AlphaGenome Grammar Discovery Agent
Can past results guide the next DNA experiment?
We built an agent that chooses pairs of DNA edits. AlphaGenome predicts their effect on DNA accessibility: how exposed DNA is to proteins. We ask whether earlier results help the agent find combinations whose effects do not simply add up.
This worked example follows TCF7, one of the twelve regions in our Jev study.
Change the highlighted base to inspect one edit.
Reference sequence: mean predicted accessibility 1.419.
T-cell DNASE prediction. The dotted line is the average over the same 1,024 bases.
The research question is whether a separate decision agent can discover strong pair interactions within fifteen pair reveals, compared with random selection and a fixed single-effect ranking. AlphaGenome predicts molecular signals; it does not choose experiments. Every example in this audience path now uses the actual TCF7 region and Jev 1.14.0 full-feedback recording from the expanded study. This example was selected after the study to explain the workflow; it is not representative evidence. Change E05 G to A and back. Reference mean is 1.4192895889282227; E05 is 1.1880651712417603. AlphaGenome input is 16,384 bases; the readout is mean DNASE over the central 1,024 bases across exact CL:0000084 T-cell tracks. Each displayed bin averages 32 real bases. Calculations use full precision; display values are rounded. These are saved model predictions, not cell measurements.
The pair is not fully explained by the two edits.
E05 and E06 each lower predicted accessibility. Adding those two reductions gives an expectation of 0.865. AlphaGenome predicts 0.808 for the pair: a further reduction of 0.057. We call that difference the interaction.
The separate effects define what simple addition would predict.
We search for large interactions in either direction, on this raw prediction scale.
Continue the same TCF7 example. Reference is 1.4192895889282227; E05 alone 1.1880651712417603; E06 alone 1.0960206985473633. The two changes are -0.2312244176864624 and -0.3232688903808594. Starting at reference and adding those changes gives 0.8647962808609009. The actual saved pair predicts 0.8077409267425537, so interaction is -0.05705535411834717. The pair lowers accessibility more than the additive expectation. The comparison stays on one zero-based scale. Interaction is F(AB)-F(A)-F(B)+F(reference); we rank its absolute magnitude. No combined profile is invented. Log1p sensitivity is separate because it changes the interaction definition.
Now the problem is choosing which pairs to test.
Ten candidate edits make 45 pairs. For each region, we predicted every pair to build an answer key, then let each chooser reveal only 15 pair results. A target is one of the five largest interactions in either direction; the score is how many of those five it finds.
Finding two of the five targets is 40% recall. A tested pair uses budget even when it is outside those five.
This measures selection within a hidden prediction landscape. Building the benchmark used all 45 pairs.
Connect the interaction just defined with the search objective. The evaluator knows the five largest absolute interactions among all forty-five pairs. Each chooser starts with the same reference and ten single-edit predictions, but can reveal only fifteen unique pair outcomes. A revealed pair counts as a discovery only when it belongs to that hidden top five. Recall is targets found divided by five. Preparing the complete benchmark costs 56 AlphaGenome outputs per region: reference, ten singles and forty-five doubles. Fifteen reveals measures selection efficiency, not physical API savings in the completed exhaustive benchmark. The twelve-region study contains sixty evaluator targets.
Jev chooses again after each recorded result.
Jev sees the known single-edit effects and its earlier pair results, then chooses one untested pair. The application checks that choice, retrieves its AlphaGenome prediction and saves the interaction before asking again. These are its first three choices at TCF7.
The first choice finds one target and uses one test.
Recorded Jev 1.14.0 full-feedback trajectory. The chooser never received the target key.
Inspect actual Jev choices 1, 2 and 3 from the same TCF7 full-feedback trajectory. E05+E06 is a target: one found with one test. E04+E06 is outside the top five: one remains found while two tests are used. E04+E05 is another target: two found with three tests. The saved interaction numbers shown at right were present in the exact input before the next choice; future outcomes and the evaluator key were absent. The application validates membership and uniqueness, charges the reveal budget and stores each observation before the next decision. This implements information flow; it does not itself establish that numerical feedback improves selection. The separate protected fresh AWS prototype uses Qwen3:1.7B and is a different policy condition. These audience controls inspect the recorded Jev study and request no inference.
A fixed ranking uses the same starting information.
The agent already knows what each edit does alone. A straightforward baseline ranks pairs by the sum of those two effect sizes, ignoring their signs. It starts with E05 and E06 here. Its order never changes after a pair result.
This sum sets the testing order. It does not predict the pair's interaction.
Both the agent and this ranking receive the reference prediction, ten single-edit predictions and fifteen pair tests.
The baseline uses abs(F(A)-F(reference))+abs(F(B)-F(reference)) with lexical pair ties. For TCF7, E05 and E06 have the two largest individual effect magnitudes: 0.2312244176864624 and 0.3232688903808594, totaling 0.5544933080673218. Their directions are negative; the ranking uses sizes. The displayed joining lengths share one physical scale. Pair outcomes cannot update this order. This comparison asks whether adaptive agent choices add value over a strong nonadaptive baseline. The following feedback control asks the narrower question of whether earlier numerical pair outcomes explain any benefit.
We compared the methods at twelve new regions.
We ran Jev and the fixed ranking with the same 15-test budget. To test feedback itself, we also ran Jev with earlier interaction values hidden, while retaining the names of tested pairs. Each region contributes five targets: 60 in total.
Filled circles mark found targets. Each five-circle group is one DNA region.
One trajectory per method at each region. These are twelve selected, non-overlapping regions.
Separate the two comparisons. Primary: Jev with full feedback versus the fixed ranking, 43 versus 40 targets out of sixty. Secondary: the same Jev interface with only earlier numerical interaction values hidden, 44 targets. Both Jev conditions retain tested pair identities and the same initial singles, candidate set and budget. Each displayed group corresponds to the same region across rows. Move the ruler to compare equal prefixes; at five tests, full feedback finds 53.3% versus 46.7% for withheld, and at ten 66.7% versus 63.3%. At fifteen, full feedback finishes one target behind withheld. This is one trajectory per condition and locus; Jev exposes no seed or temperature control. The original six-window Qwen study remains in the submitted report and is separate from this panel.
The observed gain is small; feedback remains unresolved.
Jev found 43 of 60 targets, compared with 40 for the fixed ranking. Hiding earlier interaction values produced 44, so we did not observe a final-count benefit from those values.
Three extra finds out of sixty is a five-percentage-point gain. The descriptive 95% interval runs from zero to ten points.
The run with feedback found targets sooner, but did not finish ahead. One run per condition and a selected set of regions leave the gain uncertain. This evaluates selection against model predictions; it does not establish a biological mechanism in cells.
The functioning agent loop and the selection result are distinct findings. Full-feedback Jev recovers three more targets than the fixed ranking: a mean paired gain of five percentage points, descriptive 95% locus-bootstrap interval zero to ten. Three loci improve and nine tie. Feedback is not established as the cause: the withheld condition ends at forty-four, while full feedback is forty-three; their paired contrast is -1.67 points with interval -6.67 to +3.33. Full feedback has a better early discovery curve. There is one trajectory per condition at twelve selected loci and no biological validation. The complete benchmark contains 672 actual AlphaGenome predictions and 360 successful Jev decisions. Attempt 88 failed with unknown HTTP cause and charge; one documented continuation preserved all 87 valid prior responses and its failed hold, then completed 273 new requests. No valid trajectory was discarded. The TCF7 illustration was selected after the completed study. Original formal reports and scientific records are unchanged.