MegaMolBART is a large transformer natural language processing AI model developed by AstraZeneca and NVIDIA that is trained on SMILES (Simplified Molecular-Input Line-Entry System), the chemistry text representation of a molecular formula. MegaMolBART is capable of training large chemical language models with 1 billion or more parameters using NVIDIA’s Nemo Megatron framework. It includes scripts for accelerated data pre-processing and training that implements model parallelism.

It includes scripts for accelerated data pre-processing and training that implements model parallelism. MegaMolBART achieves 98.7% novel molecule generation meaning that 99 out of 100 molecules generated are valid and have never been seen before. That's 100X more chemical space that's being explored relative to the databases. MegaMolBART is part of NVIDIA Clara discovery which is a collection of SOTA AI frameworks, GPU-optimized applications and pretrained models for accelerated computational drug discovery.

• Reaction prediction

• Molecular optimization

• De novo molecular generation

Content

content goes here