AlphaGenome 2 Explained: What It Is, How It Works, Limits and Risks

Last updated: 12 November 2026

On 11 November 2026, Google DeepMind released AlphaGenome 2, the second version of its AI model for reading the non-coding parts of our DNA. The first AlphaGenome, announced in June 2025, could take in a million letters of DNA and predict things like how strongly nearby genes would be switched on in a tissue. Version 2 adds the piece biologists kept asking for: it makes those predictions for a specific cell type or cell state, and it can be steered using a researcher’s own single-cell data.

That sounds like a narrow technical upgrade, but it goes to the heart of a hard problem. Most DNA variants linked to diseases sit outside genes, in regulatory “switches”, and a switch that matters in an immune T cell may do nothing in a liver cell. DeepMind says the new model finds nearly twice as many expression-changing variants as its earlier best setup on a large immune-cell dataset. Outside scientists quoted in DeepMind’s announcement were impressed, but the fine print matters: the model still misses most causal variants at high confidence, the technical paper is not yet peer reviewed, and it is strictly for research, not diagnosis.

In this post I explain what AlphaGenome 2 is, how it works, what the headline numbers really measure, how students can try it, where it falls short, and why I think it is one of the best case studies this year for Indian data science and GATE DA students who want to see machine learning applied to real science.

Quick summary

• What it is: A deep learning model from Google DeepMind that reads up to about 1 million DNA base pairs and predicts regulatory activity (gene expression, chromatin accessibility, histone marks, splicing and, new in v2, DNA methylation) in one chosen biological condition at a time.
• Released: 11 November 2026, via the DeepMind blog and an accompanying technical paper.
• Key new idea: “Conditioning”. Instead of fixed outputs for a list of tissues, you tell the model which tissue, cell type or cell state you care about, or describe it with your own single-cell RNA-seq data.
• Headline result: On the TenK10K immune-cell dataset (28 cell types), recall of causal variants at 90% precision rose from 16% to 30% compared with the original AlphaGenome fine-tuned on single-cell data.
• Access: Free through the AlphaGenome API for non-commercial research; commercial use is in private preview. No public weights for v2 have been announced.
• Reality check: Research use only, not validated for clinical or diagnostic use. Roughly 70% of causal variants are still not recovered at that high-precision threshold.

What is AlphaGenome 2, in plain language?

Only about 2% of the human genome codes for proteins. The rest includes vast stretches that control when, where and how much each gene is used. Think of the genes as songs and the regulatory DNA as the mixing desk: the same song can be played loudly in one cell and muted in another.

When large genetic studies (GWAS) link a region of DNA to a disease such as type 2 diabetes or inflammatory bowel disease, the hit usually lands in this regulatory territory and often contains dozens of correlated variants. Figuring out which single letter change actually causes the effect, in which cell type, is slow and expensive lab work.

AlphaGenome models try to shortcut part of that search. You give the model a DNA sequence, optionally with one letter changed, and it predicts the molecular readouts you would expect to measure in the lab. Compare the predictions for the reference and the variant sequence, and you get an estimate of the variant’s effect. AlphaGenome 2 makes that estimate specific to a cell type or state, which is where much of the biology actually happens.

How AlphaGenome 2 works

DeepMind’s paper describes a design that keeps the overall shape of the first model but rebuilds how it handles biological context.

1. Long DNA input. The model reads a sequence of roughly one million letters, so it can “see” regulatory elements far away from the gene they control.
2. U-Net-style backbone. Like AlphaGenome 1, it uses convolutional layers to detect short DNA patterns, transformer layers to relate distant parts of the sequence, and a decoder with skip connections to produce outputs at fine resolution.
3. Condition embedding. The big change. A chosen condition (for example “CD8 T cell” or “activated T cell”) is turned into a vector, and that vector modulates the network through adaptive normalisation layers at every stage, rather than being bolted on at the end.
4. Cell encoder. You can create a condition from your own single-cell RNA-seq data. DeepMind says a snapshot of a few dozen cells can be enough. The API documentation describes expression conditioning with up to 256 cells per request.
5. Multiple readouts at different resolutions. Chromatin accessibility, DNA methylation, RNA coverage, CAGE and splicing are predicted at single base-pair resolution; transcription factor binding and histone marks at 128 bp; and 3D DNA contact maps at 2,048 bp.

Because the condition is an input, the model can also do cross-modality imputation: given one kind of measurement for a cell population (say, gene activity), it can predict readouts nobody measured, such as chromatin accessibility. In the paper, imputed RNA tracks reached a correlation of 0.74 against 0.58 for Avocado, an established imputation method, and AlphaGenome 2 beat Avocado on 166 of 181 tracks.

Training data. The paper says the cell encoder was trained on over 100 million single-cell profiles across 68 tissues, and that post-training used 107 million profiles. (Some news reports quote one number and some the other; both appear in the paper for different stages.) It names public resources such as ENCODE, GTEx and CELLxGENE, adds 3,977 new single-cell pseudobulk tracks from 36 studies, and covers both human and mouse. DNA methylation is predicted across 185 human reference conditions.

