MegaMolBART is a generative model for small molecules. It learns from SMILES strings, a way to encode molecular structures as text, and can generate new strings for drug-discovery workflows. The 45-million-parameter BART model combines a bidirectional encoder, which represents existing molecules, with an autoregressive decoder, which generates structures. It was pretrained on 1.54 billion molecules from ZINC-15. section 13.2.1
The pretrained encoder produces 256-dimensional embeddings: numerical representations of molecular structure. These embeddings can capture patterns that correlate with properties such as lipophilicity, even without explicit training on those properties. A downstream supervised model can use them as features to predict properties such as solubility, permeability, or activity. section 13.2.4 section 13.2.1
The same pretrained model supports several small-molecule tasks:
- Generation samples new SMILES strings from the learned distribution, optionally conditioning on a desired property such as predicted activity against a target. section 13.2.1
- Property prediction uses molecular embeddings as features for supervised models, including models of solubility and permeability. section 13.2.1 section 13.2.4
- Retrosynthesis decomposes a target molecule into simpler precursors. MegaMolBART representations can be adapted through task-specific fine-tuning, but reliable practical systems usually depend on models built and trained for retrosynthesis rather than MegaMolBART out of the box. section 13.2.1
Screening for FADS1 inhibitors
A Receptor.AI campaign used a library of roughly 4.3 billion small molecules generated with MegaMolBART, mixing in known FADS1 ligands as positive controls. The workflow then used progressively more selective computational filters to narrow the candidates. Generating a library was the starting point for screening, not evidence that its molecules were effective inhibitors. section 13.3.2
