NVIDIA BioNeMo lets researchers fine-tune biological AI models with LoRA in hours

NVIDIA BioNeMo has released a set of ready-made "recipes" for fine-tuning foundation biological models with LoRA (Low-Rank Adaptation). The toolkit lets research teams adapt large language models for proteins and DNA to specific scientific tasks without supercomputer resources.

Foundation models in biology

Computational biology is going through a transformation similar to the one NLP went through with BERT. Models pretrained on billions of biological sequences capture statistical patterns that classical rules describe poorly but transformers pick up well.

BioNeMo Recipes works with two flagship models.

ESM2 is a protein language model from Meta, trained on UniRef50. It analyzed hundreds of millions of amino acid sequences and learned to predict the structural and functional properties of proteins; versions range from 8 million to 15 billion parameters.

Evo 2 is a DNA language model from Arc Institute, trained on 9.3 trillion nucleotides from the genomes of 128,000 species. It predicts functional regulatory elements and models the effects of genomic mutations.

Both classes of models transfer well to specialized tasks: annotating protein functions, predicting subcellular localization, and assessing the pathogenicity of variants. But fully retraining such models is expensive and slow.

Why LoRA changes the math

Instead of updating all weights, LoRA adds compact low-rank matrices to the transformer layers, while the remaining parameters stay frozen. Only these small inserts go through backpropagation.

Key numbers for biological models:

• The number of trainable parameters drops by 90–99%
• The 3-billion-parameter ESM2 with LoRA fits on 1–2 GPUs instead of dozens of A100s
• The cost of an experiment falls from thousands to single-digit dollars per GPU-hour
• Training time shrinks from weeks to a few hours
• Quality on narrow, specialized tasks is comparable to full fine-tuning

This matters a great deal in biology: lab datasets are often small, with hundreds or thousands of examples. Retraining an entire large model on that volume leads to a loss of generalization, whereas LoRA, with a minimal number of new parameters, performs noticeably better.

What BioNeMo Recipes includes

BioNeMo Recipes consists of ready-made configuration pipelines with documentation, examples and tests. The researcher picks a model, a task and a dataset, and the recipe then automatically sets the hyperparameters, configures weight loading and sets up logging.

The toolkit includes:

• Support for LoRA and full fine-tuning for ESM2 and Evo 2
• Integration with NVIDIA NeMo Framework and DGX infrastructure
• Ready-made formats for protein and genomic datasets
• Logging through Weights & Biases and automatic checkpoints

The target audience is biomedical groups and pharmaceutical companies that need to specialize a model for a particular organism, protein type or disease. Typical tasks include predicting the toxicity of therapeutic proteins, finding functional sites in the genome, assessing the pathogenicity of rare variants, and designing enzymes with specified properties.

What this means

BioNeMo Recipes lowers the barrier to entry for teams without large ML infrastructure. A pharmaceutical lab or an academic group with two GPUs can now specialize a foundation model for its own conditions. That speeds the path from scientific hypothesis to computational tool and, down the line, from discovery to therapy.

*Meta has been designated an extremist organization and is banned in Russia.