What AlphaGenome 2 can do: the reported results

All numbers below come from DeepMind’s own paper and blog. They have not yet been independently replicated.

The comparison model in the paper is the original AlphaGenome fine-tuned on single-cell data. DeepMind’s blog loosely mentions “other specialized tools” as well, without naming them.

Beyond benchmarks, DeepMind highlights several case studies:

• Inflammatory bowel disease: A 2024 Nature study traced an IBD risk region to a regulatory switch for the ETS2 gene in immune cells. AlphaGenome 2 reproduced that finding computationally and singled out a candidate causal variant.
• Vaccine response: Using single-cell data from people given COVID-19 mRNA vaccines, the model separated vaccine-activated T cells from resting ones.
• Blood cell development: It tracked how predicted regulatory activity changes as red blood cells mature.
• Ready-made resources: DeepMind released precomputed enhancer-to-gene links across 640 conditions and variant scores for about 7 million GWAS variants across 510 conditions, with examples for traits like height and type 2 diabetes.

Early reactions from outside DeepMind were positive. Professor Sarah Teichmann, a co-founder of the Human Cell Atlas, praised “how it adapts directly to new single-cell data without retraining”. Peter Donnelly, CEO of Genomics plc, said the company plans to use it in its Mystra platform. Both comments were published as part of DeepMind’s announcement, so they are endorsements rather than independent tests.

How to try AlphaGenome 2 (official method)

AlphaGenome 2 runs on Google’s servers through the AlphaGenome API. According to the official documentation:

1. Request a free API key for non-commercial use at alphagenome.google/api and read the terms.
2. Install the Python client and set your key as an environment variable (or as a Colab secret named ALPHA_GENOME_API_KEY).
3. Use the AlphaGenome 2 client, which the docs place under alphagenome.models.v2, and follow the official Colab tutorials, including the one on conditioning with single-cell data.

# In a virtual environment
pip install alphagenome

# Store your API key (never commit it to GitHub)
export ALPHA_GENOME_API_KEY="YOUR_API_KEY"

# In Python, the v2 client is created roughly like this (see the official docs for full examples):
# from alphagenome.models.v2 import dna_client
# dna_model = dna_client.create(API_KEY)

A few practical notes. DeepMind says query rates “vary with demand”, and it has not published fixed quotas or prices. Commercial users must join a private preview. DeepMind is running a feedback form until 30 November 2026, and says selected projects may get extra quota and credits. It also says AlphaGenome 2 “Skills” for its Antigravity agent platform are coming.

If you want to run a model locally, the separate alphagenome_research repository offers code and downloadable weights (via Kaggle and Hugging Face) under non-commercial terms. Those weights were released with the original AlphaGenome’s Nature paper in January 2026, and DeepMind recommends at least an NVIDIA H100 GPU. As of today, DeepMind has not announced public weights for AlphaGenome 2, so treat the API as the way to use the new model.

Limits, criticism and what the numbers do not say

This is a strong research result, but it is easy to over-read. Here is what I would keep in mind.

• It still misses most causal variants at high confidence. A recall of 30% at 90% precision means about seven in ten known causal variants are not flagged at that threshold. The paper itself says the model recovers only a minority of causal variants at high precision.
• Several news reports mis-describe the metric. Some articles say the model’s “accuracy” rose from 23% to 43%. In the paper, 43% is a recall figure at 90% sign accuracy, a different and more demanding measure. If you cite these numbers in a seminar, use the paper’s wording.
• Known blind spots. The paper notes the model tends to underestimate effects outside open (accessible) chromatin, at distant regulatory elements, and for effects that happen after transcription. DeepMind listed very long-range regulation as a challenge for version 1 too.
• Training data gaps. Short-lived, stimulated and disease cell states are under-represented in public datasets, so predictions for those states deserve extra caution.
• Correlation, not causation. The model learns from observational data. Predictions for gene knockouts or over-expression were not tested, and lab validation is still essential.
• More compute. Because each prediction is tied to one condition, scanning many cell types costs more. TipRanks, citing The Verge, reported that DeepMind said AlphaGenome 2 needs more computing than version 1.
• Not peer reviewed yet. The technical paper is a DeepMind PDF. The first AlphaGenome went through peer review and was published in Nature in January 2026; version 2 has not reached that stage.
• Closed model, limited terms. Commentators such as i-SCOOP point out that access terms matter as much as the model: free use is non-commercial, commercial pricing is not public, and reported terms restrict using outputs to train other models or to inform clinical decisions. I could not load the full terms page, so check the official text yourself.

Risks, privacy and the Indian legal angle

This is general information, not legal advice.

The most immediate risk is misuse as a diagnostic tool. DeepMind states plainly that AlphaGenome and AlphaGenome 2 are for research only and have not been validated or approved for clinical or diagnostic purposes. A student, startup or clinic that tells a patient “the AI says this variant causes your disease” would be going well beyond what the model supports.

