Predict Protein Structures and Properties with Biomolecular Large Language Models

The NVIDIA BioNeMo service is now available for early access. With the BioNeMo service, scientists and researchers now have access to pretrained biomolecular…

NVIDIA logo
NVIDIA
•
Vanessa Braunstein
•3 min read•intermediate•
--
•View Original
BERTJAXLarge Language ModelsPyTorch

Overview

The article discusses NVIDIA's BioNeMo service, a framework for training and serving biomolecular large language models (LLMs) designed for predicting protein structures and properties. It highlights the capabilities of BioNeMo in supporting AI drug discovery workflows through pretrained models and cloud API access.

What You'll Learn

1

How to utilize the BioNeMo service for predicting protein structures

2

Why ESM-1nv embeddings enhance model accuracy for protein properties

3

When to apply OpenFold for rapid 3D protein structure predictions

Key Questions Answered

What is the BioNeMo service and its primary function?
The BioNeMo service is a cloud-based framework for training and serving biomolecular large language models, specifically designed for predicting protein structures and properties. It allows scientists to access pretrained models and develop workflows for AI drug discovery.
How does ESM-1nv improve protein property predictions?
ESM-1nv, a large language model, learns patterns between amino acids to create embeddings that can be used to fit smaller supervised models for predicting protein properties like subcellular location and thermostability, resulting in higher accuracy.
What are the key features of the BioNeMo service?
Key features of the BioNeMo service include a fully managed browser-based interface, API endpoints for protein LLMs, accelerated OpenFold model for fast 3D predictions, and interactive visualization of protein structures.
What is OpenFold and how does it compare to AlphaFold-2?
OpenFold is a reproduction of DeepMind's AlphaFold-2 model for predicting 3D protein structures from amino acid sequences. It achieves similar accuracy with a median backbone prediction accuracy of 0.96 Å RMSD95 and is up to 6x faster in predictions.

Key Statistics & Figures

Median backbone prediction accuracy
0.96 Å RMSD95
Achieved by OpenFold in predicting 3D protein structures.
Speed improvement
up to 6x faster
OpenFold's enhancements in the MSA generation step allow for quicker predictions compared to previous models.

Technologies & Tools

Some links below are affiliate links. We may earn a commission if you make a purchase.

Service
Bionemo
A framework for training and serving biomolecular large language models.
Model
Openfold
A model for 3D protein structure prediction based on AlphaFold-2.
Model
Esm-1nv
A large language model for protein embeddings.
Framework
Pytorch
The framework used for implementing OpenFold.

Key Actionable Insights

1
Leverage the BioNeMo service to streamline your AI drug discovery workflows by integrating its API into your applications.
This service provides a quick start for researchers looking to utilize large language models for predicting protein structures, making it easier to develop and test new drug candidates.
2
Utilize ESM-1nv embeddings to enhance the accuracy of your downstream models focused on protein properties.
By training smaller models with ESM-1nv embeddings, researchers can achieve superior predictive performance for various protein characteristics, which is crucial for effective drug design.
3
Consider using OpenFold for rapid protein structure predictions in your research.
With its fast processing capabilities and high accuracy, OpenFold can significantly reduce the time required for structural biology studies, allowing for quicker iterations in research.