MolProphecy: Bridging medicinal chemists’ knowledge and molecular pre-trained models via a multi-modal framework

01 / Research question

Can molecular graph representations benefit from the chemical reasoning usually expressed in language?

02 / Innovation

Generate chemist-style knowledge with ChatGPT, encode it with an LLM and combine it with graph features through gated multi-head cross-attention.

03 / Mechanism

ChatGPT-generated chemist descriptions are encoded with LLaMA3 and fused with molecular graph embeddings through gated multi-head cross-attention. A limited human-review experiment refines conflicting generated knowledge.

04 / Experimental setup

Final Table 1 covers nine MoleculeNet datasets. Training/validation/test are 8:1:1: regression uses random splits; classification uses scaffold splits. Results use seeds 0–4. External solubility evaluation retains 949 molecules after removing MoleculeNet overlaps. Earlier four-dataset preprint results are separate.

Final article · online 2025, volume 2026 · Source ↗

05 / Results

FreeSolv RMSE is 0.796 ± 0.09; external solubility RMSE is 0.651 ± 0.046. The comparison retains stronger ESOL/Lipophilicity and HIV baselines.

06 / Ablation & limitations

Table 6 tests graph-only and knowledge-only variants. Table 10 improves challenge-case AUROC from 0.871 to 0.889 after chemist refinement, a limited proof of concept. Generated knowledge, possible pretraining exposure and context-dependent explanations constrain generalization; full autonomous or closed-loop chemistry is not demonstrated.

Generated knowledge is a proxy for expertise; property benchmarks alone do not prove an autonomous drug-discovery loop.

07 / My contribution

Last listed author; the contribution statement credits methodology, supervision, project administration and writing.

08 / Paper & code