When Three-Dimensional Conformer Ensembles Improve Molecular Property Prediction Beyond Two-Dimensional Fingerprints: A Systematic Study

arXiv:2606.08825v1 Announce Type: new Abstract: When do three-dimensional conformer ensembles improve molecular property prediction beyond two-dimensional fingerprints? We provide the first systematic, mechanistically grounded answer. Through ~1,000 experiments spanning 13 model configurations, 14 regression targets, and 2 classification targets across MoleculeNet, QM9, and MARCEL benchmarks, we discover selective complementarity: conformer ensemble statistics extracted via Distribution Kernel Operators (DKOs) yield statistically significant RMSE reductions on solvation-dependent properties (ESOL -11.0%, p < 10^{-9}; FreeSolv -13.5%, p < 3x10^{-5}; 10-seed paired validation) while providing no benefit for electronic or steric tasks. Three lines of evidence confirm this selectivity has a physical rather than statistical basis: improvement is larger under scaffold splits than random splits (+11.9% vs. +8.5% on ESOL), concentrates on large, flexible molecules (+18.9% for heaviest quartile), and grows monotonically with training data. We establish a four-tier performance hierarchy: end-to-end 3D GNNs (SchNet, PaiNN; 21-42% over fingerprints) >= engineered physicochemical descriptors (PMI/SASA/USR) > Morgan fingerprints + XGBoost > all neural conformer ensemble methods, confirmed by two architecturally diverse GNNs and revealing that the pre-computed feature bottleneck limits ensemble approaches. Feature attribution and mutual information analysis expose the mechanistic asymmetry: conformer mean features carry 2-8x more information per feature than fingerprint bits, yet covariance features contribute <2% of model signal, explaining why five simple scalar invariants outperform all complex covariance architectures (p < 0.001). These findings yield an empirical property taxonomy and a practical decision framework for when conformer generation is worth the investment.