In January 2026, DeepMind published AlphaGenome in Nature - a genomic AI model that can process up to one million bases of DNA in a single pass and predict gene expression, chromatin accessibility, histone modifications, transcription factor binding, and three-dimensional genome folding from sequence alone. The model's weights and code were released publicly on GitHub.

The genomics community took notice. But the prenatal diagnostics field - which stands to benefit enormously from this kind of capability - has been slower to engage with what AlphaGenome represents. Here's why it matters.

What AlphaGenome Actually Does

Most of the human genome - roughly 98% - doesn't code for proteins. For decades, this non-coding DNA was poorly understood, sometimes dismissed as "junk." We now know it contains regulatory elements that control when, where, and how much genes are expressed. Variants in these regulatory regions can cause disease just as surely as variants in protein-coding genes, but they are far harder to interpret.

AlphaGenome addresses this by learning the regulatory grammar of the genome. Given a stretch of DNA sequence, it predicts multiple layers of functional output simultaneously:

• Gene expression levels - How actively a gene is transcribed in different cell types
• Chromatin accessibility - Which regions of DNA are "open" and available for regulatory activity
• Histone modifications - Chemical marks on histone proteins that influence gene regulation
• Transcription factor binding - Where regulatory proteins attach to DNA
• 3D genome folding - How the linear DNA sequence folds in three-dimensional space, bringing distant regulatory elements into contact with their target genes

The model's one-million-base context window is significant. Many regulatory interactions happen over long genomic distances - an enhancer element can influence a gene hundreds of thousands of bases away. Previous models with shorter context windows simply couldn't capture these long-range effects.

Why This Matters for Variant Interpretation

The central challenge in clinical genetics is variant interpretation: given a DNA change found in a patient, does it cause disease or is it benign? For coding variants - changes in the protein-coding regions - we have reasonably good tools. Databases like ClinVar, predictive algorithms, and functional studies provide a framework for classification.

Non-coding variants are a different story. The American College of Medical Genetics (ACMG) guidelines for variant classification were primarily designed for coding variants. When a whole-genome sequencing test finds a variant in a regulatory region, the default classification is often "variant of uncertain significance" (VUS) - a frustrating non-answer for clinicians and patients alike.

Models like AlphaGenome offer a path toward resolving this. By predicting the functional impact of any sequence change on gene regulation, they can provide computational evidence for whether a non-coding variant is likely to be disruptive. This doesn't replace experimental validation, but it can prioritize which variants warrant further investigation and provide supporting evidence for clinical interpretation.

The Prenatal Context

Prenatal diagnostics has specific characteristics that make genomic AI particularly relevant:

Time pressure. When a prenatal test identifies a concerning variant, families and clinicians need answers quickly. A pregnancy doesn't pause while researchers debate whether a VUS is pathogenic. Computational tools that can rapidly assess variant significance are uniquely valuable in the prenatal setting.

Expanding screening scope. As NIPT moves beyond common trisomies into single-gene conditions and microdeletions, the number of variants that need interpretation grows dramatically. Natera's Fetal Focus sgNIPT already screens for 21 monogenic conditions. Future panels will be larger. The variant interpretation bottleneck will only get worse without AI-assisted tools.

Developmental gene regulation. Many prenatal conditions involve genes critical during embryonic development, where precise spatiotemporal gene regulation is essential. Non-coding variants affecting developmental enhancers or promoters may be particularly relevant in the prenatal context - and particularly suited to analysis by models trained on regulatory prediction.

De novo variants. A significant fraction of severe prenatal conditions arise from de novo (new) variants not present in either parent. These variants often lack database entries or prior literature, making computational prediction one of the few available tools for initial assessment.

The Gap Between Research and Clinical Use

It would be premature to suggest that AlphaGenome is ready for clinical prenatal diagnostics tomorrow. Several gaps need to be bridged:

Clinical validation. AlphaGenome was trained and evaluated on research datasets. Demonstrating that its predictions are reliable enough for clinical decision-making requires prospective validation studies, ideally across diverse populations.

Interpretability. Clinicians need to understand why a model makes a particular prediction, not just what the prediction is. New interpretability methods like TPCAV (a recent concept attribution approach for genomics models) are promising but still maturing.

Integration. Clinical genomics pipelines are complex ecosystems involving sequencing instruments, bioinformatics workflows, variant databases, and reporting systems. Incorporating a model like AlphaGenome requires not just running the model, but integrating its outputs into existing clinical workflows in a way that is actionable.

Training data for prenatal contexts. AlphaGenome was trained on general genomic data. Fine-tuning or evaluating the model specifically for cell types and developmental stages relevant to prenatal conditions could improve its utility. This is where synthetic genomic data - computationally generated training datasets representing prenatal-relevant scenarios - could play a role. At Eabha, we are working at this intersection of genomic AI and synthetic data generation, building tools that could help bridge the gap between foundational models like AlphaGenome and clinical prenatal applications.

What This Means Going Forward

AlphaGenome is not the end point - it's a milestone. The open release of the model's weights means that the broader community can build on it, fine-tune it for specific clinical applications, and benchmark alternative approaches against it.

For the prenatal diagnostics field specifically, the convergence of expanding NIPT panels, growing variant interpretation needs, and increasingly powerful genomic AI creates both an opportunity and an imperative. The labs and companies that figure out how to responsibly integrate these capabilities into clinical workflows will define the next era of prenatal screening.

The technology is arriving. The question now is whether the clinical, regulatory, and validation infrastructure can keep pace.