Model type
Sequence-to-function convolutional/transformer model
Model
AlphaGenome
AlphaGenome predicts regulatory activity and variant effects from long DNA sequences, with outputs for expression, splicing, chromatin and contact maps.
Sources (3)
google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps44 evaluations · 52 metric rows · 4 evaluated configurations using this model
How it worksAlphaGenome workflow
Conceptual summary of the documented data flow; optional inputs and configured downstream stages must be reported for a reproducible evaluation.
Sources (3)
google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact mapsOverview
Inputs
DNA sequence and optional variant information.
Sources (3)
google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact mapsOutputs
Predicted expression, splicing, chromatin features and contact maps.
Sources (3)
google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact mapsAccess
Official project documentation and implementation: https://github.com/google-deepmind/alphagenome_research
Sources (3)
google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact mapsSource reviewed · Automated source review, 2026-09-16. All specifications and missing details
Evaluations and results
44 evaluations · 52 metric rows. Different protocols are not a single leaderboard.
Filter evaluations
Applied filters: All linked evaluations
| Tested configuration | Protocol and dataset | Finding | Evidence and details |
|---|---|---|---|
| Configuration: AlphaGenome distilled cell-type-matched DNase scorer, Borzoi-matched locus subset | Protocol: Zero-shot CAGI5 MPRA activity-effect prediction (borzoi-ensemble-matched comparison) (AlphaGenome paper) Dataset subset: Zero-shot CAGI5 MPRA activity-effect prediction (borzoi-ensemble-matched comparison): evaluated data subset | 0.56 mean_pearsonr_all correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceCompute locus/context Pearson correlations between predicted and observed effects. The Borzoi strategy uses its reported scoring windows and modified matching; the ChromBPNet comparison excludes the unavailable TERT-GBM context. Aggregation: Mean Pearson correlation over the included locus/context comparisons; Table4 rows10 and11 have different AlphaGenome scalars and must remain separate. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 4 Variant performan'!M11 |
| Configuration: AlphaGenome Enformer-split model with 128 bp Enformer head | Protocol: histone ChIP-seq track prediction at 128 bp (AlphaGenome paper) Dataset subset: histone ChIP-seq track prediction at 128 bp: evaluated data subset | 0.71 pearsonr correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceCompare signals in aligned Enformer target bins with the dedicated head and the corresponding Enformer version. Aggregation: Pearson correlation summarized over the relevant modality tracks. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K35; 'Suppl Table 3 Track performance'!K36 |
| Configuration: AlphaGenome Borzoi-fold-1 model fine-tuned with base-resolution Borzoi RNA-seq head | Protocol: RNA expression correlation across tracks (AlphaGenome paper) Dataset subset: RNA expression correlation across tracks: evaluated data subset | 0.57 Gene TPM normalized
(cross-track pearsonr) correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceCompute per-gene cross-track correlation on normalized, gene-mean-centred expression. Aggregation: Pearson correlation across tracks. Preserve raw versus normalized Table3 rows; the inspected task paragraph does not specify the weighting used to reduce all correlations to the table scalar. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K14 |
| Configuration: AlphaGenome Borzoi-fold-1 model fine-tuned with 32 bp Borzoi-matching head | Protocol: CAGE track prediction at 32 bp (AlphaGenome paper) Dataset subset: CAGE track prediction at 32 bp: evaluated data subset | 0.74 pearsonr correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlign predicted and experimental values on the same 32-bp tracks and test intervals. Aggregation: Pearson correlation; group summaries derive from individual track correlations. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K18 |
| Configuration: AlphaGenome fold-0 track model on comparator-matched test peaks | Protocol: PRO-cap prediction on held-out peaks (AlphaGenome paper) Dataset subset: PRO-cap prediction on held-out peaks: evaluated data subset | 0.79 pearsonr log total count correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceEvaluate signal correlation, log-total-count correlation and profile JSD as separate metrics; retain only endpoints actually listed for this evaluation in Table3. Aggregation: Table3 reports task summary scalars; Extended Data Fig.3 shows per-cell-line comparisons. The inspected caption does not fully specify pooling across peaks/cell lines. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K16 |
