AI-Driven Drug Repurposing Screen
User: BioML Research (startup), AI drug discovery Budget: $12,000 (batch of 24 experiments) Outcome: 3 hits identified, 1 advancing to further study
The Problem
BioML Research uses machine learning models to predict which existing drugs might work against new disease targets. Their models generate hundreds of candidates, but computational predictions need wet lab validation.
Traditional CROs wanted $50K+ minimum contracts and 8-week timelines just to screen 20 compounds. BioML needed faster, cheaper iteration.
The Approach
BioML built an automated pipeline:
ML Prediction → Spec Generator → Litmus API → Webhooks → Model RetrainingThis case study covers one batch: 24 compounds predicted to inhibit a kinase target implicated in inflammatory disease.
Experiment Generation
BioML's pipeline auto-generates experiment specifications:
def generate_litmus_spec(compound: Compound, target: Target) -> dict:
return {
"metadata": {
"submitter_type": "ai_agent",
"agent_identifier": "bioml-repurposing-v2.1",
},
"hypothesis": {
"statement": f"{compound.name} inhibits {target.name} with IC50 ≤ 10μM",
"null_hypothesis": f"{compound.name} does not inhibit {target.name}",
},
"protocol": {
"type": "standard_template",
"template_id": "kinase-inhibition-adp-glo-v1",
},
"communication_preferences": {
"webhook_url": "https://api.bioml.io/litmus/webhook",
"notification_events": ["claimed", "completed", "issue"],
}
}Submission completed in 45 seconds. 24 experiments, total estimated cost: $11,400.
Execution Timeline
| Day | Event |
|---|---|
| 0 | 24 experiments submitted |
| 1-2 | All experiments claimed by 4 different operators |
| 3 | First issue flagged: 1 compound insoluble |
| 8 | First results received (6 experiments) |
| 14 | All 24 experiments complete |
Results Summary
| Result | Count | Compounds |
|---|---|---|
| IC50 ≤ 10μM (hit) | 3 | BML-042, BML-089, BML-156 |
| IC50 10-100μM (weak) | 5 | BML-023, BML-067, BML-091, BML-112, BML-178 |
| IC50 > 100μM (inactive) | 14 | [remaining] |
| Inconclusive (solubility) | 2 | BML-033, BML-144 |
Hit rate: 12.5% (3/24)
For comparison, BioML's previous ML model version had a 4% hit rate. The improved model (v2.1) showed meaningful improvement.
Top Hit: BML-089
| Concentration (μM) | Activity (% control) |
|---|---|
| 0 (DMSO) | 100.0 ± 3.2 |
| 1 | 72.3 ± 5.2 |
| 3 | 41.2 ± 4.6 |
| 10 | 18.7 ± 3.1 |
| 30 | 8.3 ± 2.4 |
Calculated IC50: 2.3 ± 0.4 μM
Cost Analysis
| Item | Amount |
|---|---|
| 24 experiments @ ~$475 avg | $11,400 |
| Compounds (provided by BioML) | $2,100 |
| Shipping | $180 |
| Total | $13,680 |
| Cost per compound tested | $570 |
Compared to traditional CRO quote: $52,000 for same scope.
Savings: 74%
Timeline Comparison
| Metric | Litmus | Traditional CRO |
|---|---|---|
| Contracting | Instant (API) | 2-3 weeks |
| Execution | 14 days | 6-8 weeks |
| Results format | Structured JSON | PDF report |
| Integration | Automated webhook | Manual extraction |
Model Performance Over Time
| Batch | Compounds | Hits | Hit Rate | Model Version |
|---|---|---|---|---|
| 1 | 20 | 1 | 5.0% | v1.0 |
| 2 | 20 | 1 | 5.0% | v1.2 |
| 3 | 24 | 2 | 8.3% | v2.0 |
| 4 | 24 | 3 | 12.5% | v2.1 |
| 5 | 30 | 5 | 16.7% | v2.3 |
The feedback loop is working. Wet lab validation improves computational predictions.
Lessons Learned
- Structured output enables automation: JSON results parse directly into training pipelines
- Parallelization matters: 24 experiments across 4 operators completed faster than sequential
- Validation endpoint catches errors early: 2 experiments had spec issues caught before submission
- Delayed privacy works for IP protection: 12-month delay gives time for patent filing