Do topological and quantum-inspired molecularrepresentations add value? Evidence from Africannatural-product-inspired antimalarial chemical space
Year: 2026
Type: preprint
Abstract: Topological and quantum-inspired molecular representations are increasingly proposed for drug discovery, but their predictive value relative to classical fingerprints on naturalproduct chemical space remains unclear.We benchmark a persistent-homology Topological Fingerprint (TFP), a Tucker-compressed Tensor Network Embedding (TNE), and simulated Quantum Kernel Scores (QKS) against classical baselines on 19 849 computationally generated candidates inspired by African natural-product antimalarials, with computationally predicted activity labels.Persistent homology decomposes the scaffold paradox into conserved H 1 ring features and divergent H 0 connectivity: 92.6 % of expanded molecules are whole-molecule Tanimoto-distinct from seeds while 69.3 % retain seed scaffolds.TNE compresses molecular tensors to 192 elements at 6.1× real-atom compression (reconstruction error 0.113) and retains docking-score-predictive information, competitively on PfDHFR (R 2 = 0.473 versus 0.461 for ECFP4) though below ECFP4 on PfATP4 and PfCRT.For activity prediction (n = 19 836, 5-fold cross-validation), ECFP4 achieves AUC 0.948 under random splitting and 0.822 when folds are grouped by Bemis-Murcko scaffold, the panel spanning only 632 scaffolds; the hybrid (TFP+TNE+QK) reaches 0.888, with QKS the principal contributor within that hybrid (∆ = -0.040),which nonetheless remains below ECFP4.Stacking TFP and TNE onto ECFP4 changed AUC by -0.000 2 (p = 0.084).The quantum kernel was not detectably different from a tuned RBF internally (p = 0.060) and performed lower on the ChEMBL label-source-shift panel (AUC 0.817 versus 0.847; p = 0.021).These representations add diagnostic and interpretive information under the tested protocols, but not predictive performance beyond ECFP4.
Source: ChemRxiv
Institutions: Université de Yaoundé I, Université de Sherbrooke
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7213956199