Why does the same DNA sequence behave differently in different cells? AlphaGenome 2 aims to answer this question.

Cardiomyocytes are responsible for contraction, while immune cells track down pathogens. They come from the same person and typically share virtually the same genome, yet they perform entirely different functions. The difference stems largely from how genes are utilized: which ones are activated, which ones remain silent, and to what extent they are activated.

This adds another layer of complexity to research on genetic diseases. Identifying a DNA mutation associated with a disease is far from enough. Researchers need to know in which types of cells it acts, which gene it affects, and how it alters gene activity.

AlphaGenome 2, released by Google DeepMind on November 11, 2026, aims to address these questions comprehensively. It incorporates cell type and cell state into its predictions, allowing the same segment of DNA to be analyzed separately in different biological contexts. Compared to the previous version of AlphaGenome, the key feature of this upgrade is that it shifts the focus from tissue averages to individual cells.Official Release Notes

Why Is Looking at DNA Sequences Alone Not Enough?

You can think of the genome as a large library. Different cells have similar collections, but the books they access on a daily basis vary. A change in a regulatory region may affect the content that one type of cell frequently reads, while having no noticeable effect on another type of cell.

What is often overlooked here are the two characters “调控” (regulation). Many genetic changes associated with traits and diseases do not directly alter the protein-coding sequence but instead occur in regions that control gene activity. Some regulatory regions are located far from the target gene, sometimes even separated by other genes. Therefore, searching for the gene closest in physical distance may not necessarily identify the gene that is actually affected.

The official release explains this using a research case involving intestinal immune cells: To understand a disease, it is necessary to examine both genetic variations and the affected genes and cells. Such an analysis is more specific than a general assertion that “a certain segment of DNA is associated with the disease,” but computational predictions still require validation through independent studies.Overview of Official Research

There is also a practical issue with tissue averages: significant changes in a small number of cells may become less noticeable when mixed in with a large number of other cells. For researchers, choosing the wrong scale of observation can lead to missing clues that are truly worth tracking.

Human blood viewed under a scanning electron microscope reveals blood cells and platelets of various shapes. This historical scientific photograph is used to illustrate the diversity of blood cells; it is not an output from AlphaGenome 2 nor a sample from this study. Mature red blood cells lack a nucleus and therefore cannot be used to directly illustrate “different reads of the same genome.” Photographs by Bruce Wetzel and Harry Schaefer / National Cancer Institute (NCI); public domain, source: Wikimedia Commons. Image Sources and Usage Guidelines.;Public Domain Mark

The new model asked one more question: "What kind of cell is this?"

AlphaGenome 2 uses both DNA sequences and biological conditions as inputs. These conditions can be derived from predefined tissues, cell types, or states, or extracted from the researcher’s own single-cell data. The model is thus able to make predictions based on the selected conditions, rather than relying solely on a fixed set of tissue-based outputs.

The general reader can understand it this way: In the past, it was more like examining the overall performance of many organizations; now, we can specify the object of observation and then ask how this segment of DNA functions within that object. This is a change in modeling approach; it does not mean that a complete virtual human body has been built into the software.

The model can analyze a very long segment of DNA, but its scope of analysis remains limited. It does not simultaneously account for all genetic and environmental factors in the human body, so a prediction based on a single segment of the genome cannot serve as a complete explanation of a disease.Input and Coordinate Instructions

“What Exactly Has ”Nearly Doubled”?

There is one finding in the research paper that is easily oversimplified by sensational headlines: in the TenK10K single-cell data evaluation, while maintaining a 90% precision rate, the recall rate for causally expressed variants increased from 16% in the previous-generation single-cell fine-tuning model to 30%; at a 90% accuracy rate for directional calls, the corresponding recall rate increased from 23% to 43%.AlphaGenome 2 Research Paper

Precision measures “how reliable the selected candidates are,” while recall measures “how many objects that meet the criteria are retrieved.” To illustrate with a simplified example: suppose there are 100 objects to be identified in the evaluation set; recall rates of 16% and 30% roughly correspond to retrieving 16 and 30 objects, respectively. This example is intended solely to explain the metrics and does not reflect the actual sample size used in the paper.

