{"id":401602,"date":"2026-11-12T11:18:32","date_gmt":"2026-11-12T11:18:32","guid":{"rendered":"https:\/\/bizscoreai.com\/blog\/alphagenome-2-dna-variant-cell-types\/"},"modified":"2026-11-13T11:25:41","modified_gmt":"2026-11-13T11:25:41","slug":"alphagenome-2-dna-variant-cell-types","status":"publish","type":"post","link":"https:\/\/bizscoreai.com\/blog\/alphagenome-2-dna-variant-cell-types\/","title":{"rendered":"AlphaGenome 2: Predicting DNA Variant Effects in Specific Human Cell Types"},"content":{"rendered":"<div class=\"post-quick-answer\" style=\"background:rgba(128,128,128,.08);border-left:4px solid rgba(128,128,128,.6);padding:14px 18px;margin:0 0 24px\">\n<p class=\"post-quick-answer-label\"><strong>Quick Answer<\/strong><\/p>\n<p>AlphaGenome 2 predicts how DNA variants affect gene activity in specified human cell types and states, giving researchers a research tool for prioritizing variants and planning experiments. Its reported outputs include gene expression, chromatin accessibility, histone modifications and DNA methylation, but Google DeepMind states that it has not been validated or approved for clinical or diagnostic purposes.<\/p>\n<\/div>\n<p><!--\/post-quick-answer--><\/p>\n<p>The model is part of a wider push to connect human genetics with single-cell biology. Its announcement, <a href=\"https:\/\/deepmind.google\/blog\/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease\/\" target=\"_blank\" rel=\"noopener\">published by Google DeepMind on November 11, 2026<\/a>, reports stronger variant recall than earlier systems on the TenK10K dataset and describes experiments involving inflammatory bowel disease, COVID-19 vaccination and unseen single-cell datasets. The announcement also covers AlphaGenome Atlas, a separate September 2026 medRxiv preprint, so the paper, preprint and vendor-reported results need to be considered separately.<\/p>\n<nav class=\"post-contents\" style=\"border:1px solid rgba(128,128,128,.35);border-radius:6px;padding:14px 18px;margin:0 0 28px\">\n<p><strong>In This Article<\/strong><\/p>\n<ul>\n<li><a href=\"#what-this-means-and-what-it-does-not\">What This Means and What It Does Not<\/a><\/li>\n<li><a href=\"#what-does-alphagenome-2-predict\">What Does AlphaGenome 2 Predict?<\/a><\/li>\n<li><a href=\"#how-was-alphagenome-2-trained\">How Was AlphaGenome 2 Trained?<\/a><\/li>\n<li><a href=\"#what-performance-improvements-were-reported\">What Performance Improvements Were Reported?<\/a><\/li>\n<li><a href=\"#how-does-it-handle-new-cell-types\">How Does It Handle New Cell Types?<\/a><\/li>\n<li><a href=\"#cell-type-specific-epigenomic-prediction-2024\">Cell Type-Specific Epigenomic Prediction (2024)<\/a><\/li>\n<li><a href=\"#what-did-the-model-find-in-disease-studies\">What Did the Model Find in Disease Studies?<\/a><\/li>\n<li><a href=\"#how-does-alphagenome-atlas-fit\">How Does AlphaGenome Atlas Fit?<\/a><\/li>\n<li><a href=\"#who-can-access-alphagenome-2\">Who Can Access AlphaGenome 2?<\/a><\/li>\n<li><a href=\"#what-the-evidence-shows\">What the Evidence Shows<\/a><\/li>\n<li><a href=\"#key-considerations\">Key Considerations<\/a><\/li>\n<li><a href=\"#what-to-do-now\">What to Do Now<\/a><\/li>\n<li><a href=\"#faq\">FAQ<\/a><\/li>\n<\/ul>\n<\/nav>\n<h2 id=\"what-this-means-and-what-it-does-not\" data-toc=\"1\">What This Means and What It Does Not<\/h2>\n<ul>\n<li><strong>Research scope:<\/strong> AlphaGenome 2 is intended for research use only. Google DeepMind explicitly states that neither AlphaGenome nor AlphaGenome 2 has been validated or approved for clinical or diagnostic purposes.<\/li>\n<li><strong>Evidence status:<\/strong> The original AlphaGenome findings were published in <a href=\"https:\/\/pubmed.ncbi.nlm.nih.gov\/41606153\/\" target=\"_blank\" rel=\"noopener\">Nature in January 2026<\/a>. AlphaGenome 2 performance claims in this post come from Google DeepMind\u2019s November 2026 announcement and linked research materials. AlphaGenome Atlas findings come from a September 2026 <a href=\"https:\/\/pubmed.ncbi.nlm.nih.gov\/42779988\/\" target=\"_blank\" rel=\"noopener\">medRxiv preprint<\/a>, which has not been presented here as peer-reviewed.<\/li>\n<li><strong>Experimental validation:<\/strong> A model prediction identifies a candidate effect for research. It does not by itself establish that a variant causes a disease or that its predicted activity direction is correct in a particular person.