Scaffold-Controlled Evaluation of Molecular Representations for Antimalarial Activity Prediction
4 affiliations
- 1Department of Physics
- 2University of Yaounde I
- 3Department of Medical Imaging and Radiation Sciences, Université de Sherbrooke, Sherbrooke, QC, Sherbrooke, QC, Canada
- 4Department of Biochemistry
Abstract
We tested whether graph, transformer and topological molecular representations improve antimalarial activity prediction beyond an ECFP4-random-forest baseline under chemicaldistribution shift.The scaffold split is treated as a distribution shift rather than a stricter evaluation, following recent literature on molecular out-of-distribution specification, residual similarity leakage in scaffold and Butina splits, and multi-splitter evidence that Bemis-Murcko partitions are among the milder OOD regimes relative to ECFP4-cluster or nearest-neighbour strata [1][2][3].The open question is whether the representation-task alignment of learned models compensates for their OOD extrapolation deficit relative to ECFP4, or whether the fingerprint advantage is concentrated in the out-of-distribution tail.On a frozen panel of 19 836 African antimalarial natural products, five-fold cross-validation was repeated over five seeds (25 fold-seed replicates per model and split).Under the scaffold protocol, ECFP4-RF reached mean ROC AUC 0.830 0, above GIN-TFP (0.813 8), GIN-TNE (0.809 0), GIN (0.804 7) and ChemBERTa (0.786 7); the four contrasts against ECFP4-RF remained significant after Benjamini-Hochberg correction.Two protocol elements distinguish the comparison: fold-independent restoration of pretrained transformer weights removing crossfold leakage, and controlled insertion of persistent-homology and tensor-network descriptors into a graph architecture under scaffold-held-out evaluation.Persistent-image dimensions carried the largest projection-weight magnitudes in the fusion head, but the salience did not improve ranking.A nearest-training-neighbour Tanimoto decile stratification (Section 3.5.1)answers the open question: the ECFP4 advantage is concentrated in the OOD tail (deciles D1-D4, ∆ ∈ 0.016 --0.027) and in D9 (∆ = 0.017), while mid-range D5-D8 exhibit nearparity (∆ ∈ 0.001 --0.007).Representation-task alignment therefore does not compensate for the learned models' OOD deficit on this panel; the fingerprint advantage is confined to the out-of-distribution tail.The findings characterise the OOD response of the evaluated panel, splits, architectures and training budgets and do not generalise to all learned molecular models.
Related
- Hybrid-CatBoost: an empirical comparison of molecular representations for aqueous-solubility prediction on the TDC benchmark
- Task-adaptive multimodal molecular representations for structure-sensitive property prediction
- Transformer-based molecular fragment prediction using SMILES and DeepSMILES representations in a fragment-based drug discovery pipeline.
- DeMiR: Density-Matrix-Informed Representations for Molecular Property Prediction
- Computational Evaluation of Carbazole–Chalcone Urea and Thiourea Derivatives Against EGFR: Molecular Docking, Pharmacokinetic Profiling, Target Prediction, and Molecular Dynamics Simulation
BibTeX
@misc{njafa2026scaffoldcontrolled,
title = {{Scaffold-Controlled Evaluation of Molecular Representations for Antimalarial Activity Prediction}},
author = {Jean-Pierre TCHAPET NJAFA and Penabei SAMAFOU and Fon Wilfred MBACHAM and Serge Guy NANA ENGO and Myke Vital SAO TEMGOUA},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008517/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008517/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Scaffold-Controlled Evaluation of Molecular Representations for Antimalarial Activity Prediction
AU - Jean-Pierre TCHAPET NJAFA
AU - Penabei SAMAFOU
AU - Fon Wilfred MBACHAM
AU - Serge Guy NANA ENGO
AU - Myke Vital SAO TEMGOUA
PY - 2026
DA - 2026/09/08
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008517/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008517/v1
AB - We tested whether graph, transformer and topological molecular representations improve antimalarial activity prediction beyond an ECFP4-random-forest baseline under chemicaldistribution shift.The scaffold split is treated as a distribution shift rather than a stricter evaluation, following recent literature on molecular out-of-distribution specification, residual similarity leakage in scaffold and Butina splits, and multi-splitter evidence that Bemis-Murcko partitions are among the milder OOD regimes relative to ECFP4-cluster or nearest-neighbour strata [1][2][3].The open question is whether the representation-task alignment of learned models compensates for their OOD extrapolation deficit relative to ECFP4, or whether the fingerprint advantage is concentrated in the out-of-distribution tail.On a frozen panel of 19 836 African antimalarial natural products, five-fold cross-validation was repeated over five seeds (25 fold-seed replicates per model and split).Under the scaffold protocol, ECFP4-RF reached mean ROC AUC 0.830 0, above GIN-TFP (0.813 8), GIN-TNE (0.809 0), GIN (0.804 7) and ChemBERTa (0.786 7); the four contrasts against ECFP4-RF remained significant after Benjamini-Hochberg correction.Two protocol elements distinguish the comparison: fold-independent restoration of pretrained transformer weights removing crossfold leakage, and controlled insertion of persistent-homology and tensor-network descriptors into a graph architecture under scaffold-held-out evaluation.Persistent-image dimensions carried the largest projection-weight magnitudes in the fusion head, but the salience did not improve ranking.A nearest-training-neighbour Tanimoto decile stratification (Section 3.5.1)answers the open question: the ECFP4 advantage is concentrated in the OOD tail (deciles D1-D4, ∆ ∈ 0.016 --0.027) and in D9 (∆ = 0.017), while mid-range D5-D8 exhibit nearparity (∆ ∈ 0.001 --0.007).Representation-task alignment therefore does not compensate for the learned models' OOD deficit on this panel; the fingerprint advantage is confined to the out-of-distribution tail.The findings characterise the OOD response of the evaluated panel, splits, architectures and training budgets and do not generalise to all learned molecular models.
ER -