These results are particularly promising for screening experimental candidates. Researchers have limited budgets, cell samples, and time for experiments; identifying a larger number of candidates while maintaining a high level of reliability could alter the priority order for the next round of experiments.

However, there remains a significant gap between “detecting more” and “fully understanding everything.” This cannot be rephrased to mean that diagnostic accuracy has doubled, nor can it be inferred that disease risk can now be accurately predicted. Model evaluation addresses statistical performance within a specific task, whereas clinical issues also involve patient background, disease progression, and other evidence.

New cell statuses can be imported, but data quality remains important

A cell’s identity is not fixed forever. Gene activity changes depending on whether an immune cell is in a resting or activated state. When faced with new cohorts, new treatment conditions, or rare cellular states, researchers often prefer to use the data they have just measured rather than wait for a new model that covers that scenario.

According to official documentation, AlphaGenome 2 can describe cellular states using existing single-cell data, eliminating the need to retrain the model for each new dataset. While this broadens the scope of research, the quality of predictions remains dependent on data quality and model performance.Overview of Official Research

When interpreting these results, it is important to distinguish between phenomena already observed in the data and relationships inferred by the model. Avoiding retraining does not eliminate the need for independent validation; furthermore, findings derived from a limited sample cannot be directly generalized to the entire population.

Another noteworthy capability is the ability to infer signals that have not yet been measured. This can help researchers formulate hypotheses about what to measure next. However, the presence of a predicted curve in the interface does not mean that the corresponding experiment has already been conducted. The platform explicitly labels interpolated results that have not been validated against a baseline as exploratory results; readers must distinguish between predictions and measurements.Explanation of Push-Compensate Results

How Far Are We From Understanding the Disease?

The research paper also acknowledges that, even under high accuracy requirements, the new model still identifies only a subset of candidates; distant regulatory effects, certain post-transcriptional mechanisms, and diseases or transient cellular states not adequately covered by the training data remain challenges. Furthermore, the state relationships learned from observational data cannot automatically be equated with causal relationships observed in intervention experiments.Discussion Section of a Research Paper

These limitations affect the interpretation of the results. Even if a model does not predict a significant effect, this does not mean that a particular genetic variation is insignificant. The model may not yet account for the correct cellular states or may not be sensitive to the relevant mechanisms; the insights it provides still require independent evidence to support them.

From this perspective, AlphaGenome 2 is more likely to first transform pre-experimental preparations: by organizing a very long list of candidates into a more evidence-based priority list, allowing scientists to allocate their limited resources to questions that are more promising and easier to validate.

Who can use it, and what are the limitations?

AlphaGenome 2 is currently available for non-commercial research through its official API, while offering a private preview for commercial research. Access scope, quotas, and terms of use are specified in the official documentation; the term “open for research” should not be interpreted as implying unconditional or unlimited use.API Documentation

For the general reader, it is even more important to note whether subsequent studies have specified the data sources, comparison groups, and scope of validation. The success of individual cases cannot replace independent evaluations across different populations and diseases.

The scope of use must also be clearly defined. According to the official FAQ, direct API output and Atlas scores are limited to non-commercial research and may not be used for training machine learning models or clinical decision-making. At the time of AlphaGenome 2“s release, access for commercial research was in private preview; enterprises must go through official commercial channels, and ”free for research” should not be interpreted as permitting arbitrary commercial use.Usage Restrictions

AlphaGenome 2 delves deeper into a crucial question: exactly which cells and under what conditions a given genetic variation exerts its effects, and what those effects are. The truly promising future developments will depend on whether these predictions can be validated in more independent experiments and help researchers uncover mechanisms that were previously obscured by tissue-level averages.