Discovery of Novel Polycyclic Aromatic Inhibitors of Anaplastic Lymphoma Kinase (ALK) via Integrated QSAR Modeling, Machine Learning, and Pharmacophore-Guided Virtual Screening with AI-Assisted Mechanistic Interpretation of Structure–Activity Relationships
Year: 2026
Type: preprint
Abstract: Anaplastic Lymphoma Kinase (ALK) is a validated oncogenic driver in non-small cell lung cancer, but acquired resistance, most critically the solvent-front mutation G1202R, erodes the durability of all approved inhibitors, requiring new scaffolds without cross-resistance liabilities. An interpretable GFA-MLR QSAR model (R² = 0.60, Q² = 0.58, r²pred = 0.56) and a Genetic-Algorithm-optimized machine learning classifier ensemble (Random Forest, XGBoost, k-NN, Naïve Bayes) were built independently from the same descriptor set of 287 ALK inhibitors (225 descriptors, 109 pharmacophore hypotheses; training n = 230, test n = 57), with GA-based feature selection applied uniformly before training each classifier. Random Forest (accuracy = 98.7%, κ = 0.98) selected two of the same high-frequency pharmacophores (A-T1-1, J-T5-1) identified by the GFA-MLR equation, despite no cross-reference between the two branches during model development. A linear regression model and a nonlinear ensemble classifier therefore converged independently on the same structural determinants of ALK potency. The physicochemical rationale underlying each GFA-MLR coefficient was then derived through a structured, AI-assisted interpretation framework anchored to the known ALK binding-site architecture. Together, the dual-branch convergence and the AI-assisted coefficient interpretation constitute a cross-validated alternative to single-model QSAR interpretation, a strategy that, to our knowledge, has not previously been applied in kinase inhibitor discovery. Pharmacophore-guided NCI screening yielded 28 candidates; eight (28.6%) inhibited ALK at 10 µM, and three, NSC 290954, NSC 196403, and NSC 715055 (IC₅₀ = 2.15, 4.02, 5.95 µM), spanning anthraquinone, aminoquinoline, and 4-anilinoquinazoline scaffolds distinct from approved inhibitors, were confirmed by dose–response assay, docking/molecular dynamics (ATP-hinge binding, Met1199 engagement), and NCI-60 cellular profiling.
Source: Research Square
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225789021