Alphagenome-ft: a fine-tuning package for alphagenome & Perturbation Assay Fine-Tuning Results

Hi all,

We’re excited to share alphagenome-ft, a lightweight pip-installable, wrapper for structured fine-tuning of AlphaGenome — along with results on perturbation assays using encoder-only adaptation.

We built alphagenome-ft to make the process of adapting alphagenome to new datasets modular, explicit, and easy to extend.

The package supports:

  • Adding custom or predefined output heads

  • Heads-only (linear probe), LoRA, or full fine-tuning

  • Progressive unfreezing (encoder / transformer / decoder)

  • Attribution methods (e.g. grad×input, ISM)

  • All without modifying the original AlphaGenome codebase

In addition to the wrapper itself, we used it to fine-tune AlphaGenome on short perturbation assays (~300 bp inputs). By extracting and adapting the convolutional encoder for short sequences, we achieve state-of-the-art performance on lentiMPRA and STARR-seq while reducing inference cost by ~500× compared to the full model.

The encoder-only setup also transfers well to zero-shot regulatory variant prediction (CAGI5), and we observed that frozen encoders can generalize slightly better out-of-distribution. We additionally identify a simple protocol improvement: matching the aggregation window to assay length yields consistent gains over the original protocol.

The wrapper and detailed write-ups are available here:
Blog (wrapper): https://genomicsxai.github.io/blogs/2026-003/
Blog (perturbation fine-tuning): https://genomicsxai.github.io/blogs/2026-002/

We’d very much welcome feedback from the community — especially from others experimenting with fine-tuning or adapting AlphaGenome to new assays.

2 Likes