Introduction

AI models are changing how we think about and access information on an unprecedented scale. These methods, ranging from regression, classification, and even generation, allow the exploration of data-driven domains in unique ways. BioNeMo seeks to serve as a platform for accessibility to multiple bio-related AI tools to serve researchers in their challenges. The BioNeMo framework facilitates centralized model training, optimization, fine-tuning, and inferencing for protein and molecular design.

BioNeMo Framework

BioNeMo Framework provides versatile functionalities for developing, training and deploying large scale bio-based models. BioNeMo allows users to build biomolecular models by providing access to pre-trained models, creating workflows to fit downstream task models from embeddings, and generating biomolecules that meet user-specified criteria based on the fit model. Built for supercomputing scale, the framework allows developers to easily configure and deploy distributed multi-node jobs with minimal code.

The underpinnings of the biological research framework rely on NeMo, a framework initially established for conversational AI methods. NeMo provides a robust environment for working with large learning models, including Megatron models. The BioNeMo Framework provides enhancements to PyTorch Lighting, such as hyperparameter configuarbility with YAML files and checkpoint management. Users can conveniently and quickly train models using these features, test them for desired tasks, and integrate them alongside existing applications.

Some of the key features of BioNeMo Framework are:

• Development and training of large transformer models using NVIDIA’s Megatron framework.

• Easy to configure multi-GPU, multi-node training with data parallelism, model parallelism, and mixed precision.

• Model training recipes that can be readily implemented on DGX compute infrastructure.

• Logging with Tensorboard and Weights and Biases to monitor the model training process.

BioNeMo Models: Overview

Model	Modality	Uses	Trained/Converted Checkpoints on NGC
OpenFold	Protein	Protein Structure Prediction	Public checkpoint fine-tuned by BioNeMo
DiffDock Score Model	Protein + Molecule	Generation of Ligand Poses	Public checkpoint converted to BioNeMo format
DiffDock Confidence Model	Protein + Molecule	Generation of Ligand Poses	Public checkpoint converted to BioNeMo format
EquiDock DIPS Model	Protein	Protein-Protein Complex Formation	BioNeMo checkpoints pre-trained from scratch
EquiDock DB5 Model	Protein	Protein-Protein Complex Formation	BioNeMo checkpoints pre-trained from scratch
ESM-2nv 650M	Protein	Representation Learning	Public checkpoint converted to BioNeMo Format
ESM-2nv 3B	Protein	Representation Learning	Public checkpoint converted to BioNeMo Format
ESM-1nv	Protein	Representation Learning	BioNeMo checkpoints pre-trained from scratch
ProtT5nv	Protein	Representation Learning	BioNeMo checkpoints pre-trained from scratch
MegaMolBART	Small Molecule	Representation Learning + Molecule Generation	BioNeMo checkpoints pre-trained from scratch
MolMIM	Small Molecule	Representation Learning + Molecule Generation	BioNeMo checkpoints pre-trained from scratch

For more information about the models included in BioNeMo Framework, refer to the Model Cards linked in the table above or the original publications referenced in the respective model descriptions. One can also go to the linked NGC Model Catalog Pages to download the model checkpoints, integrate it into your own BioNeMo FW workflows. You can use the tutorials provided in the documentation to learn how to plug in a bionemo model and run training/inference or downstream tasks.

Refer to the Quickstart Guide for details on how to get started with BioNeMo Framework.