| Configuration: AlphaGenome Enformer-split model with 128 bp Enformer head | Protocol: ATAC track prediction at 128 bp (AlphaGenome paper) Dataset subset: ATAC track prediction at 128 bp: evaluated data subset | 0.72 pearsonr correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome Enformer-split model with 128 bp Enformer head: ATAC track prediction at 128 bp Compare signals in aligned Enformer target bins with the dedicated head and the corresponding Enformer version. Aggregation: Pearson correlation summarized over the relevant modality tracks. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K32; 'Suppl Table 3 Track performance'!K33 |
| Configuration: AlphaGenome fold-1 ensemble, splice-junction head | Protocol: Human splice-junction classification and count prediction (AlphaGenome paper) Dataset subset: Human splice-junction classification and count prediction: evaluated data subset | 0.82 pearsonr correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceFor classification, contrast read-supported junctions with unobserved donor–acceptor pairs. For count prediction, correlate log(1+x)-transformed predicted and observed counts only on nonzero observed junctions. Aggregation: Classification: auPRC per tissue averaged across tissues. Quantitative endpoint: Pearson correlation of log-transformed counts; do not pool these two metric rows. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K8 |
| Configuration: AlphaGenome ensemble of four fold-1 models | Protocol: Human splice-site classification: annotation-derived (AlphaGenome paper) Dataset subset: Human splice-site classification: annotation-derived: evaluated data subset | 0.82 auPRC dimensionless · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome ensemble of four fold-1 models: Human splice-site classification: annotation-derived Compare probabilities with the selected binary splice-site labels at genomic positions. Aggregation: Compute auPRC separately for four strand/site classes and average the four values. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K4; 'Suppl Table 3 Track performance'!K5 |
| Configuration: AlphaGenome distilled all-fold student | Protocol: Polyadenylation-QTL causality (AlphaGenome paper) Dataset subset: Polyadenylation-QTL causality: evaluated data subset | 0.63 PAS_10000_average_auprc dimensionless · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome distilled all-fold student: Polyadenylation-QTL causality Summarize predicted RNA coverage near PASs, derive the largest allelic change in proximal-versus-distal usage, average tracks and classify positive versus matched negative variants. Unscored variants receive zero as specified by the paper. Aggregation: Average auPRC over100 random positive-to-negative matching permutations; Table4 endpoint is PAS_10000. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 4 Variant performan'!M9 |
| Configuration: AlphaGenome distilled all-fold student | Protocol: Author MFASS splice-disruption prediction (AlphaGenome paper) Dataset subset: Author MFASS splice-disruption prediction: evaluated data subset | 0.51 all_tissues_auprc dimensionless · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome distilled all-fold student: Author MFASS splice-disruption prediction Use the MFASS-specific donor/acceptor and junction scoring procedure, sum splicing components, then average across tissues. Compare with the source’s exon-inclusion disruption label. Aggregation: auPRC for the splice-disrupting label defined by the MFASS exon-inclusion threshold. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 4 Variant performan'!M8 |
| Configuration: AlphaGenome fold-0 track model on comparator-matched test peaks | Protocol: DNase prediction on held-out peaks (AlphaGenome paper) Dataset subset: DNase prediction on held-out peaks: evaluated data subset | 0.79 log1p_count_pearsonr correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome fold-0 track model on comparator-matched test peaks: DNase prediction on held-out peaks Evaluate signal correlation, log-total-count correlation and profile JSD as separate metrics; retain only endpoints actually listed for this evaluation in Table3. Aggregation: Table3 reports task summary scalars; Extended Data Fig.3 shows per-cell-line comparisons. The inspected caption does not fully specify pooling across peaks/cell lines. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K23 |
| Configuration: AlphaGenome fold-0 contact-map model, bilinearly resized to Orca 4 kb bins | Protocol: chromatin contact-map prediction (AlphaGenome paper) Dataset subset: chromatin contact-map prediction: evaluated data subset | 0.79 Per-cell type
