Does Geometry Matter? Comparing Fingerprint, Graph, and 3D Pretraining for Molecular Property Prediction

Materials discovery requires identifying molecules whose electronic properties match a target application across a chemical space too large for exhaustive evaluation. Density-functional theory (DFT) offers the accuracy needed to rank candidates but is far too expensive at screening scale, while semiempirical GFN2-xTB supplies optimized geometries and approximate electronic structure at a small fraction of the cost. We assemble a 30-million-molecule dataset of combinatorially generated conjugated candidates, each with a GFN2-xTB-optimized geometry and computed electronic properties, averaging 24.4 heavy atoms with roughly 90% between 15 and 35. We then ask which molecular representation maps this dataset into accuracy on OCELOT, a curated benchmark of π-conjugated molecules with DFT targets for HOMO, LUMO, and the gap. Pretraining a fingerprint encoder yields little: even with the full dataset and end-to-end fine-tuning it fails to beat a published descriptor baseline, and more pretraining makes its frozen LUMO predictions worse. Pretraining 3D encoders on the same molecules by coordinate denoising instead cuts error by up to 22% relative to random initialization, and our best 3D model outperforms every fingerprint and graph configuration tested, lowering test MAE by 11%, 9%, and 7% on the three properties relative to the strongest of them.