A transformer-based molecular optimization framework. Optimize any molecule using an arbitrary scoring function while maintaining its integrity and purpose.
The optimization framework consists of three elements - attention loss, latent space distance, and a user-defined scoring function.
- Attention loss helps us in maintaining integrity of the molecule. Chemformer's attention heads penalize modifications which lead to an invalid molecule.
- Latent space distance helps us in sticking to the purpose of the original molecule - it measures the semantic difference between our modified molecule and the sourceo ne.
- User-defined objective can be any function which accepts SMILES and returns a scalar.
- Input: SMILES embedded using a Chemformer (
Chemformer embedding.ipynb) - Candidate generation: handles reranking of the candidates, chooses the best of them, and runs them through the benchmark. (
Optimize molecule.Sampler- currently only a greedy sampler is supported, feel free to add more!) - Attention loss: prediction of the most fitting modifications (
Optimize molecule.MolecularOptimizer.get_transformer_ll) - Latent space distance: cosine distance between embeddings of the source molecule and the candidate. Currently it's intertwined with scoring function, sorry :/
- Scoring: whatever you want it to be, get crazy.
Optimize moleculenotebook shows two examples.- A chance that a molecule will be a Mu-receptor antagonist (
score_fn), based on a neural network predictions (thanks to dr Sabina Podlewska for providing data) - Possibility of DILI-related injury, based on an XGBoost algorithm (gradient-less! Data from CAMDA challenge 2020)
- A chance that a molecule will be a Mu-receptor antagonist (
Sample optimizations of a molecule to be a better Mu-receptor antagonist (prediction format: [binding probability, 1-antagonist probability])
DILI chance optimization (don't pay too much attention to this one, we had a poor classifier)