@4kb pearsonr correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceCompare predicted and observed contact maps for each held-out interval and cell type. Aggregation: Mean Pearson correlation across held-out interval evaluations; cell types remain explicit before reporting the summary. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K40 |
| Configuration: AlphaGenome Borzoi-fold-1 model fine-tuned with base-resolution Borzoi RNA-seq head | Protocol: RNA expression correlation across genes (AlphaGenome paper) Dataset subset: RNA expression correlation across genes: evaluated data subset | 0.52 Gene TPM normalized
(cross-gene pearsonr) correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceCompute per-track cross-gene correlation on raw log-expression and, separately, normalized/mean-centred expression. Aggregation: Pearson correlation across genes. Preserve raw versus normalized Table3 rows; the inspected task paragraph does not specify the weighting used to reduce all correlations to the table scalar. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K15 |
| Configuration: AlphaGenome fold-0 track model on comparator-matched test peaks | Protocol: ATAC prediction on held-out peaks (AlphaGenome paper) Dataset subset: ATAC prediction on held-out peaks: evaluated data subset | 0.86 pearsonr correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome fold-0 track model on comparator-matched test peaks: ATAC prediction on held-out peaks Evaluate signal correlation, log-total-count correlation and profile JSD as separate metrics; retain only endpoints actually listed for this evaluation in Table3. Aggregation: Table3 reports task summary scalars; Extended Data Fig.3 shows per-cell-line comparisons. The inspected caption does not fully specify pooling across peaks/cell lines. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K28 |
| Configuration: AlphaGenome distilled all-fold student | Protocol: Splice-site-region variants splicing-based classification (AlphaGenome paper) Dataset subset: Splice-site-region variants splicing-based classification: evaluated data subset | 0.57 auprc_max_abs_track_aggregation dimensionless · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome distilled all-fold student: Splice-site-region variants splicing-based classification Score each variant from its predicted splicing changes and compare with the selected ClinVar pathogenicity labels. Aggregation: auPRC in the category, retaining its own class prevalence and sampling scheme. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 4 Variant performan'!M4 |
| Configuration: AlphaGenome Borzoi-fold-1 model fine-tuned with 32 bp Borzoi-matching head | Protocol: histone ChIP-seq track prediction at 32 bp (AlphaGenome paper) Dataset subset: histone ChIP-seq track prediction at 32 bp: evaluated data subset | 0.69 pearsonr correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlign predicted and experimental values on the same 32-bp tracks and test intervals. Aggregation: Pearson correlation; group summaries derive from individual track correlations. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K34 |
| Configuration: AlphaGenome distilled all-fold student | Protocol: European-ancestry LCL caQTL effect-size prediction (AlphaGenome paper) Dataset subset: European-ancestry LCL caQTL effect-size prediction: evaluated data subset | 0.59 pearsonr correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome distilled all-fold student: European-ancestry LCL caQTL effect-size prediction Use the source’s local log-fold-change scorer, select the task-specific cell/assay tracks and compare against the source-provided QTL labels or measured effects. Aggregation: Pearson correlation against reported effect sizes of the causal/significant QTL set. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 4 Variant performan'!M27; 'Suppl Table 4 Variant performan'!M28 |
| Configuration: AlphaGenome distilled all-fold student | Protocol: Zero-shot GTEx splicing-outlier prediction (AlphaGenome paper) Dataset subset: Zero-shot GTEx splicing-outlier prediction: evaluated data subset | 0.22 auprc dimensionless · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome distilled all-fold student: Zero-shot GTEx splicing-outlier prediction Evaluate the sequence-derived scores directly on the held-out variants. Aggregation: Pooled auPRC across GTEx tissues after assigning positive variants their observed tissues and negatives tissues sampled to match the positive tissue distribution. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 4 Variant performan'!M6 |
