Fine-tuning AlphaGenome’s splicing heads for RNA-seq samples, with reusable training code

anshulkundaje · x · 2026-07-29

What the post covers

The authors describe how they fine-tuned and evaluated AlphaGenome’s splicing heads for RNA-seq-derived transcriptomic signals.

The goal is to adapt AlphaGenome to new RNA-seq samples that were not seen in pretraining.
The workflow predicts sample-specific multimodal transcriptomic patterns at genomic intervals held out during fine-tuning.
The case study uses two RNA-seq samples from López-Oreja et al. (2023): one with the SF3B1 K700E mutation and one wild type.
The post emphasizes reusable preprocessing, training code, debugging lessons, and how different pieces of the pipeline came from AlphaGenome versus the authors’ own implementation.

The article is aimed at a domain audience and is positioned as a practical guide for reproducing and extending the setup.

More from Models