipation
An early read on likely adverse events, informing endpoint selection and inclusion/exclusion criteria before any patient is exposed.
Biomarker identification
A panel of candidate biomarkers linking your mechanism to predicted indications, giving early pharmacodynamic readouts to confirm target engagement and monitor efficacy.
Competitive differentiation
A benchmark against standard-of-care and comparators, showing where your mechanism is genuinely differentiated and where combination or sequencing opportunities exist.
Scientific boundaries
This validation covers therapeutic indications and adverse-event profiles. A complete clinical profile also needs modeling of pharmacokinetic and pharmacodynamic properties and sub-population responses, which we do not claim here.
Want this run on your molecule?
We can apply the same blinded, mechanism-first analysis to your preclinical asset and review it with you in a 30-minute readout.
References
- 1. Dowden, H., & Munro, J. (2019). Trends in clinical success rates and therapeutic focus. Nature Reviews Drug Discovery, 18(7), 495–496.
- 2. Sertkaya, A., Beleche, T., Jessup, A., & Sommers, B. D. Costs of Drug Development and R&D Intensity in the US, 2000–2018. JAMA.
- 3. DiMasi, J. A., Grabowski, H. G., & Hansen, R. W. (2016). Innovation in the pharmaceutical industry: New estimates of R&D costs. Journal of Health Economics, 47, 20–33.
- 4. Huang, K., Chandak, P., Wang, Q. et al. (2024). A foundation model for clinician-centered drug repurposing. Nature Medicine, 30, 3601–3613. (baseline benchmark reference)