<\/li>\n<li><strong>Comparison status:<\/strong> The benchmark percentages below are reported comparisons from the AlphaGenome 2 announcement. The announcement does not present them as independent validation of general clinical superiority.<\/li>\n<\/ul>\n<h2 id=\"what-does-alphagenome-2-predict\" data-toc=\"1\">What Does AlphaGenome 2 Predict?<\/h2>\n<p>AlphaGenome 2 takes a one-million-letter DNA sequence and a target biological condition, which can be a tissue, cell type or cell state. The condition may come from a reference catalog or be generated from a researcher\u2019s single-cell dataset through a cell encoder.<\/p>\n<p>Its conditional design replaces the original model\u2019s parallel predictions for broad tissues and cell lines with condition-specific outputs. This architecture is intended to distinguish how closely related cells read the same DNA sequence differently. The model can also use one measured type of cellular data to estimate missing measurements, a process called cross-modality imputation. For example, a single-cell gene activity profile can be used to predict chromatin accessibility, which describes how open or accessible a DNA region is to molecular machinery.<\/p>\n<p>Beyond gene expression, the amount of activity produced by a gene, AlphaGenome 2 predicts chromatin accessibility, histone modifications and DNA methylation. The original peer-reviewed AlphaGenome paper also covered transcription initiation, transcription factor binding, chromatin contact maps, splice site usage, and splice junction coordinates and strength.<\/p>\n<h2 id=\"how-was-alphagenome-2-trained\" data-toc=\"1\">How Was AlphaGenome 2 Trained?<\/h2>\n<p>According to <a href=\"https:\/\/deepmind.google\/blog\/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease\/\" target=\"_blank\" rel=\"noopener\">Google DeepMind\u2019s November 11, 2026 announcement<\/a>, AlphaGenome 2 was trained on bulk and single-cell measurements from human and mouse samples. The training data contained more than 107 million individual cell profiles across 68 tissues. The announcement says this data allowed the model to resolve activity in distinct heart muscle cells and blood immune cells rather than relying only on broad tissue averages.<\/p>\n<p>The earlier AlphaGenome study was published in <a href=\"https:\/\/pubmed.ncbi.nlm.nih.gov\/41606153\/\" target=\"_blank\" rel=\"noopener\">Nature in January 2026<\/a>. That peer-reviewed paper reported that the model matched or exceeded the strongest available external models in 25 of 26 evaluations of variant effect prediction. It also reported that AlphaGenome recapitulated mechanisms of clinically relevant variants near the TAL1 oncogene.<\/p>\n<h2 id=\"what-performance-improvements-were-reported\" data-toc=\"1\">What Performance Improvements Were Reported?<\/h2>\n<p><strong>Vendor-reported, not independently validated:<\/strong> On the TenK10K dataset, a single-cell map covering 28 primary immune cell types, AlphaGenome 2 identified 30% of causal variants. The comparison systems, including the original AlphaGenome fine-tuned on single-cell data and other specialized tools, identified 16%. Directional gene effects reached 43%, compared with 23% for the comparison systems. These figures come from Google DeepMind\u2019s <a href=\"https:\/\/deepmind.google\/blog\/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease\/\" target=\"_blank\" rel=\"noopener\">November 2026 announcement<\/a>.<\/p>\n<p><strong>Vendor-reported:<\/strong> On genome regions outside the training data, AlphaGenome 2 explained significantly more variation in gene expression between cell types within a tissue. The announcement also says the model tracked continuous gene-expression changes during development, including the maturation of human blood stem cells into red blood cells and immune lineages.<\/p>\n<p><strong>Vendor-reported reproduction:<\/strong> The model reproduced a previous finding that a variant in a distant regulatory switch drives inflammatory bowel disease by overactivating ETS2 in gut macrophages. According to the announcement, AlphaGenome 2 singled out the causal variant and selectively predicted increased ETS2 activity in gut macrophages, with other cell types and genes largely unaffected.<\/p>\n<h2 id=\"how-does-it-handle-new-cell-types\" data-toc=\"1\">How Does It Handle New Cell Types?