Abstract
<jats:p>Computational drug discovery for neglected diseases often faces a scaffold paradox: candidate molecules achieve extreme structural diversity (92.6 % Tanimoto diversity) while preserving conserved ring systems essential for bioactivity (69.3 % scaffold recovery). Classical extended-connectivity fingerprints do not adequately capture this topological phenomenon. To address it, we introduce three quantum-inspired representations—the Topological Fingerprint (TFP) based on persistent homology, the Tensor Network Embedding (TNE, 6.1× real-atom mean compression), and simulated Quantum Kernel Scores (QKS)—and evaluate them on a library of 19 836 African antimalarial candidates. A corrected classical-only benchmark shows that ECFP4 remains the strongest single descriptor (AUC 0.948), followed by AP (0.940), BPF (0.939), FCFP4 (0.918), MACCS (0.905) and PHCO (0.896); TFP (0.876) and TNE (0.722) capture complementary but lower signal. The hybrid descriptor (TFP+TNE+QK) achieves AUC 0.888, significantly below ECFP4 (p < 0.000 1) but above the standalone quantum-inspired descriptors, with QKS the principal positive contributor (ablation ∆ = −0.040). The quantum kernel shows no statistically significant difference from a gamma-tuned RBF baseline at any sample size (p = 0.060 at n = 19 849; p = 0.419 at n = 5 000), while significantly outperforming the linear kernel at scale (p ≤ 0.000 6). Crosspaper validation shows that H1 count correlates with resistance resilience scores (Spearman ρ = 0.312, p = 0.005 7, n = 77; pilot n = 14: ρ = 0.947, p < 0.000 1), yet this association is mediated by molecular size: after controlling for molecular weight the correlation collapses (ρpartial ≈ 0, p > 0.7), indicating that H1 count functions primarily as a size proxy rather than an independent topological predictor.</jats:p>