Share some training scripts and evaluation usage notebook

Help & Support
1

Hi, thank you for open-sourcing the AlphaGenome research repository.

While exploring the codebase (GitHub - google-deepmind/alphagenome_research: Research code accompanying AlphaGenome), I noticed that there is currently no dedicated README or documentation describing the training and evaluation workflows. Although some evaluation utilities and notebooks are available, the overall usage (e.g., how to train, how to run evaluation, expected inputs/outputs, recommended configs) is not clearly documented.

Would it be possible to provide:

  • A README section or separate documentation covering the training workflow

  • Instructions for running evaluation scripts

  • Example commands and expected outputs

Clear documentation for these workflows would greatly help researchers and students who want to reproduce results or build upon this work.

If anything like this already exists or is planned (e.g. in Colabs or external docs), pointers would be appreciated!

Thanks again for sharing this project.

2

Hi Jiahao,

Thanks for reaching out.

Please see the AlphaGenome manuscript for details on the evaluation and training workflows. regarding fine-tuning on custom datasets, we are working on releasing utilities to support this soon.

You can also access the new colab notebook for finetuning on a set of BigWigs here

Kind regards,
Tumi

3

Hi @Tumi_Makgatho , thanks for the update and for sharing the new finetuning and evaluation examples in the AlphaGenome repository (alphagenome_research/colabs at main · google-deepmind/alphagenome_research · GitHub). I really appreciate the effort your team has put into making the project more accessible to the community.

I have a quick question and suggestion. Since my background is in machine learning, I’m very interested in reproducing the results reported in the AlphaGenome manuscript and also using the framework as a benchmark to compare different models. Would it be possible to release a standardized evaluation script that implements all the key details and settings described in the paper? Having a unified and officially supported evaluation pipeline would make it much easier for researchers to reproduce the reported results and ensure fair and consistent comparisons across methods.

I believe this would be very valuable for the community and could help accelerate research and adoption. Thanks again for your work, and I look forward to future updates!

4

Hi Jiahao,

We appreciate your interest in benchmarking against AlphaGenome and your suggestion regarding a standardised evaluation script.

Currently, we are unable to release a standardized evaluation script. However, to ensure you can still achieve fair and consistent comparisons, we have publicly released the AlphaGenome predictions and targets for the majority of our key evaluations.

Kind regards,
Tumi