NOVA is taking SOTA models to the next level
@metanova_labs optimized version of Boltz-2 is now available on our Github
Boltz-2 by @MIT and @RecursionPharma is a groundbreaking model that excels at simultaneously predicting complex 3D structures and binding affinities, but is very computationally expensive. An average sample submitted to NOVA Compound takes around 1 min 6 s to score on an A100 GPU with the out-of-the-box version of Boltz-2.
Our optimized version improves validation efficiency by decreasing inference time per sample to 41s, meaning an improvement of 61% in throughput under the same conditions.
For comparison, @nvidia offers an inference microservice for Boltz-2 that increases throughput by 50% on the average sequence length of monoamine transporters. To deploy this in production, however, it requires a license that costs $4,500 USD/year/GPU. Our miners, and the broader open source community, now have access to a faster version (under an MIT license) than the closed-source alternative.
Faster and accessible to all. This is how we accelerate drug discovery.
šgithub.com/metanova-labs/nova/...ā¦
#Bittensor #SN68 #drugdiscovery #opensource #deSci
