Google DeepMind Releases AlphaGenome 2 for Cell-Type-Specific Variant Effects

Google DeepMind’s latest drop, AlphaGenome 2, is a classic case of a hyperscaler throwing a shiny new API at the bio‑informatics community while keeping the real meat— the model weights— locked behind a non‑commercial curtain. As someone who runs a lean hosting operation and has spent a decade wrestling with production‑grade DNA models, I see both the promise and the pitfalls. The model reads a full megabase of DNA, conditions on a 256‑dimensional cell‑state embedding, and spits out a laundry list of functional readouts—from RNA coverage to CpG methylation. On paper it beats its predecessor and a few open‑source rivals, but the delivery model and licensing raise eyebrows for any independent provider looking to offer this as a service.

What AlphaGenome 2 actually does

At its core AlphaGenome 2 is a U‑Net style architecture: a convolutional encoder, an 11‑block transformer, and a decoder. The key innovation is the conditional embedding that represents a cell state. This 256‑dimensional vector is injected via FiLM layers throughout the encoder and transformer, which the DeepMind team found to be the main driver of performance gains. Conditioning only the decoder or output layers barely moved the needle, so the heavy lifting happens early in the network.

The model accepts a 1 Mb DNA window—exactly 1,048,576 base pairs—plus the cell‑state embedding, and returns a suite of functional predictions: RNA coverage, CAGE, splicing, chromatin accessibility, histone marks, transcription‑factor binding, contact maps, and, newly, CpG methylation. Unlike AlphaGenome 1, which forced users to pick from a fixed catalog of thousands of tracks, AlphaGenome 2 delivers one condition per call, meaning you can ask for any cell type on the fly, provided you can generate the embedding.

Training pedigree and data diet

The training regime was a four‑stage marathon on 2,048 TPU v4 chips. First, a bulk‑data pre‑training round was followed by ingestion of 3,977 single‑cell RNA and ATAC pseudobulk tracks from 36 studies. Next came a condition‑refinement stage that tuned only the embeddings and FiLM layers. Finally, DeepMind distilled eight teacher models into a single student using mutated input sequences and added a post‑training step that maps a massive scRNA‑seq encoder—trained on over 100 million human profiles from CELLxGENE—into the frozen model’s condition space. The API docs let you feed up to 256 cells per call, turning raw counts into the required embedding.

This pipeline is impressive in scale, but the reliance on proprietary TPU infrastructure and the lack of disclosed parameter counts (the paper only mentions 235 M for AlphaGenome 1) make it hard for smaller outfits to reproduce or audit the process. The model’s weights are not released; they’ll come later, and only under a non‑commercial license, which is a red flag for anyone hoping to host the model locally.

Benchmark bragging rights

DeepMind’s own numbers show AlphaGenome 2 pulling ahead on several fronts. On GTEx tissues the median Pearson correlation for cross‑tissue expression changes rose from 0.61 to 0.71, and within tissues the explained cell‑type variance jumped from 23 % to 34 % for scRNA‑seq. The headline result is on the TenK10K single‑cell eQTL benchmark across 28 immune cell types: at 90 % precision the model recovered 30 % of causal variants, versus 16 % for AlphaGenome 1 fine‑tuned on the same data. Direction‑of‑effect recall climbed from 23 % to 43 %.

On held‑out ENCODE‑rE2G CRISPRi data, enhancer‑gene AUPRC hit 0.78, outpacing AlphaGenome 1’s 0.73 and a Borzoi ensemble’s 0.61. The advantage widened for distal enhancers (100 kb to 2,500 kb away), where AlphaGenome 2 scored 0.62 against 0.48. For assay imputation, it beat the Avocado baseline on 166 of 181 held‑out tracks. In an unseen Parkinson’s cohort, recall reached 41 % versus 17 % for a cohort‑tuned Scooby model. All these figures come from DeepMind’s own evaluation, so they should be taken as a baseline rather than an industry‑wide gold standard.

Access model: API‑only and non‑commercial

AlphaGenome 2 is currently only reachable via a hosted API. An API key is free for non‑commercial research, and DeepMind has published Colab tutorials and a Python package (alphagenome.models.v2) to make it easy to call. Commercial research is in a private preview, and the weights are slated for release “later” under a non‑commercial license. This means you cannot download the model to run on your own GPU farm, nor can you embed it in a SaaS offering without a bespoke agreement.

For independent hosting providers, this is a deal‑breaker. Running a high‑throughput genomics service on a third‑party API adds latency, cost per call, and a dependency on Google’s uptime. It also forces you into Google’s pricing model, which historically has been opaque and prone to price spikes for heavy usage. If you’re already paying for GPU instances, you might be better off sticking with open‑source alternatives that you can containerise and scale behind your own load balancers.

Business‑risk lens for founders

From a risk perspective, AlphaGenome 2’s API‑only stance introduces vendor lock‑in. You’re betting that Google will keep the service stable, affordable, and that the licensing terms won’t change. The non‑commercial restriction also blocks any direct monetisation of the model’s predictions, forcing you to build value elsewhere—perhaps by adding downstream analytics, visualisations, or integration with clinical pipelines.

Moreover, the model’s input size (1 Mb) and per‑call embedding generation (up to 256 cells) imply substantial compute per request. If you’re serving thousands of variants per day, the API costs could balloon quickly. Without transparent pricing, you’re flying blind. In contrast, open‑source models like Avocado or Borzoi can be run on commodity GPUs, giving you full control over cost and scaling.

Practical steps for independent teams

If you want to experiment with AlphaGenome 2, start with the free non‑commercial API key and the provided Colab notebooks. Use the precomputed atlases—one covering enhancer‑gene links across 640 primary cell conditions and another scoring about 7 million common and GWAS variants across 510 conditions—to benchmark your own pipelines without incurring API calls. This can give you a sense of the model’s predictive power before you decide whether to pursue a commercial partnership.

At the same time, keep an eye on open‑source alternatives. The benchmark tables show that AlphaGenome 2 outperforms Borzoi and Scooby on several tasks, but the margins are not astronomical. A well‑tuned open‑source model on your own hardware can achieve comparable results for many use‑cases, especially if you can tailor the training data to your domain.

Bottom line and actionable advice

AlphaGenome 2 is a technical tour de force: a 1 Mb conditional DNA model that lifts single‑cell eQTL recall from the mid‑teens to thirty percent at high precision, and pushes enhancer‑gene prediction into the high‑70s AUPRC range. Yet the delivery via a closed API and the non‑commercial licensing model make it a risky proposition for independent hosting providers and biotech startups looking to build scalable services.

My advice: use the free API to explore the model’s capabilities, but plan to either negotiate a commercial agreement with Google early or fall back on open‑source stacks you can run in‑house. Build your value‑add around downstream interpretation, integration, or visualisation, not around the raw model predictions. And always keep a close eye on the cost per call—hidden fees can eat into margins faster than a DDoS attack on a small data centre. In a world where hyperscalers love to tout performance but hide the price tag, staying pragmatic and keeping control of your stack is the only way to survive.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Reporting is based on the source material cited below. Sources: MarkTechPost; marktechpost.com; Global1.News (12 November 2026).

By Allan Ali, Global1.News