MolMAE

Most molecular foundation models learn from SMILES strings, 2D molecular graphs, or 3D atomic coordinates. But a molecule does not present its atoms to the world — it presents its surface. Solvation, electrostatic complementarity, membrane permeation, and ligand-protein recognition all happen at the solvent-excluded interface, which is only implicit in a bond graph or a coordinate list. MolMAE is a masked autoencoder built to make it explicit.

Developed by Jiaqing Li in the Department of Industrial and Molecular Pharmaceutics at Purdue University and posted to bioRxiv in July 2026, MolMAE takes three complementary views of the same molecule: a surface point cloud, a 3D molecular graph, and SMILES-derived fragment and functional-group tokens. Its central idea is functional-group-aligned co-masking — rather than masking each modality independently, the model picks one chemically meaningful functional group and hides the corresponding region in every view at once. The surface patches near that group, the atoms that compose it, the fragment tokens they generate, and the functional-group token itself all disappear together, so reconstruction requires relating local geometry, electronics, topology, and chemical semantics to one another.

MolMAE is framed as complementary to, not a replacement for, atom-coordinate-based 3D pretraining. It beats Uni-Mol on solubility, blood-brain-barrier permeability, and BACE bioactivity, while Uni-Mol remains stronger on hydration free energy, lipophilicity, and clinical toxicity — a split the paper attributes to MolMAE's single-conformer surface.

#Key Features

• Surface as the primary signal: Solvent-excluded surfaces are generated with NanoShaper and reconstructed as the leading pretraining objective, with graph, fragment, and functional-group reconstruction supplying complementary supervision.
• Physicochemical field reconstruction: Beyond geometry, the surface branch predicts electrostatic potential, Fukui-related reactivity descriptors, Gaussian and mean curvature, and a distance-weighted local atom-type distribution — so two patches with identical shape but different chemistry are not conflated.
• Chemically aligned masking: A single selected functional group defines the masked region across all four views, masking roughly 20% of graph atoms and 15% of surface patches within the same chemical neighborhood.
• Staged objective schedule: Surface coordinate and atom/bond reconstruction are activated first, followed by quantum descriptors, stereochemistry, fragment semantics, and finally cross-modal alignment losses, preventing noisy semantic targets from dominating early training.
• Verified representational complementarity: Centered kernel alignment of 0.44 and nearest-neighbor Overlap@10 of 0.32 against Uni-Mol show the two encoders organize chemical space differently.

#Technical Details

The pretraining corpus was curated from ChEMBL 36: small molecules with valid canonical SMILES, parent forms only, at least one standardized activity measurement against a human single-protein target (assay confidence ≥7; IC50, Ki, Kd, or EC50), then RDKit sanitization, InChIKey deduplication, and lead-like filters on molecular weight, cLogP, hydrogen-bond donors and acceptors, rotatable bonds, and ring count. That yielded 261,840 molecules, of which 261,448 survived conformer generation, NanoShaper surface construction, and AIMNet2-based quantum descriptor mapping.

Each surface is tokenized into patches: patch centers are chosen by farthest point sampling with a count of clip(round(N/8), 64, 256) for a molecule with N surface points, and each center gathers its 32 nearest points. Patch features combine center-relative coordinates, normals, quantum descriptors, multi-scale curvature, and nearby-atom context, and are projected to a 192-dimensional hidden space in a Transformer encoder. The graph branch reconstructs atom type, bond type, chirality label, and a continuous chirality-geometry scalar; the sequence branch runs Transformer encoders over hashed fragment and functional-group vocabularies. Surface-graph cross-attention fuses the branches. On seven MoleculeNet tasks under matched scaffold splits across three seeds, MolMAE reaches an ESOL RMSE of 0.740 against Uni-Mol's 0.867, a BBBP ROC-AUC of 93.72% (8.67 points above Uni-Mol), and 85.86% on BACE.

#Applications

The pretrained encoder serves as a general-purpose feature extractor for property prediction on small labeled datasets, either fine-tuned end to end or frozen behind a lightweight head — the linear-probe setting where MolMAE leads on all three regression benchmarks. That makes it most relevant to early-stage drug discovery teams screening lead-like compounds for aqueous solubility, permeability, and target bioactivity, where labeled data are scarce and surface polarity and exposed shape drive the endpoint. Concatenating MolMAE and Uni-Mol embeddings improves linear-probe performance over either alone, suggesting value as an additional descriptor channel rather than a replacement.

#Impact

MolMAE is a single-author preprint that has not yet been peer reviewed, and no code or weights have been released, so its results cannot currently be reproduced independently. Its 261K-molecule corpus is also small next to the multi-million-compound datasets behind most chemical foundation models. The contribution is the ablation-supported case that molecular surfaces carry pretraining signal beyond graphs and atomic coordinates: removing the surface branch degrades ESOL to 0.785 RMSE, geometry-only surfaces recover 0.773, and adding physicochemical fields reaches 0.740. The costs are real: surface generation, descriptor mapping, and dense point-cloud processing make pretraining and fine-tuning more expensive than coordinate-based approaches, and one conformer per molecule limits properties governed by conformational ensembles.