Terray Therapeutics uses NVIDIA DGX Cloud to train generative AI foundation models for small-molecule drug discovery

Overview

Terray Therapeutics leverages NVIDIA DGX Cloud and NVIDIA Base Command Platform to train COATI, a multimodal encoder-decoder foundation model for chemistry, reducing model training time from a week to a day and improving infrastructure utilization by over 4x compared to alternate cloud services.

The challenge

The chemical compound space is functionally infinite, with over 10^60 possible drug-like molecules. Typical drug discovery programs are highly inefficient and fundamentally constrained, able to explore only a few dozen to a few hundred compounds per week. When Terray was first developing COATI using a mix of on-premises GPU-based servers and traditional cloud services, provisioning and configuration of distributed training runs became increasingly challenging and tedious as models scaled up.

The solution

Terray Therapeutics leverages NVIDIA DGX Cloud and NVIDIA Base Command Platform to train COATI, a multimodal encoder-decoder foundation model for chemistry, pretrained on a dataset of hundreds of millions of small molecules, which converts chemical structures into numerical representations that can be used to generate molecules with desired properties. Terray uses a hybrid approach, training and building models on DGX Cloud and deploying and running inference on its on-prem cluster with NVIDIA RTX A6000 GPUs, drawing on dedicated NVIDIA AI experts to optimize workloads and monitor telemetry.

Reported business value

NVIDIA DGX Cloud improved infrastructure utilization by over 4x versus alternate cloud services, reduced model training time from a week to a day, and took less than one day to onboard onto DGX Cloud. Terray can now train multiple COATI variants in parallel to find the optimal pretrained embedding. Separately, integrating NVIDIA's cuEquivariance technology into Terray's proprietary TerraBind models accelerated computational performance by 3-4x and achieved an average 70% cost savings compared to PyTorch across all model sizes.