| Configuration: AlphaGenome Enformer-split model with 128 bp Enformer head | Protocol: TF ChIP-seq track prediction at 128 bp (AlphaGenome paper) Dataset subset: TF ChIP-seq track prediction at 128 bp: evaluated data subset | 0.58 pearsonr correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome Enformer-split model with 128 bp Enformer head: TF ChIP-seq track prediction at 128 bp Compare signals in aligned Enformer target bins with the dedicated head and the corresponding Enformer version. Aggregation: Pearson correlation summarized over the relevant modality tracks. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K38; 'Suppl Table 3 Track performance'!K39 |
| Configuration: AlphaGenome Enformer-split model with 128 bp Enformer head | Protocol: DNase track prediction at 128 bp (AlphaGenome paper) Dataset subset: DNase track prediction at 128 bp: evaluated data subset | 0.87 pearsonr correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome Enformer-split model with 128 bp Enformer head: DNase track prediction at 128 bp Compare signals in aligned Enformer target bins with the dedicated head and the corresponding Enformer version. Aggregation: Pearson correlation summarized over the relevant modality tracks. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K26; 'Suppl Table 3 Track performance'!K27 |
| Configuration: AlphaGenome fold-0 track model on comparator-matched test peaks | Protocol: ATAC prediction on held-out peaks (AlphaGenome paper) Dataset subset: ATAC prediction on held-out peaks: evaluated data subset | 0.85 log1p_count_pearsonr correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome fold-0 track model on comparator-matched test peaks: ATAC prediction on held-out peaks Evaluate signal correlation, log-total-count correlation and profile JSD as separate metrics; retain only endpoints actually listed for this evaluation in Table3. Aggregation: Table3 reports task summary scalars; Extended Data Fig.3 shows per-cell-line comparisons. The inspected caption does not fully specify pooling across peaks/cell lines. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K29 |
| Configuration: AlphaGenome fold-0 track model on comparator-matched test peaks | Protocol: DNase prediction on held-out peaks (AlphaGenome paper) Dataset subset: DNase prediction on held-out peaks: evaluated data subset | 0.75 pearsonr correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome fold-0 track model on comparator-matched test peaks: DNase prediction on held-out peaks Evaluate signal correlation, log-total-count correlation and profile JSD as separate metrics; retain only endpoints actually listed for this evaluation in Table3. Aggregation: Table3 reports task summary scalars; Extended Data Fig.3 shows per-cell-line comparisons. The inspected caption does not fully specify pooling across peaks/cell lines. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K22 |
| Configuration: AlphaGenome distilled all-fold student | Protocol: African-ancestry LCL caQTL classification (AlphaGenome paper) Dataset subset: African-ancestry LCL caQTL classification: evaluated data subset | 0.56 auPRC dimensionless · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome distilled all-fold student: African-ancestry LCL caQTL classification Use the source’s local log-fold-change scorer, select the task-specific cell/assay tracks and compare against the source-provided QTL labels or measured effects. Aggregation: auPRC over causal/noncausal labels. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 4 Variant performan'!M19; 'Suppl Table 4 Variant performan'!M20 |
| Configuration: AlphaGenome fold-0 track model on comparator-matched test peaks | Protocol: PRO-cap prediction on held-out peaks (AlphaGenome paper) Dataset subset: PRO-cap prediction on held-out peaks: evaluated data subset | 0.53 jsd dimensionless · lower Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceEvaluate signal correlation, log-total-count correlation and profile JSD as separate metrics; retain only endpoints actually listed for this evaluation in Table3. Aggregation: Table3 reports task summary scalars; Extended Data Fig.3 shows per-cell-line comparisons. The inspected caption does not fully specify pooling across peaks/cell lines. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 3 Track performance'!K17 |
| Configuration: AlphaGenome distilled all-fold student | Protocol: Splicing-QTL causality (AlphaGenome paper) Dataset subset: Splicing-QTL causality: evaluated data subset | 0.76 tissue_weighted_mean_auprc dimensionless · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · source checkedMethods, coverage and sourceAlphaGenome distilled all-fold student: Splicing-QTL causality Compare predicted splicing effects for causal sQTLs and matched negatives; keep the wide and proximal splice-distance analyses distinct. Aggregation: auPRC per tissue, averaged with variant-count weights. AlphaGenome Nature 2026 supplementary comparison tables · 'Suppl Table 4 Variant performan'!M5 |
Source checking is not independent reproduction. Release 2026-09-23-2b89723c6dd9.
