AlphaGenome: one model to read the regulatory genome

Most of the genome doesn’t code for proteins, and almost every common disease variant lives out there in the non-coding 98%. The hard part of my PhD isn’t finding those variants, it’s working out what they do. So when a single model claims it can read a megabase of DNA and predict splicing, expression, chromatin and 3D contacts all at once, it lands right on the question I care about most. This is my journal-club read of AlphaGenome, and why I think it changes the baseline for non-coding variant interpretation.

The two trade-offs it sets out to break

Before AlphaGenome, sequence-to-function models forced you to pick a side of two trade-offs.

The first is context length versus resolution. Short-context models can be base-pair precise but they can’t see a distal enhancer sitting 500 kb away. Long-context models can see the enhancer but predict in coarse bins, smearing out exactly the single-base effects that matter for splicing. The second is specialisation versus breadth. A model trained on one assay can be excellent at that assay and completely blind to a variant’s effect in another modality.

AlphaGenome’s design goal is to refuse the choice, to be long-context and high-resolution, generalist and competitive with specialists. That framing is what makes it worth a careful read rather than a benchmark-table skim.

What it actually is

At a glance: it takes a 1 Mb DNA window and predicts 11 modalities, 5,930 human tracks and 1,128 mouse tracks, at base resolution where it needs to be, including 2D outputs for chromatin contact maps. The architecture is a U-Net-like backbone: convolutions pick up local sequence grammar, transformers carry long-range enhancer–promoter dependencies, and the model keeps both 1D sequence embeddings and 2D pairwise embeddings so it can emit Hi-C-style contacts.

The training is a two-stage trick worth understanding, because it’s the part most relevant to anyone who wants to use a model like this at scale.

Stage one is pretraining on observed genomic tracks, under 4-fold genomic splits so they can measure generalisation to unseen intervals. Stage two is distillation: a single student model learns from an ensemble of teachers on augmented and mutated sequences. That second stage is the quietly important one, it keeps ensemble-like accuracy while collapsing everything into one model you can call once per variant. If you’ve ever tried to scan millions of variants through an ensemble, you know why that matters.

The scoreboard, read sceptically

The headline numbers are genuinely strong: state-of-the-art on 22 of 24 genome-track tasks and 25 of 26 variant-effect tasks. But a generalist winning broadly can hide weak spots by averaging over easy benchmarks, so the parts I find more convincing are the places it beats specialists on their home turf.

• Splicing. AlphaGenome models splicing at three levels, splice-site probability, splice-site usage under competition, and explicit donor–acceptor junction formation, and predicts all three plus RNA-seq coverage together. The splice-junction head is the novel piece: most predictors score donors and acceptors independently, while AlphaGenome models the pairing. It reaches state-of-the-art on six of seven splicing-variant tasks. The one loss (MFASS, narrowly behind Pangolin) is worth noting precisely because the authors don’t hide it.
• Expression / eQTLs. It improves tissue-weighted sign prediction over Borzoi and stays useful across distance bins, though, honestly, distal variants are still hard, which the paper admits.
• Chromatin contacts. This is where it’s unusual. Its 2D pair representations let it predict Hi-C / Micro-C-style maps directly, and it beats Orca especially on cell-type differences, the differences that actually carry regulatory meaning.

The TAL1 case study is the one that sold me on the “interpret modalities jointly” claim: the model predicts variant-linked accessibility, expression and chromatin-mark changes consistent with neo-enhancer activation, and in-silico mutagenesis points at the motif-level positions driving it. That’s not just prediction, it’s a mechanistic hypothesis you could take to the bench.

The ablations are the honest bit

I always read the ablations first now, because they tell you whether the design choices are real or cosmetic. Two results stand out. Base-pair target resolution matters specifically for splicing and barely for coarse assays, so the “high resolution where needed” claim is earned, not decorative. And longer context helps: training and inference at 1 Mb is best overall. Removing modalities degrades many downstream variant tasks, which supports the whole multimodal premise, though the gains concentrate early (expression and accessibility) with diminishing returns as you pile on more.

Where it stops, and why that’s the interesting part for a PhD

The authors are refreshingly explicit about limits, and the limits map almost perfectly onto open problems I could work on:

• Distal regulation beyond ~100 kb is still difficult. A lot of cancer enhancer hijacking lives exactly in that hard zone.
• Tissue- and condition-specific effects aren’t fully captured, and cancer is a condition-specific rewiring of regulation.
• Training is still biased toward protein-coding genes, with thin non-coding-gene coverage.
• It’s benchmarked on reference genomes, not personal genomes, and it predicts molecular consequences, not phenotype or disease outcome.

That last point is the one I keep coming back to. AlphaGenome is a phenomenal hypothesis-generation engine, variant prioritisation plus mechanism, but the jump from “this variant changes accessibility” to “this variant drives this tumour” is still ours to make. That gap is where integrative work, conservation priors, and disease-specific data have to come in, and it’s a large part of how I think about my own objective: not replacing models like this, but wrapping them in the cancer-specific context they’re missing.

My one-line take

AlphaGenome narrows two trade-offs that have shaped regulatory genomics for a decade, and it does it convincingly enough that I now treat it as a baseline to beat rather than a tool to admire. The splice-junction head and the distillation strategy are the contributions I expect to age well. The limitations aren’t footnotes, they’re a roadmap, and a flattering amount of that roadmap overlaps with the questions I get to spend my PhD on.