Hi all,
We’re excited to share a new Colab that demonstrates how to finetune AlphaGenome on custom genomic tracks. The finetuning consists of training new heads, using a frozen AlphaGenome base model.
The notebook covers the full finetuning workflow end-to-end:
-
Setting up the data pipeline: loading genomic intervals and corresponding track values.
-
Model initialization: loading the pretrained AlphaGenome checkpoint and initializing new output heads.
-
Training loop running with JAX/Haiku.
-
Inference: using your finetuned model to generate predictions and visualize results against ground truth.
We intentionally kept the training loop and recipe simple and hackable – the goal is to give you a clean starting point that you can easily adapt to your own use cases and experimental setups. We’re really looking forward to seeing how the community uses this!
The notebook is also available as a Vertex AI recipe for cloud-based training.
We’d love to hear your feedback! Let us know if you run into any issues or have suggestions for improvements!
Vincent (on behalf of the AlphaGenome team)