AlphaGenome’s Authors on Predicting the Effects of DNA Variants

Google DeepMind product manager Dhavi Hariharan introduces genomics lead Žiga Avsec and researchers Natasha Latysheva, Tom Ward and Jun Cheng. They discuss AlphaGenome, a DNA sequence-to-function model described in their Nature paper.

Avsec describes the team’s goal as deciphering genomic sequences. One measure of progress is whether a model can predict how a small DNA change affects the molecular processes that read that sequence.

Cheng hopes better variant predictions can help research into genetic disease. He previously worked on AlphaMissense, which addresses protein-coding variants; AlphaGenome extends attention to the much larger noncoding portion of the genome.

Ward says the multidisciplinary team brings together different backgrounds. For him, genetics is compelling because a small change in DNA can have consequences across complex biological systems.

Avsec recalls earlier models that traded sequence length against resolution. Predicting gene expression required a longer input to capture relevant regulatory context, but came at the expense of some fine detail.

Latysheva contrasts AlphaGenome with earlier general and specialized models. Its combination is a long input of up to a million DNA bases, many output types and single-base resolution for many outputs.

A single model lets researchers examine a variant through several molecular lenses, she says, rather than coordinating separate tools for each layer. Higher benchmark accuracy is a research result, not proof that every individual prediction is correct.

Ward explains why this combination is difficult: long sequences, fine-grained predictions and multiple modalities all increase memory and computational demands when the model learns to predict experimental measurements from DNA.

The team split adjacent sequence regions across multiple TPUs and enabled the devices to communicate. Each device could process a shorter part without losing the ability to exchange information across the sequence.

Avsec recalls an early prototype that processed long sequences at high resolution nearly as quickly as an earlier training loop on synthetic data. Turning the idea into a working model required considerably more engineering.

Training also required feeding large amounts of data fast enough. Ward says some modalities were extremely sparse, so efficient compression and decompression helped remove a data-loading bottleneck.

The team also reviewed datasets for quality and diversity. Once data loading worked, researchers could iterate on the model and extend it from one-dimensional tracks to more complex outputs.

Cheng explains RNA splicing as bringing together non-contiguous pieces of genetic information needed for a functional protein. Variants can disrupt this process; modeling splice junctions required sparse two-dimensional outputs.

Latysheva describes contact maps as matrices representing interactions between DNA regions. The genome folds in three dimensions, and proximity between regulatory regions can affect gene expression.

The researchers report that adding contact maps and splicing did not noticeably degrade the other prediction modalities. Avsec suggests this may reflect the related biological processes that the measurements capture.

Evaluation has two parts: predicting molecular signals for DNA sequences the model has not seen, and assessing the effect of variants. For the latter, the model predicts a reference sequence and a changed sequence separately, then compares the outputs.

A single prediction can produce enormous outputs, making variant scoring slow. Engineers built an API that aggregates results on the same devices running the model, enabling many scoring operations in parallel.

The team divided evaluation across people and modalities and reused existing community benchmarks where possible. Cheng says they examined both molecular measurements and some organism-level outcomes, including examples where the model did poorly.

In one reported example, the model ranked certain experimentally studied cancer driver variants above random controls. The authors present this as a benchmark result, not a diagnosis for a patient.

The team says the work from a prototype to publication took under two years. It released an API alongside the preprint so researchers could make and visualize predictions without setting up the model and accelerator hardware themselves.

Cheng hopes researchers can use the model to prioritize potentially relevant variants and study how genomic regions regulate genes in different cell types. Those are research hypotheses that require further experimental investigation.

Users requested support for insertions, deletions, structural variants, embeddings and fine-tuning. The team says some insertion/deletion functionality already existed and user feedback helped uncover and fix a bug.

Latysheva says the model then covered human and mouse data, while researchers wanted additional species, modalities and cell types. The team uses this feedback to choose future work.

The group wants to summarize large numbers of variant scores to help researchers prioritize candidates, then inspect tissue- and cell-specific predictions in detail. Ward also discusses making much larger analyses more practical.

The researchers discuss releasing model weights and enabling work with embeddings, fine-tuning and new data. They also see potential in single-cell atlases for predicting effects in specific cell types.

The authors invite researchers to read the paper and share use cases and gaps. [Google DeepMind states](https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/) that AlphaGenome predictions were not designed or validated for direct clinical use.