The second issue is data protection. The cell-conditioning feature invites researchers to send their own single-cell data to an API. Single-cell and genetic data from human volunteers is personal data. In India, the Digital Personal Data Protection Act, 2023, together with the DPDP Rules notified in November 2025 and being phased in, governs how such data can be processed, and research involving human participants is also expected to follow ICMR’s ethical guidelines and institutional ethics committee approval. Before uploading any patient-derived data to a foreign cloud service, check consent forms, ethics approval and your institution’s data policy. Using public, de-identified datasets is the safe default for learning.

The third issue is population fairness. Predictions are only as broad as the training data. Indian populations have historically been under-represented in large genomic resources, which is one reason projects like GenomeIndia, whose data from 10,000 genomes was made available through the Indian Biological Data Centre in January 2025, matter. Researchers should test whether AlphaGenome-style predictions hold up on variants common in Indian populations rather than assume they transfer.

Why this matters for Indian students and educators

AlphaGenome 2 is a near-perfect example of modern machine learning meeting a real scientific problem, and nearly every idea in it maps to something in the syllabus.

• GATE DA: Machine Learning and Deep Learning. Convolutional networks, transformers, U-Net encoder-decoder designs, skip connections and conditioning layers (adaptive normalisation, FiLM) are all here, applied to a sequence of a million tokens.
• GATE DA: Probability and Statistics. Precision, recall, “recall at 90% precision”, correlation coefficients and the difference between correlation and causation are exactly what you need to read this paper honestly.
• Linear Algebra. Embeddings for cell conditions are vectors in a learned space; understanding them is understanding the model.
• Data engineering and DBMS. Working with AnnData objects, genome coordinates and massive public datasets is a serious data-handling skill.
• AI ethics. Research-only limits, consent and population bias are live debates, not textbook abstractions.

On careers: computational biology and bioinformatics roles are growing at Indian research institutes, pharma and biotech companies, and genomics startups, and they reward people who combine solid ML with enough biology to ask good questions. You do not need a biology degree to start. A practical path is to learn basic molecular biology (DNA, RNA, transcription, gene regulation), work through the official AlphaGenome Colab tutorials with a free non-commercial key, and then reproduce a small analysis, such as scoring a published variant, and compare it with the original study. Educators can use the benchmark table above as a ready-made class exercise on evaluation metrics.

Frequently asked questions

Is AlphaGenome 2 free?

For non-commercial research, yes, through the AlphaGenome API with a free key. Commercial research use is in private preview, and DeepMind has not published pricing.

Is AlphaGenome 2 open source?

No. The client library and the research code for the original model are public, and the original model’s weights can be downloaded under non-commercial terms, but DeepMind has not announced public weights for AlphaGenome 2.

How is AlphaGenome 2 different from AlphaGenome 1?

Version 1 predicted a fixed set of outputs for bulk tissues and cell lines. Version 2 takes a chosen condition as input, can be conditioned on your own single-cell data, adds DNA methylation predictions, and performs better on DeepMind’s single-cell variant benchmarks.

Can a doctor use AlphaGenome 2 to diagnose a genetic disease?

No. DeepMind says the models are not validated or approved for clinical or diagnostic purposes. They are tools to prioritise hypotheses for laboratory testing.

Does it work for species other than humans?

The model is trained on human and mouse data. It is not designed for plants or other animals.

Do I need a GPU to use it?

Not for the API, which runs on Google’s servers. Running the downloadable research model locally is a different matter: DeepMind recommends at least an NVIDIA H100.

The bottom line

AlphaGenome 2 is a meaningful step: it moves DNA sequence models from “what does this variant do in a tissue?” to “what does it do in this particular cell, right now?”, and it lets researchers bring their own single-cell data to the question. The reported gains on immune-cell benchmarks are large. But it still finds a minority of causal variants at high confidence, its paper awaits peer review, and it is a research tool, not a diagnostic test. For Indian students, the bigger lesson is that the most exciting AI work is increasingly happening at the meeting point of machine learning and another discipline. The fundamentals you study for GATE DA are exactly the ones you need to understand, use and critique a model like this.

Sources and further reading

• Google DeepMind blog: AlphaGenome 2, deciphering how cells read DNA
• Technical paper: AlphaGenome 2 predicts regulatory activity and variant effects across cell states (PDF)
• AlphaGenome API documentation
• GitHub: google-deepmind/alphagenome_research
• Google DeepMind blog: the original AlphaGenome (June 2025)
• Nature: the AlphaGenome paper (2026)
• Nature: the ETS2 inflammatory disease study (2024)
• i-SCOOP: AlphaGenome 2 and why its access terms matter
• Interesting Engineering: AlphaGenome 2 finds nearly 2x more disease-linked variants
• Unite.AI: AlphaGenome 2 predicts cell-state variant effects
• SiliconANGLE: DeepMind releases new robotics and gene prediction models
• TipRanks: Google’s new AI model identifies nearly twice as many disease-linked DNA changes
• Gigazine: AlphaGenome published in Nature and weights released (January 2026)
• ThePrint: GenomeIndia data made publicly available
• India Briefing: DPDP Rules 2025 notified

Written by Piyush Wairale (M.Tech, IIT Madras), founder of PiyushAI Edtech. Benchmark figures are as reported by Google DeepMind and the publications above on or before 12 November 2026 and have not been independently replicated.