Related configurations, pipelines and services
These configurations, services and pipelines use this model within their own configurations. Their results, where available, are not assigned to the underlying model.
- AlphaGenome absolute multimodal scores → RandomForestClassifier Pipeline · 1 metric rows
- AlphaGenome variant scores plus expressed-splice-site feature → Explainable Boosting Classifier Pipeline · 1 metric rows
- AlphaGenome distilled multimodal DNase, RNA-seq and histone ChIP-seq features → per-locus LassoCV Pipeline · 1 metric rows
- ENCODE-rE2G extended logistic regression plus AlphaGenome K562 RNA-seq input × gradient feature Pipeline · 1 metric rows
Use this model
How it works, versions and access
Versions and evaluated configurations
- AlphaGenome distilled cell-type-matched DNase scorer, ChromBPNet-matched locus subset · family
- AlphaGenome fold-1 ensemble, splice-junction head · family
- AlphaGenome Borzoi-fold-1 model fine-tuned with base-resolution Borzoi RNA-seq head · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome fold-0 contact-map model, bilinearly resized to Orca 4 kb bins · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome Borzoi-fold-1 model fine-tuned with 32 bp Borzoi-matching head · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome ensemble of four fold-1 models · family
- AlphaGenome fold-1 splice-site-usage model · family
- AlphaGenome fold-0 track model on comparator-matched test peaks · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome Borzoi-fold-1 model fine-tuned with base-resolution Borzoi RNA-seq head · family
- AlphaGenome distilled all-fold student · family
- AlphaGenome distilled cell-type-matched DNase scorer, Borzoi-matched locus subset · family
- AlphaGenome Enformer-split model with 128 bp Enformer head · family
- AlphaGenome distilled all-fold student · family
How it works
How it works
AlphaGenome progressively downsamples DNA with convolutional blocks, then uses a transformer tower and pairwise interaction blocks to represent long-range context. A U-Net-style decoder restores sequence resolution using skip connections. Modality-specific heads predict one-dimensional genomic tracks, splicing outputs and two-dimensional contact maps.
Sources (3)
google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact mapsVersions and reproducibility
The paper distinguishes fold-specific evaluation models, all-fold teachers and distilled students; a family name alone does not choose one of these configurations. Up to 1 million base pairs; single-base outputs for the modalities described in the README.
Sources (3)
google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact mapsStrengths, limitations and unresolved questions
Strengths and limitations
Strengths supported by sources
- Combines several molecular readouts in one sequence model, with variant scoring utilities.
Sources (3)
google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps
Limitations and conditions
- The paper identifies remaining challenges beyond 100kb, context-specific variant effects and non-coding genes. Species coverage is human/mouse and personal-genome prediction was not benchmarked in the study.
Sources (3)
google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps
Profile review details
Inspected pinned official documentation, relevant implementation files and named primary-paper sections. Claims are limited to those artifacts. Remaining field extraction and identity conflicts are explicit; no new performance claims, model runs or human review are implied.
