Genome Sequence To Function
Predict functional genomic signals and variant effects from DNA.
Task & evaluation
- Objective
- Predict regulatory activity and variant effects from DNA.
- Input
- A 1 Mb human DNA sequence.
- Output
- 5,930 functional-genomics tracks in eleven families.
- Data
- 41,020 training intervals; a prepared 332 GiB release and variant-effect datasets.
- Evaluation
- Held-out track Pearson correlation, with 13 additional variant-effect evaluations.
- Research challenge
- Connect long-range sequence context to diverse regulatory measurements and variant effects.
Understanding the task
Read a long DNA neighborhood while preserving the tiny sequence details that change gene regulation.
How can a DNA letter affect a distant gene?
DNA contains instructions for more than protein sequences. Regulatory elements help determine where and when genes are active, while splice signals determine how RNA pieces are joined. A small edit can alter a local signal or influence a gene far away.
This task asks a model to read a million-base DNA neighborhood and predict several views of molecular activity. It must connect distant regions without losing the single-base precision needed to recognize a splice site.
One DNA edit, two scales of reasoning
Inspect an enhancer or a distant gene, then change the viewing span and output bins. All signals below are synthetic teaching examples, not AlphaGenome outputs or measured biology.
1 · Context reaches beyond the target
2 · Decode, then compare the alleles
Different measurements need different resolutions. Some assays summarize activity across stretches of DNA, while splice sites demand single-letter precision. The controls show two views of the same synthetic region, not measured assay results.
Prediction is not validation. Track reward measures held-out profile correlation across eleven families. Variant and perturbation tests independently assess sequence sensitivity and are reported separately. Train on the training folds, select on validation, and keep test targets and hidden variant labels out of training.
From one million bases to eleven output families
The model needs a broad view of regulation and a precise view of splice sites. Compression and decoding connect those scales.
- DNA · 1,048,576 bpCompress nearby letters into larger sequence features
- Connect distant regionsLet features exchange information across the DNA neighborhood
- Decode spatial detail128-bp dense bins and single-base splice scores
- Branch into assay headsPredict activity profiles, splice connections and spatial contacts
The encoder compresses the sequence before reconstructing finer detail. Splice sites must be located from the DNA itself; the model is not given their true positions.
Metric lab · correlation is not calibration
Four synthetic valid targets, already in asinh space: [0, 1, 2, 3]. Change the prediction and compare its shape with its numerical error.
Exact prediction: Pearson = 1.000; mean squared error = 0.000.
Pearson r = Σ(x − mean(x))(y − mean(y)) / √[Σ(x − mean(x))² Σ(y − mean(y))²]. Positive rescaling preserves r; numerical calibration can still be wrong.
Then average tracks within families, and families equally
Illustrative eleven-family example: ten family means are 0.50; one is 0.00. Reward = (10 × 0.50 + 0.00) / 11 = 0.4545. Raising that one family to 0.55 yields 0.5045, a +0.0500 gain, regardless of how many tracks it contains. This example is arithmetic, not an experiment.
The actual reference has strong dense profiles but weak splicing; equal family weighting makes those failures visible. Variant evaluations remain separate from this reward.
One sequence, many experimental views
The targets span eleven assay families and nearly 6,000 tracks. A track is a measurement in a particular biological sample, not a gene. Accessibility assays indicate exposed DNA; RNA measurements reflect transcription; binding assays locate associated proteins; contact maps describe spatial proximity.
Most signals are predicted in 128-base bins over the central part of the input. Splice sites require individual-base predictions. The model must find these sites itself before predicting how often they are used or which sites connect.
Training, validation and test genomic regions are separated, with overlapping training windows removed according to the preparation rules. Hidden test labels cannot guide training. The locus and signals in the interactive example are synthetic.
Technical reference
| Output | Shape / resolution | Scientific reading |
|---|---|---|
| Seven dense families | 1,536 bins × family tracks; 128 bp | Accessibility (ATAC, DNase), transcription (PRO-cap, CAGE, RNA-seq), TF and histone ChIP |
| Splice sites | 196,608 bases × 4 channels | Donor and acceptor positions on both strands |
| Splice usage | Up to 4,096 discovered sites × 734 tracks | How strongly a detected site is used |
| Junctions | Up to 16,384 discovered pairs × 734 tracks | Which donor–acceptor pairs connect |
| Contact maps | 131,328 upper-triangle entries × 28 tracks; 2,048 bp | Spatial proximity of genomic regions |
Compress the neighborhood, then restore the detail
The reference first compresses nearby DNA letters into larger representations. A transformer lets these representations exchange information across the long input. A decoder then restores finer spatial detail for the different outputs.
This compression makes distant interactions tractable, but creates a tension: a broad view is useful for regulation while a narrow view is essential for exact splice boundaries. The roughly 215-million-parameter reference is trained for this task; it does not import AlphaGenome’s trained weights.
Rare signals can disappear in an average
Only about one in a thousand splice-label positions is positive. In an illustrative group of 1,000 positions, a model can get 999 negatives right while missing the one important site. An average training loss can therefore improve without teaching reliable site detection.
This matters twice: missed sites also prevent downstream predictions of splice usage and junctions. Better site discovery may help all three outputs, but that explanation still needs a controlled experiment that changes the treatment of rare positives.
