In my previous article (Rational Drug Design), I shared a session presented at NVIDIA GTC from Isomorphic Labs.

The focus of the session was Rational Drug Design With AlphaFold 3, where the team provided insights as to how they develop general AI models focused on biology.

As part of our strategic partnership with Google, we have been fortunate enough to work with the exceptional team at DeepMind. We actively leverage their specialist AI models (such as AlphaFold) and collaborate on future opportunities.

Although I expect the DeepMind team to continue to lead the field, delivering “bleeding-edge” capabilities, the proprietary nature of these specialist AI models can be restrictive.

Therefore, as part of our strategy to support drug development, we leverage proprietary and open AI models to predict compound structures and optimise synthesis to accelerate the discovery process.

Recently, we partnered with WWT to test a new multi-model Structural Prediction platform, leveraging the NVIDIA Inference Microservices (NIM) architecture, which is a set of optimised inference microservices that accelerate the deployment of AI models on any NVIDIA-accelerated infrastructure.

Synthesising large molecules like Monoclonal Antibodies (mAbs) without prior structural knowledge is a costly, time-consuming gamble. Without the confirmed binding affinity, effort can be wasted on non-viable candidates.

Therefore, predicting compound structures upfront identifies high-potential drug candidates and eliminates non-starters, drastically cutting the synthesis time and expense.

With this goal in mind, we tested the open source Boltz-2 structural biology foundational model, formulated based on complexes not in the training data.

The results were impressive, with improved scale and time to prediction.

Prediction throughput increased 260-fold from 3 complexes (default baseline) to 780 complexes in parallel between two separate tools.

Prediction speed increased 5.5x on average (default baseline) from 11 minutes per complex to 2 minutes per complex.

Finally, we validated that the accuracy of Boltz-2 is comparable to that of AlphaFold2, while being 52 times faster.

We are currently working on a case study with WWT, where we plan to share more details of our testing.

However, the initial findings support our strategy to industrialise our use and scale of computational science, targeting proprietary and open AI models, to accelerate drug discovery.

enterprise architecture ea enterprise_architecture artificial_intelligence ai machine_learning ml generative_ai large_language_model llm natural_language_processing nlp transformer data prompt_engineering google deepmind alphafold nvidia nvidia_inference_microservices nim boltz-2 life_sciences health healthcare lifeinwork