<\/h2>\n<p><strong>Vendor-reported:<\/strong> AlphaGenome 2 can generalize to new single-cell experiments without retraining. The announced workflow accepts a single-cell RNA-sequencing snapshot from a few dozen cells, compresses its gene activity into a cell-state representation and uses that representation to condition predictions for an unseen state. Google DeepMind reports that the model can match or outperform models trained or fine-tuned directly on those datasets.<\/p>\n<p>In the announcement\u2019s COVID-19 mRNA vaccine example, AlphaGenome 2 distinguished vaccine-activated immune T cells from resting cells. It also reported that a variant linked to severe COVID-19 disrupted gene activity more strongly in activated cells, while the effect was missed in unactivated cells.<\/p>\n<h2 id=\"cell-type-specific-epigenomic-prediction-2024\" data-toc=\"1\">Cell Type-Specific Epigenomic Prediction (2024)<\/h2>\n<p>A <a href=\"https:\/\/www.nature.com\/articles\/s41467-024-54441-5\" target=\"_blank\" rel=\"noopener\">2024 Nature Communications study on cell type-specific epigenomic prediction<\/a> explains why cell-specific modeling matters for interpreting genetic variation. The study states that variants associated with regulatory elements and quantitative trait loci tend to have cell type-specific effects, even when they occur in non-coding regulatory regions that cannot be connected directly to functional outcomes.<\/p>\n<p>The same study identifies xQTL mapping as the most suitable current approach for this problem. This type of population study measures genomic variation among individuals and its correlation with regulatory elements. AlphaGenome 2 addresses a related modeling problem by predicting variant effects for a chosen cell condition, but its announced results remain distinct from the evidence described in the 2024 study.<\/p>\n<h2 id=\"what-did-the-model-find-in-disease-studies\" data-toc=\"1\">What Did the Model Find in Disease Studies?<\/h2>\n<p>Google DeepMind\u2019s precomputed AlphaGenome 2 scores cover trait-associated variants across hundreds of human cell and tissue conditions. The company says these scores can help identify candidate variants and cell types for traits ranging from height to type 2 diabetes. It also says the model can predict whether a variant is expected to raise or lower gene activity when sufficiently confident. These are research prioritization and interpretation outputs, not clinical test results.<\/p>\n<p>The model also extends enhancer-to-gene mapping into primary human cell types, including brain astrocytes. The announcement says this can help connect variants in distant regulatory regions, called enhancers, with their target genes in cells where disease may develop.<\/p>\n<h2 id=\"how-does-alphagenome-atlas-fit\" data-toc=\"1\">How Does AlphaGenome Atlas Fit?<\/h2>\n<p><strong>Preprint-only:<\/strong> <a href=\"https:\/\/pubmed.ncbi.nlm.nih.gov\/42779988\/\" target=\"_blank\" rel=\"noopener\">AlphaGenome Atlas, published as a medRxiv preprint on September 20, 2026<\/a>, extends AlphaGenome predictions across the human genome. The resource covers every possible single-nucleotide variant and many observed insertions and deletions, with a unified AlphaGenome Variant Impact score.<\/p>\n<p>The preprint reports state-of-the-art performance across diverse benchmarks and improved prioritization of deleterious non-coding variants. Its authors also report that the combined resource helped resolve an epileptic encephalopathy case, increased statistical power to detect rare non-coding variants associated with population-level traits, and improved the mechanistic interpretation of variants with genetic and clinical significance. These are preprint findings, not clinical validation.<\/p>\n<h2 id=\"who-can-access-alphagenome-2\" data-toc=\"1\">Who Can Access AlphaGenome 2?<\/h2>\n<p>Google DeepMind\u2019s <a href=\"https:\/\/deepmind.google\/blog\/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease\/\" target=\"_blank\" rel=\"noopener\">November 11, 2026 announcement<\/a> states that AlphaGenome 2 and interactive Colab tutorials for expression conditioning are available for non-commercial research through the AlphaGenome API. Commercial research is available in private preview. The announcement also says AlphaGenome 2 Skills are planned for Google Antigravity, its agentic development platform.<\/p>\n<p>Peter Donnelly, CEO of Genomics Limited, said in the announcement that human genetics is one of the most powerful tools in drug discovery and that GWAS is central to it. He also described the difficulty of moving from an association signal to a drug-development requirement and said DNA sequence prediction models can help bridge that gap.