01 / Atlas
The atlas
APromoter stores promoter sequences with predicted activity profiles across tissue and cell contexts. Search accepts promoter names, genes and panel terms. A record page reports its sequence, source annotations and activity profile.
Compare activity values within a fixed model, biological context, reporter cassette and reference construct.
02 / Data organization
Records and panels
Each promoter record links a sequence to its provenance and model-derived measurements. Panels group target and reference contexts into fixed comparison sets for tissue preference and enrichment analyses.
- Activity
- The reporter readout returned by a registered VirDLA configuration.
- Selectivity
- Activity in target contexts compared with activity in the panel's reference contexts.
- TBOX
- A versioned, machine-readable reporter construct definition.
03 / VirDLA
Virtual reporter assays
VirDLA is the assay layer between genomic foundation models and promoter analysis. It assembles a candidate sequence with a registered reporter construct, executes the selected model adapter and converts model signals into a relative reporter readout.
The trunk-activation route extracts tissue-relevant information from internal model tensors. Compact input contexts accelerate repeated sequence evaluation. APromoter uses the assay to build atlas profiles, while VirEvo uses it as an optimization oracle.
04 / Prediction
Prediction workflows
Single context
Evaluate one sequence with a selected model, tissue and reporter cassette. The result includes the registered activity readout and model-specific output details.
Batch sequences
Compare several sequences under one fixed assay configuration. Keeping the context and cassette constant makes the resulting values directly comparable.
Context panel
Evaluate one sequence across a released panel. The output orders contexts by activity and calculates target-to-reference selectivity from the panel's reference tissues.
05 / Analysis and design
Analysis and sequence design
Contribution enrichment
A deletion scan removes windows from a promoter and rescales the requested fitness measure. Positive impact indicates that deleting the window reduced fitness; negative impact indicates that deletion improved it. The result is an attribution over the submitted sequence under one explicit assay configuration.
VirEvo
VirEvo repeatedly calls VirDLA to search promoter sequence space. A run records its starting sequence, mutation settings, assay configuration and objective. VirEvo returns ranked sequence candidates for experimental testing.
Sequence inspection
APSCS calculates sequence statistics, panel activity and transcription-factor motif matches under the selected background model.
06 / Model scope
Registered model backends
The public release describes VirDLA configurations for AlphaGenome, APenformer and Borzoi. Each adapter has its own supported contexts, input geometry and output mapping. Raw outputs remain model-specific.
A reproducible result records the model backend, model version, tissue or panel, TBOX identifier and software configuration.
07 / Validation
Validation and interpretation
VirDLA configurations are compared with SHARPR-MPRA measurements from HepG2 and K562. Each reported correlation names its model, assay configuration and sequence set.
APromoter also compares literature-curated promoters with their assigned target and reference contexts. These analyses test whether the atlas recovers the expected direction of tissue preference.
The benchmark source is Ernst J. et al., Genome-scale high-resolution mapping of activating and repressive nucleotides in regulatory regions, with original data under GEO accession GSE71279.
08 / Resources
Downloads
The download center contains the promoter sequence pool used for database scanning, project-prepared SHARPR-MPRA test sets and released TBOX definitions. File descriptions distinguish original source data from subsets prepared for project benchmarks.
09 / Access
Public and compute-limited access
The atlas, About page, VirDLA overview, manuscript and this guide are public. Prediction and sequence-design workflows include compute-intensive tasks and require an API key. Contact the administrator at contact@apromoter.bio; please write from your institutional email address.