Stable record: catalog-model-alphagenomeSpecifications
Inputs, training, access and other details
Explanatory profile: source reviewed · Automated source review, 2026-09-16. Review applies to the cited claims; unresolved fields are listed below. Numerical results retain their own review status.
| Property | Description and evidence |
|---|---|
| Model type | Sequence-to-function convolutional/transformer modelSources (3)google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps |
| Architecture | U-Net-inspired sequence backbone combining convolutional local processing with transformer blocks for longer-range interactions; one-dimensional track heads and two-dimensional contact-map representations.Sources (3)google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps |
| Inputs | DNA sequence and optional variant information.Sources (3)google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps |
| Outputs | Predicted expression, splicing, chromatin features and contact maps.Sources (3)google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps |
| Parameters | Approximately 450M trainable parameters, including encoder, sequence transformer, pairwise blocks, decoder and prediction heads.Sources (3)google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps |
| Known versions | The paper distinguishes fold-specific evaluation models, all-fold teachers and distilled students; a family name alone does not choose one of these configurations.Sources (3)google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps |
| Training data | Human and mouse molecular datasets; separate fold-specific models for held-out reference-interval evaluation and all-fold teachers for student distillation.Sources (3)google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps |
| Training cutoff | ENCODE RNA-seq and chromatin metadata were downloaded 9–17 January 2025; the contact-map source was accessed 4 March 2021. These are component retrieval dates, not one universal latest-experiment cutoff.Sources (3)google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps |
| Context limits | Up to 1 million base pairs; single-base outputs for the modalities described in the README.Sources (3)google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps |
| Weights licence | Non-commercial AlphaGenome model terms; not the Apache licence covering source code.Sources (3)google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps |
| Access | Official project documentation and implementation: https://github.com/google-deepmind/alphagenome_researchSources (3)google-deepmind/alphagenome_research: README.md; alphagenome: Journal full-text XML; alphagenome-supp: Publisher supplementary archive · Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps |
| Code licence | Apache-2.0Sourcesgoogle-deepmind/alphagenome_research: LICENSE · LICENSE: licence text |
Evidence
Source checking verifies the cited claim or transcription. It does not establish independent reproduction.
Evidence table
Inspect claims, sources and review details
Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.
One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.
58 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Diagram caption Conceptual summary of the documented data flow; optional inputs and configured downstream stages must be reported for a reproducible evaluation. Individual claims | alphagenome: Journal full-text XML Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Retrieved page snapshot; no immutable publisher revision supplied | source checked automated source review · 2026-09-16 Audit detailsInspected pinned official documentation, relevant implementation files and named primary-paper sections. Claims are limited to those artifacts. Remaining field extraction and identity conflicts are explicit; no new performance claims, model runs or human review are implied. Field: Source artifact SHA-256: Hash scope: SHA-256 of retrieved original artifact bytes Format: original_artifact |
| Diagram caption Conceptual summary of the documented data flow; optional inputs and configured downstream stages must be reported for a reproducible evaluation. Individual claims | google-deepmind/alphagenome_research: README.md Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: 0db53bd4352c66d1e00a049a81da373a066e6670 | source checked automated source review · 2026-09-16 Audit detailsInspected pinned official documentation, relevant implementation files and named primary-paper sections. Claims are limited to those artifacts. Remaining field extraction and identity conflicts are explicit; no new performance claims, model runs or human review are implied. Field: Source artifact SHA-256: Hash scope: SHA-256 of retrieved original artifact bytes Format: original_artifact |
| Diagram caption Conceptual summary of the documented data flow; optional inputs and configured downstream stages must be reported for a reproducible evaluation. Individual claims | alphagenome-supp: Publisher supplementary archive Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Retrieved page snapshot; no immutable publisher revision supplied | source checked automated source review · 2026-09-16 Audit detailsInspected pinned official documentation, relevant implementation files and named primary-paper sections. Claims are limited to those artifacts. Remaining field extraction and identity conflicts are explicit; no new performance claims, model runs or human review are implied. Field: Source artifact SHA-256: Hash scope: SHA-256 of retrieved original artifact bytes Format: original_artifact |
Diagram steps
