# CRITICAL EXAMINATION: AI-Enabled Drug Discovery Brief

## 1. STRONGEST CLAIM (AND WHY IT'S LIKELY OVERSTATED)

**The "100x cost reduction" claim ($400M → $3M) is the most aggressive assertion and requires immediate challenge.**

### Operational Definition Problems:
- **What exactly constitutes "preclinical costs"?** This term is doing enormous work here. Does it include:
- Failed candidates along the way?
- Platform development/infrastructure costs (amortized or excluded)?
- Personnel costs for the AI/ML teams?
- Licensing fees for training data?
- The $400M baseline—is this an industry average, median, or cherry-picked comparator?

- **What counts as "discovery timeline"?** The 4.5 years → 18 months comparison:
- Does this start from target identification or from program initiation?
- Is the comparator for the *same indication* (IPF) or a general industry average?
- IPF has known biology and validated targets—this isn't a novel target class.

### Why This Matters:
The $3M figure almost certainly excludes platform R&D costs that Insilico has spent hundreds of millions developing. This is like saying "marginal cost of a Tesla is $X" while ignoring factory construction. **Label: UNVERIFIED without third-party audit of cost methodology.**

---

## 2. MISSING DATA POINTS (Critical Gaps)

### Missing Baseline #1: Phase II Success Rate Comparison
- Industry Phase II success rate: ~30% historically
- **What is the Phase II success rate for AI-discovered drugs specifically?**
- INS018_055 reaching Phase II is a *process milestone*, not an *outcome milestone*
- We need: Success/failure rates at each phase for AI-discovered vs. traditional drugs (n>20 minimum)

### Missing Baseline #2: Time-to-Market and Approval Data
- Zero AI-discovered drugs have reached Phase III completion or FDA approval
- **What's the denominator?** How many AI-discovered candidates have *failed* in trials?
- Recursion's 50+ petabytes and 2.2M weekly experiments—what's the *output* in approved therapies? (Currently: zero)

### Missing Comparison:
- No comparison to computational chemistry approaches that *aren't* branded as "AI" but use similar methods (e.g., traditional QSAR, molecular dynamics)
- **Demand:** Head-to-head comparison of AI platforms vs. sophisticated non-AI computational approaches on identical targets

---

## 3. COMPETING EXPLANATIONS / ALTERNATIVE INTERPRETATIONS

### Alternative A: Selection Bias in Target Choice
AI companies may be selecting "easier" targets with well-characterized biology (like IPF with known TGF-β pathways) where traditional methods would also succeed faster. The speed improvement may reflect **target selection strategy**, not AI capability.

### Alternative B: Survivorship Bias in Reported Metrics
We're hearing from companies that reached Phase II. **Where are the AI drug discovery companies that failed?**
- Atomwise's early partnerships?
- BenevolentAI's clinical setbacks?
- The denominator problem is severe.

### Alternative C: Cost Shifting, Not Cost Reduction
The $3M figure may represent costs shifted to:
- Earlier platform development (sunk costs)
- Partner organizations (Roche-Genentech paying for validation)
- Future phases (problems deferred, not solved)

---

## 4. FALSIFICATION TESTS

### Test 1: Blinded Retrospective Analysis
Take 10 drugs that failed in Phase II historically. Run them through current AI platforms *without revealing outcomes*. Can the AI predict failures? If not, the "acceleration" may just be faster failure.

### Test 2: Cost Audit by Independent Party
Commission a third-party accounting firm to conduct full