Skin, Decoded
← Contents

Straight answer

Does AlphaGenome 2 predict anything about your skin?

Answered in plain English, with sources

No — not about your skin, and not about you. AlphaGenome 2 predicts molecular signals from a DNA sequence in a cell type you choose. It has no idea what you eat, what you put on your face, or whether you have eczema.

AlphaGenome 2 is a model of how DNA is read, not a model of skin. If you hand it a stretch of DNA and a cell type, it returns predicted molecular activity: how much a nearby gene is expressed, whether that region is open or closed, how the transcript is spliced, and how methylated it is. That is the whole output.

It does not take a photograph, a routine, a product, a symptom or a question. It cannot tell you whether a cream will work for you, whether your skin is “genetically old”, or what you should buy. Those are questions about a person in an environment. AlphaGenome 2 is a question about a sequence in a dish.

What it actually predicts

InputA DNA sequence, plus one cell type or tissue
OutputPredicted gene expression, chromatin accessibility, splicing, methylation and 3D contacts
UnitOne variant at a time, one cell type at a time
Not includedProducts, routines, lifestyle, age, symptoms, diagnosis

The value is in regulatory variants — changes in the 98% of the genome that does not code for protein. Those variants often sit hundreds of thousands of letters from the gene they control, and their effect depends entirely on which cell type is reading the DNA. That is why the cell type is an input and not an afterthought.

Why “your skin” is a harder question than it sounds

Skin is not one tissue. A keratinocyte in the outer layer, a melanocyte producing pigment, a fibroblast building collagen and a macrophage patrolling for damage all read the same genome differently. A prediction is only meaningful for the cell type you asked about. Ask about the wrong one and the answer is confidently irrelevant.

The honest limit

A variant can be predicted to change expression in one cell type and do nothing in another. AlphaGenome 2 reports what it predicts for the cell type you selected. It does not tell you which cell type in your skin matters for your concern — that is a biology question, and often an open one.

What would have to be true for it to say something about skin

01

A variant worth asking about

Not a whole genome, not a consumer test's summary. One change, whose effect is unknown.

02

The right cell type

The prediction is conditional on it. Skin has many, and the relevant one differs by question.

03

A measured effect, then validation

A prediction is a hypothesis. It becomes knowledge when an experiment in real cells agrees with it.

What it must not be used for

  • Deciding whether a product will work on your skin
  • Diagnosing anything, or replacing a dermatologist
  • Estimating personal risk of a disease from a consumer DNA file
  • Any commercial product that calls the model on a customer's behalf — the free API licence forbids it

DeepMind states plainly that the model is a research tool, not validated or approved for clinical or diagnostic use, and that outputs must not be used to train other machine-learning models. That is not fine print to work around; it is the boundary that makes the work publishable.

So what is it good for?

Finding the mechanism behind a genetic signal. The clearest example is a region on chromosome 21 associated with inflammatory bowel disease that sits in a gene desert — no gene nearby at all. AlphaGenome pointed at an enhancer that acts on ETS2 in macrophages, a result later confirmed experimentally. That is the shape of a real contribution: a region with no obvious explanation, a cell type, and a mechanism that can be tested.

Bottom line

AlphaGenome 2 can tell a researcher which cell type a variant probably acts in and what it probably does there. It cannot tell you anything about your skin. Any product claiming otherwise is using the model's reputation, not its output.

Read next

Sources

AlphaGenome 2: Deciphering how cells read DNA to better understand diseaseGoogle DeepMind, 11 November 2026

Announcement, architecture, benchmark figures and limitations.

AlphaGenome 2 preprintbioRxiv, November 2026 · doi:10.64898/2026.11.11.765857

Not yet peer-reviewed.

AlphaGenome Terms of Service and documentationGoogle DeepMind

Licensing: non-commercial research use, no clinical decision-making, no training other models.

A disease-associated gene desert directs macrophage inflammation through ETS2Stankey C.T. et al., Nature 630, 447–456 (2024)

The chr21q22 / ETS2 / IBD finding reproduced by AlphaGenome 2.

Not medical advice. General education only. If you have a skin concern, see a dermatologist. AlphaGenome and AlphaGenome 2 are research tools, not validated for clinical use.