| alphagenome: Journal full-text XML Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Retrieved page snapshot; no immutable publisher revision supplied | source checked automated source review · 2026-09-16 Audit detailsInspected pinned official documentation, relevant implementation files and named primary-paper sections. Claims are limited to those artifacts. Remaining field extraction and identity conflicts are explicit; no new performance claims, model runs or human review are implied. Field: Source artifact SHA-256: Hash scope: SHA-256 of retrieved original artifact bytes Format: original_artifact |
Diagram steps
| google-deepmind/alphagenome_research: README.md Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: 0db53bd4352c66d1e00a049a81da373a066e6670 | source checked automated source review · 2026-09-16 Audit detailsInspected pinned official documentation, relevant implementation files and named primary-paper sections. Claims are limited to those artifacts. Remaining field extraction and identity conflicts are explicit; no new performance claims, model runs or human review are implied. Field: Source artifact SHA-256: Hash scope: SHA-256 of retrieved original artifact bytes Format: original_artifact |
Diagram steps
| alphagenome-supp: Publisher supplementary archive Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Retrieved page snapshot; no immutable publisher revision supplied | source checked automated source review · 2026-09-16 Audit detailsInspected pinned official documentation, relevant implementation files and named primary-paper sections. Claims are limited to those artifacts. Remaining field extraction and identity conflicts are explicit; no new performance claims, model runs or human review are implied. Field: Source artifact SHA-256: Hash scope: SHA-256 of retrieved original artifact bytes Format: original_artifact |
| Diagram title AlphaGenome workflow Individual claims | alphagenome: Journal full-text XML Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Retrieved page snapshot; no immutable publisher revision supplied | source checked automated source review · 2026-09-16 Audit detailsInspected pinned official documentation, relevant implementation files and named primary-paper sections. Claims are limited to those artifacts. Remaining field extraction and identity conflicts are explicit; no new performance claims, model runs or human review are implied. Field: Source artifact SHA-256: Hash scope: SHA-256 of retrieved original artifact bytes Format: original_artifact |
| Diagram title AlphaGenome workflow Individual claims | google-deepmind/alphagenome_research: README.md Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: 0db53bd4352c66d1e00a049a81da373a066e6670 | source checked automated source review · 2026-09-16 Audit detailsInspected pinned official documentation, relevant implementation files and named primary-paper sections. Claims are limited to those artifacts. Remaining field extraction and identity conflicts are explicit; no new performance claims, model runs or human review are implied. Field: Source artifact SHA-256: Hash scope: SHA-256 of retrieved original artifact bytes Format: original_artifact |
| Diagram title AlphaGenome workflow Individual claims | alphagenome-supp: Publisher supplementary archive Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Retrieved page snapshot; no immutable publisher revision supplied | source checked automated source review · 2026-09-16 Audit detailsInspected pinned official documentation, relevant implementation files and named primary-paper sections. Claims are limited to those artifacts. Remaining field extraction and identity conflicts are explicit; no new performance claims, model runs or human review are implied. Field: Source artifact SHA-256: Hash scope: SHA-256 of retrieved original artifact bytes Format: original_artifact |
| Model type Sequence-to-function convolutional/transformer model Individual claims | alphagenome: Journal full-text XML Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Retrieved page snapshot; no immutable publisher revision supplied | source checked automated source review · 2026-09-16 Audit detailsInspected pinned official documentation, relevant implementation files and named primary-paper sections. Claims are limited to those artifacts. Remaining field extraction and identity conflicts are explicit; no new performance claims, model runs or human review are implied. Field: Source artifact SHA-256: Hash scope: SHA-256 of retrieved original artifact bytes Format: original_artifact |
Sources and history
View linked audit checks and correction history
Release 2026-09-23-2b89723c6dd9 · Record review: discovered
5 source records and release history
- AlphaGenome official resource · Original source · Not reported
- google-deepmind/alphagenome_research: README.md · Original source · 0db53bd4352c66d1e00a049a81da373a066e6670
- alphagenome: Journal full-text XML · Original source · Retrieved page snapshot; no immutable publisher revision supplied
- alphagenome-supp: Publisher supplementary archive · Original source · Retrieved page snapshot; no immutable publisher revision supplied
- google-deepmind/alphagenome_research: LICENSE · Original source · 0db53bd4352c66d1e00a049a81da373a066e6670
Technical metadata and extraction receipts
Stable ID: catalog-model-alphagenome
- areas
- dna-genomes
- method types
- foundation model
- entity level
- family
- version
- API / released weights
- reported name
- AlphaGenome
- access
- Rate-limited, non-commercial API requires a key. Downloadable weights require accepting non-commercial model terms; local inference recommends an H100 GPU.