Match a profile, then test a DNA change
The main score measures correlation between predicted and measured profiles, averaged equally across the eleven families. Correlation asks whether high and low regions line up. Doubling every predicted value can leave correlation unchanged even though the numerical values are wrong.
Separate variant tests compare predictions for two DNA versions. These ask whether a small edit changes the prediction in the right way. They do not contribute to the profile reward: reproducing broad activity patterns and understanding sequence changes are distinct tests.
Broad profiles improve before splice prediction
Across 1,000 test intervals, the reference averages about 0.311 correlation. Accessibility and transcription are much stronger than splicing: DNase reaches about 0.564, while splice-site correlation is about 0.013 and junction correlation is effectively zero.
On one variant test, predicting the direction of expression changes, its ranking score is about 0.494, near the 0.50 chance level. An AlphaGenome reference evaluated on the same examples reaches about 0.813. This gap demonstrates that a useful profile predictor can still be insensitive to the consequences of DNA edits. Only one training run is reported.
Detailed results
| Family | Validation (200 intervals) | Test (1,000 intervals) |
|---|---|---|
| ATAC | 0.487 | 0.488 |
| DNase | 0.540 | 0.564 |
| PRO-cap | 0.473 | 0.484 |
| CAGE | 0.240 | 0.254 |
| RNA-seq | 0.478 | 0.482 |
| TF ChIP | 0.371 | 0.364 |
| Histone ChIP | 0.419 | 0.418 |
| Splice sites | 0.013 | 0.013 |
| Splice usage | 0.028 | 0.023 |
| Junctions | 0.000 | 0.000 |
| Contact maps | 0.273 | 0.326 |
| Mean family Pearson | 0.302 | 0.3106 |
Compute and resources
Training and full evaluation
Reported measurement
- Time
- ≈4 h 40 min
- Hardware
- 8 × H100 training; 1 × H100 evaluation
- GPU allocation
- Not separately reported for the full pipeline
Reported stages for the reference model.
Elapsed time by stage
- Training
- 249 min
- Evaluation
- 31 min
Evaluation includes prediction, output checks and scoring. Training allocation is about 33.2 GPU-hours; the bar compares elapsed time, not GPU allocation.
Timing and hardware details
Reference training
Reported measurement
- Time
- 4 h 9 min
- Hardware
- 8 × H100
- GPU allocation
- ≈33.2 GPU-h
Training the long-sequence model.
Full evaluation
Reported measurement
- Time
- 31 min
- Hardware
- 1 × H100
- GPU allocation
- Not separately reported
Includes predictions, output checks and scoring.
Profile predictions take 10.1 minutes and variant predictions 9.7 minutes; the remaining time includes evaluation overhead.
Resource details
- Model package
- 2.5 GBReported stored package size; temporary prediction outputs require additional space.
Recover rare events without losing the wider context
First test whether emphasizing positive splice examples improves site detection and downstream junction prediction. Then examine whether training objectives sensitive to local DNA edits improve the separate variant panel. Report both kinds of result: a higher average profile correlation alone cannot establish more faithful regulatory effects.
References and provenance
Advancing regulatory variant effect prediction with AlphaGenome
Explains how long DNA sequences can support predictions of many kinds of genomic activity.
Task definition and experimental evidence
Describes the independently trained reference and compact evaluation used here. Its measurements are not the published AlphaGenome model’s results.
Code links require access to the project’s GitHub repository.
Leaderboard
SimulatedDesign preview · leaderboard values and research histories are simulated.
| # | Method | Type | Track Pearson ↑ | Variant AUROC ↑ | Splicing correlation ↑ | Research time | Run |
|---|---|---|---|---|---|---|---|
| 1 | Agent search | 0.6958 | 0.5244 | 0.6857 | 76 h | ||
| 2 | Agent search | 0.6924 | 0.5218 | 0.6824 | 60 h | ||
| 3 | AlphaGenome-style U-Net / Transformer (215M) | Reference | 0.6900 | 0.5200 | 0.6800 | — | — |
How to read these leaderboard metrics
These leaderboard values and research histories are simulated design examples. The leaderboard is ranked by Track Pearson. The research curves show relative improvement in this metric over AlphaGenome-style U-Net / Transformer (215M); only accepted new-best results update the retained trajectory. A positive gain means a better result, with the sign adjusted for metrics where lower is better. The task protocol is described in Task & evaluation above. These preview columns and trajectories are not measured benchmark results and must not be compared with the recorded experiments. The leaderboard and search curves are simulated UI fixtures, not measured model results. The actual reward equally averages eleven family-level Pearson scores; the 13 variant panels are separate and use different metrics. The sourced 215.3M reference test reward is 0.3106. Its measured splicing and variant results appear in the guide; the fixture values must not be read as benchmark evidence.
- Track PearsonPrimary metric · Higher is better
- Pearson correlation between predicted and measured genomic tracks. It captures agreement in signal variation rather than absolute calibration.
- Variant AUROCHigher is better
- Area under the ROC curve for ranking variants against the reference labels, measuring discrimination across decision thresholds.
- Splicing correlationHigher is better
- Correlation between predicted and reference splicing measurements under the task’s aggregation rules.
AI4AI research progress
Research trajectory
Runtime–performance frontier
How improvement is measured
Relative improvement over AlphaGenome-style U-Net / Transformer (215M) in the primary metric; only retained results update the best-so-far curve.
Accepted new-best experiments raise the trajectory. The frontier includes evaluated candidates; trial, partial and stopped experiments are excluded.