M2UMol

Filtering DrugBank for molecules RDKit can turn into a graph yields 11,571 compounds. All of them have a 2D topological graph, 10,581 have curated biochemical annotations, 9,468 a 3D conformer, and only 5,899 a textual description — leaving just 4,325 with all four. Pre-training schemes that require complete modalities therefore learn from barely a third of the corpus, and hit the mirror-image problem at deployment, where a chemist screening a new compound has a SMILES string and nothing else.

M2UMol, from Huazhong Agricultural University and The Ohio State University and published in Nature Communications in February 2026, treats the other modalities as teachers rather than as inputs. Instead of aligning the 2D encoder against one fused "everything else" vector, it runs three independent contrastive objectives — 2D against 3D, against text, and against biochemical features — each mediated by a modal-specific adapter that generates a pseudo representation of the target modality from the 2D graph alone. A molecule qualifies for pre-training as long as it has its 2D graph and at least one of the other three, which is what lifts the usable corpus from 4,325 to 11,571. A modality classifier trained alongside the contrastive losses pushes the three generated vectors into distinguishable clusters, so the adapters learn modality-specific knowledge rather than a generic average.

What ships is consequently a unimodal artifact with multimodal knowledge folded in. At fine-tuning time the 2D encoder embeds the graph, the adapters synthesise the 3D, text and biochemical vectors, and a multi-head attention module fuses all four; the text and 3D encoders are discarded entirely.

#Key Features

• Multi-to-uni knowledge transfer: Three separate 2D-to-modality contrastive objectives replace the usual anchor-versus-fused alignment, so the modality count is not capped at two and none must be present for every molecule.
• Adapters that stand in for the missing modality: Three projection heads generate pseudo 3D, text and biochemical representations from the 2D embedding, recovering the multimodal fusion step when only a graph is available.
• Biochemical features as a first-class modality: Targets, enzymes and drug categories are one-hot encoded into a "Bio" modality, injecting curated pharmacology that structure- and text-only pre-training misses.
• Frozen-encoder fingerprints: The shipped package loads the checkpoint in eval mode and returns a fixed 128-dimensional vector straight from a SMILES string, as a drop-in replacement for a handcrafted fingerprint.
• Interpretability tooling: A bundled tool takes a bare SMILES and reports attention-weighted functional groups plus nearest neighbours retrieved across the generated modalities.

#Technical Details

The 2D encoder is a graph transformer: three GPS layers at 128 channels, each combining GINEConv message passing with 4-head multi-head attention, over random-walk positional encodings of walk length 20. During pre-training the text modality is encoded by Microsoft's PubMedBERT and the 3D modality by ComENet from the DIG library; both are further pre-trained rather than frozen. Pre-training runs 150 epochs on a single RTX 3090 in roughly 11 hours — orders of magnitude cheaper than methods pre-trained on millions of molecules.

On the eight MoleculeNet property-prediction benchmarks under scaffold splits, M2UMol averages 76.59 ROC-AUC against 74.98 for MoleculeSDE, 74.62 for MolFM, 71.34 for MolCLR and 69.83 for MegaMolBART, and takes the top result on five of eight datasets (ToxCast 66.3, SIDER 66.0, MUV 80.5, HIV 79.0, BACE 84.3); SMILES-sequence models such as MoLFormer-XL trail on the same splits. An ablation without multimodal pre-training scores 73.78, isolating the transfer as the source of the gain. On drug–drug interaction prediction over 191,570 interactions among 1,700 drugs and 86 event types, it gains 9.66% and 6.84% in accuracy over baselines under cold-start and scaffold splits. Swapped into the DrugBAN framework for drug–target interaction, it improves on DrugBAN by 2.60% AUC and 2.48% AUPR averaged over BindingDB (49,199 interactions) and BioSNAP (27,464).

#Applications

The immediate use is as a general molecular featurizer for teams that have SMILES and little else: virtual screening, ADMET and toxicity prediction, drug–drug interaction flagging and drug–target scoring all consume the same 128-dimensional embedding, and the encoder is small enough to fine-tune on a single GPU. The bundled analysis tool serves medicinal chemists directly, surfacing which substructures drove a prediction and which annotated compounds a new molecule resembles.

#Impact

M2UMol argues that careful pre-training design can substitute for corpus scale — 11k curated multimodal molecules and eleven GPU-hours matching or beating models trained on millions. That matters for the many biological domains where multimodal data is curated, small and patchy. The limitations are real: DrugBank confines pre-training to approved and investigational drug-like chemistry rather than the broader space ZINC- or PubChem-scale models see, three separate checkpoints ship because molecular encodings differ slightly across the property, interaction and target tasks, and the released code hard-codes CUDA device zero. The code is MIT-licensed on GitHub with a Zenodo archive; the pre-trained weights and processed corpus are distributed from a HuggingFace repository.