- method type
- foundation model
- historical missing metadata
- checkpoint revision: not_yet_extracted; training data: not_yet_extracted; licence: not_yet_extracted
- metadata review scope
- historical_missing_metadata preserves the original discovery state. Current descriptive evidence and missingness are recorded in profile.facts; numerical-result review is separate.
- entity classification
- review date: 2026-09-17; rationale: The cited profile describes a named learned biological predictor or representation model/family. Preserve this identity separately from task-specific fitting, individual checkpoints, pipelines and hosted access.; source ids: evidence-official-578439da3f3a3a476b7f; evidence-official-56e5abfb5f12f1cd3b20; evidence-official-bc0327ef2ea6107e1773; source locator: Paper: Unifying DNA sequence-to-function model and Discussion; README.md: Model weights and licensing; Supplementary Methods: Model (p.9), ENCODE RNA-seq Data (p.3), Contact maps; ambiguities: None recorded
Related records
- family: AlphaGenome distilled cell-type-matched DNase scorer, ChromBPNet-matched locus subset
- family: AlphaGenome fold-1 ensemble, splice-junction head
- family: AlphaGenome Borzoi-fold-1 model fine-tuned with base-resolution Borzoi RNA-seq head
- uses model: AlphaGenome absolute multimodal scores → RandomForestClassifier
- family: AlphaGenome distilled all-fold student
- family: AlphaGenome distilled all-fold student
- uses model: AlphaGenome variant scores plus expressed-splice-site feature → Explainable Boosting Classifier
- family: AlphaGenome fold-0 contact-map model, bilinearly resized to Orca 4 kb bins
- family: AlphaGenome distilled all-fold student
- family: AlphaGenome distilled all-fold student
- family: AlphaGenome distilled all-fold student
- family: AlphaGenome Borzoi-fold-1 model fine-tuned with 32 bp Borzoi-matching head
- family: AlphaGenome distilled all-fold student
- family: AlphaGenome distilled all-fold student
- family: AlphaGenome distilled all-fold student
- family: AlphaGenome ensemble of four fold-1 models
- family: AlphaGenome fold-1 splice-site-usage model
- family: AlphaGenome fold-0 track model on comparator-matched test peaks
- family: AlphaGenome distilled all-fold student
- family: AlphaGenome distilled all-fold student
- family: AlphaGenome distilled all-fold student
- family: AlphaGenome distilled all-fold student
- family: AlphaGenome distilled all-fold student
- uses model: AlphaGenome distilled multimodal DNase, RNA-seq and histone ChIP-seq features → per-locus LassoCV
- family: AlphaGenome Borzoi-fold-1 model fine-tuned with base-resolution Borzoi RNA-seq head
- family: AlphaGenome distilled all-fold student
- family: AlphaGenome distilled cell-type-matched DNase scorer, Borzoi-matched locus subset
- uses model: ENCODE-rE2G extended logistic regression plus AlphaGenome K562 RNA-seq input × gradient feature
- family: AlphaGenome Enformer-split model with 128 bp Enformer head
- family: AlphaGenome distilled all-fold student