PROTECT

In this project, we aim to explore the potential of transformer models to optimize molecular properties and/or generate new molecules with desired non-toxic properties.

Transformer models are well-known for capturing long-distance relationships within an input sequence. This ability recruits them as promising candidates for capturing dependencies within the studied small molecules that can be redeemed important for understanding and controlling for toxic properties. The transformer architecture provides generalizability by being pre-trained on large unsupervised dataset then fine-tuned on small downstream datasets. It also comes in different flavors that makes suitable to perform molecular property prediction, optimization, and/or generation.