Welcome to the Genomics × AI Blog

Genomics X AI Editors · 2026-02-19

Adapting AlphaGenome to MPRA data

Alan Murphy, Alejandra Durán, Peter K. Koo · 2026-02-20

This post provides a high-level overview of how to use the AlphaGenome and Enformer repositories to extract modular convolutional encoders for short sequences — including links to the GitHub repositories — and summarises the results we achieved on perturbation assays. Foundation sequence-to-function models like AlphaGenome and Enformer are trained on ~1 Mb genomic windows to predict thousands of regulatory tracks. We show that their most transferable component is the convolutional encoder that learns local cis-regulatory grammar. By extracting this encoder from the long-range transformer and decoder modules, we: achieve state-of-the-art performance on lentiMPRA, STARR-seq, and CAGI5 benchmarks reduce inference cost by ~500× generalise across assays, species, and architectures This reframes foundation genomics models as modular regulatory representation engines, reusable for short perturbation sequences (100–300 bp) and regulatory design workflows. Code: AlphaGenome fine-tuning utilities | Full analysis and experiments

Fine-tuning AlphaGenome in native JAX/Haiku

Alan Murphy, Masayuki Nagai, Alejandro Buendia, Anshul Kundaje, Peter K. Koo · 2026-02-25

This post introduces alphagenome-ft, a lightweight Python package for fine-tuning AlphaGenome using native JAX/Haiku. We highlight workflows for adding custom prediction heads differing fine-tuning strategies freezing/unfreezing parameters attribution approaches Here we focus on general workflows applicable to genome-scale assays and custom heads. For fine-tuning the encoder for short sequences such as MPRA, see this post. Code: AlphaGenome fine-tuning utilities

Porting AlphaGenome to PyTorch

Danila Bredikhin, Alejandro Buendia, Martin Kjellberg, Christopher Zou, Xinming Tu, Anshul Kundaje · 2026-03-10

This post introduces alphagenome-pytorch, a faithful port of Google DeepMind's AlphaGenome model to PyTorch. We reproduce the full AlphaGenome architecture in PyTorch and release weights for fold-specific and distilled models. We also verify numerical equivalence of predictions across all output tracks with the original JAX checkpoint from DeepMind. Our package exposes a simple inference API that slots into any PyTorch project without requiring JAX, XLA, or TPU-specific tooling. In this post, we review the initial release for the package and walk through two use cases it enables: 1. Inference within existing PyTorch pipelines, which allows for genome-wide inference across tracks 2. Variant effect prediction and in silico mutagenesis (ISM) Contributions: Initial Development Team: Danila Bredikhin (Lead), Martin Kjellberg, Christopher Zou Finetuning and Validation: Danila Bredikhin, Alejandro Buendia, Xinming Tu Blog Post: Alejandro Buendia, Danila Bredikhin Code: alphagenome-pytorch

Benchmarking AlphaGenome on NVIDIA GPUs: latency, memory, and feasibility across sequence lengths

Xinming Tu, Alejandro Buendia, Anshul Kundaje, Sara Mostafavi · 2026-04-15

AlphaGenome pushes genomic modeling to sequence lengths up to 1 Mb, but practical adoption still comes down to a simple question: what fits on the GPU you actually have, and how long will it take? This post benchmarks the official JAX implementation and the community PyTorch port across seven NVIDIA GPUs: H200 (141 GB), H100 (80 GB), A100 (80 GB), L40S (48 GB), L40 (48 GB), A40 (48 GB), and RTX 6000 (24 GB). We report inference, heads-only finetuning, and full-weights finetuning on real genomic workloads. Key takeaways: Inference up to 1 Mb is feasible on every tested GPU from 48 GB upward; 1 Mb typically requires about 35-41 GB of peak memory Heads-only finetuning fits on every tested GPU up to 1 Mb, making it the accessible adaptation path for most labs Full-weights finetuning is memory-bound: 1 Mb is comfortable on H200, borderline on H100, reaches 524 kb on 80 GB cards otherwise, and tops out at 262 kb on 48 GB cards Above about 131 kb, memory scaling is close to linear for all three workloads

DOI Workflow Validation Post

Genomics X AI Editors · 2026-05-18

This short post validates the Genomics x AI production DOI workflow for newly accepted blog posts.

Zenodo Community Token Validation Post

Genomics X AI Editors · 2026-05-22

This short post validates the Genomics x AI Zenodo account and community deposit workflow.