AlphaGenome Atlas and the Grammar of Regulation

A model predicts measurements, not meanings

The foundation is AlphaGenome, introduced in June 2025 and described in a peer-reviewed Nature paper published on January 28, 2026. It accepts about one million DNA bases and predicts molecular readouts across several experimental modalities. These include gene expression, chromatin accessibility, and RNA splicing. Chromatin is DNA packaged with proteins; accessibility describes how available a region is to molecular machinery. Splicing removes selected stretches from an RNA transcript and joins the remaining pieces. A variant can influence these processes without changing a protein-coding sequence.

The technical achievement is combining long context with fine output resolution. Nearby letters can specify a short binding motif, while a distant regulatory element can influence a gene. A narrow window may recognize the motif but omit its wider surroundings. A coarse output may retain the surroundings while blurring a splice boundary. AlphaGenome addresses that tension in one framework. Its paper reports matching or exceeding the strongest external comparator in 25 of 26 variant-effect evaluations. This is a benchmark result across defined tasks, rather than a statement that every regulatory question is solved.

The useful counterfactual is deliberately small

A sequence-to-function model supports a straightforward comparison. Predict a molecular readout for a reference sequence, substitute a letter, predict again, and calculate the difference. The altered sequence is a computational counterfactual. It asks what the learned relationship between sequence and measurement implies when one input changes. The original AlphaGenome release explains this reference-versus-alternative approach. It does not require pretending that a cell has already been edited.

Consider an illustrative experiment in thought. Suppose two variants both reduce a predicted expression signal by one unit. One does so only in a liver-related readout; the other does so across several tissue readouts and changes a predicted splice boundary as well. The shared number would hide a substantial distinction. A researcher would want the complete pattern, the measurement scale, and the context in which the readout was learned. The thought experiment is not a reported result. It shows why preserving several output types can matter more than producing a single impressive-looking confidence value.

A ranking score is an index into a richer record

Atlas adds a prioritization layer called the AlphaGenome Variant Impact, or AVI, score. Its accompanying 2026 technical manuscript describes integrating predicted regulatory effects with additional information, including AlphaMissense protein-impact predictions, conservation measures, and coding features. The score therefore should not be interpreted as a pure expression-change meter. It is an aggregate constructed to help distinguish variants worth examining. Feature attributions then relate a score to contributions from different categories of information.

The distinction resembles a library catalogue. A relevance ranking can help locate a book without replacing the book. Here, the underlying molecular tracks and feature contributions are the material a researcher must inspect. A high score could arise for different biological reasons. Ranking two changes similarly does not establish that they share a mechanism, a tissue effect, or an experimental consequence. The technical manuscript is separate from the January Nature publication on the base model; readers should not transfer the earlier paper's peer-review status automatically to every later Atlas claim.

What precomputation changes in a working laboratory

The September release describes Atlas as an approximately one-petabyte resource with molecular predictions, an AVI score, feature attributions, and a collection of recurring sequence motifs. Making results searchable changes the cost of posing a question. A laboratory can inspect candidate variants before committing scarce assay capacity. It can also compare nearby positions rather than treating each unusual letter as an isolated mystery. The release reports collaborations in which prioritized predictions were followed by experimental checks; those checks are a distinct evidential step.

An independent way to appreciate the scale is simple arithmetic. A stretch of one thousand DNA positions permits three alternative letters per position, giving three thousand single-substitution questions. If an experimental team can examine only a modest subset, a ranking system can allocate attention. However, ranking also determines what remains invisible. A workflow that tests only the top predictions may repeatedly confirm a model's familiar patterns while missing systematic failures below its cutoff. Reserving some experimental capacity for surprising, low-ranked, and deliberately contrasting cases would produce more informative feedback.

A second illustrative comparison concerns selection. Suppose a screen can test twenty variants and the model supplies a ranked list of two hundred. Testing all twenty highest scores estimates success among preferred candidates. Testing fifteen preferred candidates and five sampled from other ranks additionally reveals something about the ranking itself. Neither design answers every question, but the latter can expose a blind spot that the former would never observe. Experimental allocation is therefore part of evaluating the atlas, not merely an administrative step after prediction.

Coordinates and context are part of the scientific claim

The public AlphaGenome software development kit exposes region coordinates, reference and alternate bases, requested outputs, and tissue-related ontology terms. Those are meaningful scientific inputs. An answer about one genomic interval and one tissue annotation is not automatically an answer about every cell. The repository also distinguishes querying precomputed Atlas results from requesting fresh model predictions. Access methods and usage terms need to be checked for the intended research setting.

Reproducibility begins before plotting. A defensible analysis should record the genome reference, coordinate convention, interval, alternative allele, selected outputs, model or resource version, and any score transformation. Imagine two collaborators disagreeing because one position was shifted by one base. Both graphs could look polished while describing different substitutions. A plain record of inputs would resolve the disagreement faster than an explanation about model intelligence. The same discipline matters when converting a continuous predicted track into a summary number: the aggregation window and normalization can change which variants look prominent.

The next advance is a better disagreement

My interpretation is that Atlas becomes most valuable when it makes competing explanations easy to state. If an expression change appears without a corresponding accessibility change, that pattern suggests a different follow-up from a prediction involving an altered splice junction. A model need not deliver a final causal account to improve the choice of experiment. It can sharpen the hypothesis sufficiently that the next measurement rules out more possibilities.

A strong follow-up would specify its variants and success criteria in advance, include reference and comparison controls, and measure the proposed molecular effect in a relevant biological setting. Failures would deserve as much attention as successes, especially when they cluster by tissue, sequence context, or variant type. The practical destination is a cycle of prediction, perturbation, measurement, and revision. AlphaGenome's documentation restricts these predictions to research and theoretical modeling rather than clinical decision-making. The scientific opportunity is substantial precisely because the atlas remains revisable: it gives researchers a large, structured set of propositions that the living system can test.