Geroscience Compound Atlas + Predictive Bench

Coverage Holes in Historical Chemical Repositories

Scientific humility requires exposing data absence. ChEMBL is historically optimized for approved oral small molecules and kinase/GPCR campaigns. Peptides, natural extracts, and gray-market compounds represent systematic coverage holes.

BPC-157 (Synthetic Pentadecapeptide)

Reason for Hole: Modality exclusion. Pentadecapeptides are excluded from standard small-molecule high-throughput binding screens. Zero structured binding assays exist in ChEMBL.
Scientific Status: Preclinical tissue-healing literature in rodents (E1); marketed heavily in gray-market wellness communities without mammalian lifespan proof (E0).

Epitalon (Synthetic Tetrapeptide)

Reason for Hole: Modality & provenance. Ala-Glu-Asp-Gly peptide with Russian literature citations; no curated target engagement entries in ChEMBL.
Scientific Status: Telomerase elongation claims in cultured human fibroblasts (E1); widely sold as an anti-aging elixir without replicated ITP lifespan validation (E0).

FakeGeroCompoundX99 (Negative Control)

Reason for Hole: Unresolvable test entity. Intentionally inserted into the seed watchlist to verify that our ingestion pipeline detects and reports resolution failures rather than silently hallucinating identifiers.
Scientific Status: Unresolved (404); correctly flagged as unmapped.

Benchmark Model Comparison — Scaffold Split vs Random Split

Evaluating L2-regularized Logistic Regression (Baseline) vs HistGradientBoosting (Contender) across 3 seeds (42, 123, 456) on 2048-bit Morgan circular fingerprints + standardized RDKit 2D descriptors on ChEMBL mTOR kinase activity (N = 561 molecules).

Split Strategy	Model Pipeline	AUROC	AUPRC	Recall @ 5% FPR	Brier Score
Bemis-Murcko Scaffold (Honest)	Baseline (Logistic Regression)	0.9735 ± 0.0203	0.9929 ± 0.0057	0.8496 ± 0.1207	0.0486 ± 0.0238
Bemis-Murcko Scaffold (Honest)	Contender (HistGradientBoosting)	0.9434 ± 0.0294	0.9855 ± 0.0095	0.7187 ± 0.1550	0.0820 ± 0.0392
Stratified Random (Diagnostic)	Baseline (Logistic Regression)	0.8912 ± 0.0482	0.9780 ± 0.0106	0.6528 ± 0.0856	0.0922 ± 0.0223
Stratified Random (Diagnostic)	Contender (HistGradientBoosting)	0.8766 ± 0.0467	0.9748 ± 0.0117	0.6667 ± 0.1623	0.1159 ± 0.0147

Key Takeaways for Hiring Managers & Reviewers:

Linear Baseline Generalization: Regularized linear models (Logistic Regression) outperform non-linear tree ensembles on sparse circular fingerprint representations ($p = 2057$), avoiding overfitting to dominant scaffold clusters.
Zero Scaffold Leakage: Enforcing Bemis-Murcko clustering ensures test compounds share zero scaffold topology with training compounds (Train ∩ Test = ∅).

Top Highest-Confidence False Positive Predictions

Compounds with inactive ground truth (pChEMBL < 6.0) predicted as active with highest confidence under scaffold split.

ChEMBL ID	Predicted Prob	True pChEMBL	Murcko Scaffold	Chemical & Pharmacophore Analysis

Constrained Molecular Generator (Genetic Algorithm)

Optimizing the frozen Phase 3 ChEMBL mTOR surrogate model while evaluating QED as an adjustable bias control (λ_QED) and penalizing assay interference (PAINS).

Generated Structures

200

100% Valid • 100% Unique

Novelty vs Training

99.5%

199 / 200 unobserved InChIKeys

Internal Diversity

0.655

1 - mean pairwise Tanimoto

Mean mTOR Prob

0.991

Surrogate active probability

QED Bias Sensitivity Analysis (λ_QED)

Evaluating how the molecular generator behaves when historical small-molecule oral drug-likeness priors are removed vs enforced.

QED Weight (λ_QED)	Objective Character	Mean mTOR Prob	Mean QED	Internal Diversity	Novelty Rate
λ = 0.0	Pure Target Affinity (Unconstrained)	1.000	0.059	0.563	100.0%
λ = 0.2	Balanced Tradeoff (Primary)	0.991	0.754	0.655	99.5%
λ = 0.5	Oral Drug-Likeness Constrained	0.985	0.635	0.670	96.5%

Chemical Space Takeaway (PCA Embedding):

Principal Component projection (PC1: 13.3%, PC2: 9.1% variance) demonstrates that GA-evolved candidates explore the mTOR training active chemotype manifold adjacent to benchmark binders, rather than drifting into unphysical junk space. When unconstrained (λ_QED = 0.0), molecules drift toward high molecular weight macrocyclic configurations (mean QED 0.059), demonstrating the restrictive nature of historical small-molecule heuristics.