← Back to Part 1 | Home

Paying the Entropic Cost in Advance: Matched Molecular Pair Analysis, ML Potency Benchmarks & Reaction Planning

1. The Non-Small Cell Lung Cancer Case: Crizotinib to Lorlatinib

Google DeepMind demonstrates this principle using two FDA-approved tyrosine kinase inhibitors (TKIs) targeting ALK/ROS1 oncogenic fusions:

In Crizotinib, 5 single bonds allow the molecule to adopt hundreds of inactive solution conformations. Lorlatinib closes the terminal ring into a 12-membered macrocycle. The Antigravity agent demonstrated that closing this ring not only eliminates the rotatable bond penalty, but also significantly enhances predicted penetration across the blood-brain barrier—a critical therapeutic requirement for secondary brain metastases.

2. Matched Molecular Pair Analysis (MMPA) Without Hallucinations

To prove that cyclization was the active causative factor rather than accidental side-chain substitutions, the Antigravity agent automatically conducted Matched Molecular Pair Analysis (MMPA):

// Antigravity Agent MMPA Transformation Rule:
Linear Substructure [R1-NH-CH2-CH2-R2] 
  ==> Cyclized Macrocycle [R1-N(Ring)-CH2-CH2-R2]
Result: Delta_pIC50 = +1.18 log units (~15.1x potency gain)
Rotatable Bond Reduction: -5 bonds
Conformation Penalty Delta: -4.2 kcal/mol

Because MMPA requires strictly identical contexts around the changing core, Antigravity’s RDKit-grounded skill ensures that pairs differing by confounding functional group modifications are excluded from the causal estimate.

3. Why Random Train/Test Splits Fail in Chemistry ML

A notorious pitfall in computational drug discovery is the "random split illusion." When machine learning models are trained on randomly split chemical databases, compounds with nearly identical core scaffolds appear in both training and test sets. The model simply memorizes the scaffold, achieving misleadingly high R² scores (>0.90) that plummet to zero when tested against novel scaffolds in the wet lab.

The Antigravity agent enforced rigorous data hygiene:

• Chronological Splits: Training on older published patents/papers (pre-2022) to predict the potency of recently discovered analogs (post-2024), simulating real prospective discovery.
• Bemis-Murcko Scaffold Splits: Ensuring that no molecular scaffold in the test set shares a ring architecture with the training set, giving an unvarnished estimate of out-of-distribution generalizability.

4. De Novo Macrocycle Design with Retrosynthetic Verification

Finally, Antigravity did not stop at predictive scoring. It generated novel macrocyclic candidates and paired each design with:

1. Atom-mapped reaction transforms: Verifying that each intermediate can be synthesized using known organic reactions (such as Suzuki-Miyaura cross-coupling followed by intramolecular macrolactamization).
2. Electron-pushing arrow diagrams: Graphically illustrating the curly-arrow mechanistic steps for the cyclization stage so bench chemists can immediately assess feasibility.
3. Predicted target potency: Re-scoring candidate designs against the validated scaffold-split model.