<\/p>\n<h2 id=\"what-the-evidence-shows\" data-toc=\"1\">What the Evidence Shows<\/h2>\n<ul>\n<li><strong>Peer-reviewed:<\/strong> The January 2026 <a href=\"https:\/\/pubmed.ncbi.nlm.nih.gov\/41606153\/\" target=\"_blank\" rel=\"noopener\">Nature paper<\/a> reported that the original AlphaGenome matched or exceeded external models in 25 of 26 variant-effect evaluations.<\/li>\n<li><strong>Vendor-reported:<\/strong> The November 2026 announcement reported 30% causal-variant identification and 43% directional gene-effect identification on TenK10K, compared with 16% and 23% for the comparison systems.<\/li>\n<li><strong>Preprint-only:<\/strong> The September 2026 <a href=\"https:\/\/pubmed.ncbi.nlm.nih.gov\/42779988\/\" target=\"_blank\" rel=\"noopener\">AlphaGenome Atlas preprint<\/a> reports improved prioritization of deleterious non-coding variants and contributions to a rare-disease case and population-level trait analyses.<\/li>\n<li><strong>Not independently validated:<\/strong> The announcement\u2019s COVID-19 vaccine, inflammatory bowel disease and new-cell-state examples are research results reported by Google DeepMind. The cited materials do not establish independent clinical validation.<\/li>\n<\/ul>\n<h2 id=\"key-considerations\" data-toc=\"1\">Key Considerations<\/h2>\n<ul>\n<li>Confirm that the relevant result comes from the peer-reviewed Nature paper, the AlphaGenome 2 announcement or the AlphaGenome Atlas preprint.<\/li>\n<li>Check the access route and usage terms before using AlphaGenome 2 in commercial research.<\/li>\n<li>Review licensing for each input dataset, variant annotation source and precomputed output.<\/li>\n<li>Compare model predictions with relevant wet-lab measurements from the cell type and condition under study.<\/li>\n<li>Record the model version, reference genome, cell-state input and dataset identifiers used in the analysis.<\/li>\n<li>Describe the outputs as research predictions rather than diagnostic or clinical evidence.<\/li>\n<\/ul>\n<h2 id=\"what-to-do-now\" data-toc=\"1\">What to Do Now<\/h2>\n<ol>\n<li>Review the non-commercial API terms and private-preview requirements for commercial work.<\/li>\n<li>Check the licenses attached to every dataset, genome reference, variant annotation and downloaded score.<\/li>\n<li>Confirm whether the supporting result is peer-reviewed, preprint-only or vendor-reported.<\/li>\n<li>Compare candidate variant predictions with wet-lab results from the relevant cell type and condition.<\/li>\n<li>Document the model version, input sequence, cell-state representation and source datasets.<\/li>\n<li>Label model outputs as research findings and exclude them from diagnostic or clinical evidence claims.<\/li>\n<\/ol>\n<h2 id=\"faq\" data-toc=\"1\">FAQ<\/h2>\n<h3>What Is AlphaGenome 2?<\/h3>\n<p>AlphaGenome 2 is a DNA sequence model that predicts how variants affect gene activity in a specified tissue, cell type or cell state. Google DeepMind announced it on November 11, 2026 for research use.<\/p>\n<h3>What Does AlphaGenome 2 Predict?<\/h3>\n<p>It predicts gene expression, histone modifications, chromatin accessibility and DNA methylation for a selected biological condition. It can also estimate missing measurements when one type of cellular measurement is supplied.<\/p>\n<h3>Can AlphaGenome 2 Work With New Cell Types?<\/h3>\n<p>Google DeepMind reports that AlphaGenome 2 can condition predictions on an unseen cell state using a single-cell dataset from a few dozen cells, without retraining the model. The announcement says it can match or outperform models trained or fine-tuned directly on those datasets.<\/p>\n<h3>Is AlphaGenome 2 Approved for Clinical or Diagnostic Use?<\/h3>\n<p>No. Google DeepMind states that AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes.<\/p>\n<h3>How Can Researchers Access AlphaGenome 2?<\/h3>\n<p>Google DeepMind says AlphaGenome 2 is available for non-commercial research through the AlphaGenome API. Commercial research is available in private preview.<\/p>\n<div class=\"source-disclaimer\"><em>This article reports on publicly available information from the sources listed below. We have no reason to believe it is inaccurate, but we have not independently verified every claim, and details may change as the story develops.