CAGE-QMol: A Constraint-Aware Quantum-Inspired Optimization Framework for Brain-Penetrant Multi-Target Alzheimer’s Drug Discovery
TL;DR
A mathematically grounded penalty selection rule for constrained drug-discovery QUBOs, a single end-to-end pipeline from raw public data to ranked, 3D-embedded leads, and reproducible head-to-head benchmarks between classical, quantum-inspired, and QAOA-simulated optimizers on a real Alzheimer’s task are contributed.
Abstract
Alzheimer’s disease remains one of the hardest disorders to drug, and most computational pipelines still tackle one target at a time and ignore brain penetration until late. We bring all of that into a single optimization problem. Our framework, called CAGE-QMol(Constraint-Aware Quantum-inspired Molecular optimization), turns the search for a brain-penetrant, dual BACE1/AChE inhibitor into a constrained quadratic unconstrained binary optimization (QUBO) and solves it with an ensemble of classical, quantum-inspired, and Quantum Approximate Optimization Algorithm (QAOA) backends. The pipeline begins with three real public datasets—MoleculeNet BACE, ChEMBL CHEMBL4822 (BACE1) and CHEMBL220 (AChE), and the TDC BBB_Martins blood–brain-barrier set—which together yield 12,465 unique molecules with at least one measured endpoint. Multi-task ExtraTrees predictors trained on Morgan ECFP4 fingerprints and physicochemical descriptors deliver scaffold-split test-set metrics of R2=0.624 (MAE=0.587) for BACE1 and R2=0.383 (MAE=0.793) for AChE. The optimizer combines these predictions with a Lipinski-based feasibility cone, a TDC-derived BBB classifier, and similarity-driven diversity into a constrained QUBO. We adapt the classical exact-penalty rule, λ>ΔS/δg, which guarantees every global minimizer of the penalized energy is feasible, and specialize it so that the multiplier is computed from the data rather than hand-tuned. A 10-qubit PennyLane QAOA circuit is benchmarked against exact enumeration, simulated annealing, genetic search, Bayesian TPE, and random search across ten seeds; the QUBO formulation lets a genetic solver match the exact ground state on every seed, while ablations show that removing the QUBO selection collapses the mean therapeutic score from 6.242 to 5.976 (p<10−3, paired t-test). Top-50 candidates exhibit a mean BBB probability of 0.716, a mean QED of 0.768, and 100% Lipinski feasibility, with leading scaffolds (tetrahydroisoquinolinone, methoxy-tetrahydronaphthalene-urea, indanone-piperidine) reproducing motifs found in published dual BACE1/AChE inhibitor families. This paper contributes (i) a mathematically grounded penalty selection rule for constrained drug-discovery QUBOs, (ii) a single end-to-end pipeline from raw public data to ranked, 3D-embedded leads, and (iii) reproducible head-to-head benchmarks between classical, quantum-inspired, and QAOA-simulated optimizers on a real Alzheimer’s task.