<\/em><\/div>\n<section class=\"post-sources\">\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/deepmind.google\/blog\/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease\/\" target=\"_blank\" rel=\"noopener nofollow\">deepmind.google<\/a><\/li>\n<li><a href=\"https:\/\/pubmed.ncbi.nlm.nih.gov\/41606153\/\" target=\"_blank\" rel=\"noopener nofollow\">pubmed.ncbi.nlm.nih.gov<\/a><\/li>\n<li><a href=\"https:\/\/pubmed.ncbi.nlm.nih.gov\/42779988\/\" target=\"_blank\" rel=\"noopener nofollow\">pubmed.ncbi.nlm.nih.gov<\/a><\/li>\n<li><a href=\"https:\/\/www.nature.com\/articles\/s41467-024-54441-5\" target=\"_blank\" rel=\"noopener nofollow\">nature.com<\/a><\/li>\n<\/ul>\n<\/section>\n<p><script type=\"application\/ld+json\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"FAQPage\",\"mainEntity\":[{\"@type\":\"Question\",\"name\":\"What Is AlphaGenome 2?\",\"acceptedAnswer\":{\"@type\":\"Answer\",\"text\":\"AlphaGenome 2 is a DNA sequence model that predicts how variants affect gene activity in a specified tissue, cell type or cell state. Google DeepMind announced it on November 11, 2026 for research use.\"}},{\"@type\":\"Question\",\"name\":\"What Does AlphaGenome 2 Predict?\",\"acceptedAnswer\":{\"@type\":\"Answer\",\"text\":\"It predicts gene expression, histone modifications, chromatin accessibility and DNA methylation for a selected biological condition. It can also estimate missing measurements when one type of cellular measurement is supplied.\"}},{\"@type\":\"Question\",\"name\":\"Can AlphaGenome 2 Work With New Cell Types?\",\"acceptedAnswer\":{\"@type\":\"Answer\",\"text\":\"Google DeepMind reports that AlphaGenome 2 can condition predictions on an unseen cell state using a single-cell dataset from a few dozen cells, without retraining the model. The announcement says it can match or outperform models trained or fine-tuned directly on those datasets.\"}},{\"@type\":\"Question\",\"name\":\"Is AlphaGenome 2 Approved for Clinical or Diagnostic Use?\",\"acceptedAnswer\":{\"@type\":\"Answer\",\"text\":\"No. Google DeepMind states that AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes.\"}},{\"@type\":\"Question\",\"name\":\"How Can Researchers Access AlphaGenome 2?\",\"acceptedAnswer\":{\"@type\":\"Answer\",\"text\":\"Google DeepMind says AlphaGenome 2 is available for non-commercial research through the AlphaGenome API. Commercial research is available in private preview.\"}}]}]}<\/script><\/p>\n<hr style=\"margin:2.5em 0 1em;opacity:.35\" \/>\n<p style=\"font-size:.85em;opacity:.7\">See our <a href=\"https:\/\/bizscoreai.com\/blog\/disclaimer\/\">editorial disclaimer<\/a> for how our articles are produced.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>AlphaGenome 2 predicts how DNA variants affect gene activity in specific cell types, with research access available through the AlphaGenome API.<\/p>\n","protected":false},"author":1,"featured_media":401601,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"rank_math_title":"AlphaGenome 2: Predicting DNA Variant Effects","rank_math_description":"AlphaGenome 2 predicts DNA variant effects in specific cell types, with reported benchmarks, research access details and key evidence limitations.","rank_math_focus_keyword":"alphagenome 2","footnotes":""},"categories":[1],"tags":[],"class_list":["post-401602","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-news"],"elementor_data":null,"elementor_edit_mode":null,"_links":{"self":[{"href":"https:\/\/bizscoreai.com\/blog\/wp-json\/wp\/v2\/posts\/401602","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/bizscoreai.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/bizscoreai.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/bizscoreai.com\/blog\/wp-json\/wp\/v2\/comments?post=401602"}],"version-history":[{"count":2,"href":"https:\/\/bizscoreai.com\/blog\/wp-json\/wp\/v2\/posts\/401602\/revisions"}],"predecessor-version":[{"id":401619,"href":"https:\/\/bizscoreai.com\/blog\/wp-json\/wp\/v2\/posts\/401602\/revisions\/401619"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/bizscoreai.com\/blog\/wp-json\/wp\/v2\/media\/401601"}],"wp:attachment":[{"href":"https:\/\/bizscoreai.com\/blog\/wp-json\/wp\/v2\/media?parent=401602"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/bizscoreai.com\/blog\/wp-json\/wp\/v2\/categories?post=401602"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/bizscoreai.com\/blog\/wp-json\/wp\/v2\/tags?post=401602"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}