APromoter / VirDLA

From model capability to virtual assay

VirDLA

A Zero-Shot Virtual Assay Harness for Multi-Omics Foundation Models

Turn multi-omics foundation models into virtual reporter assays. VirDLA extracts assay-relevant signals from model representations to compare promoter activity across tissues and power iterative sequence design.

From foundation model
to virtual experiment.

  1. 01 / Capability

    MFM

    Sequence representations and molecular predictions.

  2. 02 / Execution

    VirDLA harness

    Construct assembly, model calls and assay readout.

  3. 03 / Observation

    Virtual assay readout

    Relative reporter activity in a defined context.

  4. 04 / Iteration

    VirEvo

    Candidate search, optimization and compression.

Ask an experimental question.

How does this candidate behave relative to its reporter reference in the selected tissue or cell context?

Define

Sequence + assay design

Choose the candidate sequence, reporter reference, supported construct configuration and target tissue or panel.

T1 test and internal reference channels The candidate promoter drives Firefly luciferase in the test channel. HSV-TK drives Renilla luciferase in the internal reference channel. The channels are separated by a cHS4 insulator within one T1 cassette. Arrows show promoter-to-reporter direction. TestFirefly / luc2ReferenceRenilla cHS4insulatorCandidate promoter → FireflyHSV-TK promoter → RenillaT1 cassette · test channel + internal reference channel
T1 dual-reporter channel schematic. Red and blue mark the promoters; grey marks the reporter regions.
Execute

A model-backed protocol

VirDLA assembles the construct, locates reporter regions and invokes the selected model through its registered adapter.

VirDLA harness assay configuration + biological context
  1. 01 / AssembleReporter cassetteCandidate + fixed reporter context
  2. 02 / PredictFoundation modelSequence representations or molecular tracks
  3. 03 / Read outModel adapterReporter regions → relative activity
The foundation model supplies predictions. VirDLA defines the construct, executes the assay and computes its readout.
Read

A contextual result

The adapter maps predictions or representations to a relative reporter readout. Panel analysis compares target and reference contexts.

Compare activity across biological contexts.

VirDLA estimates relative reporter activity and compares target and reference contexts to prioritize candidates.

Relative activity
A virtual reporter readout produced by the selected backend and assay configuration.
Panel selectivity
A comparison across defined target and reference contexts, supporting candidate prioritization.
Reproducibility
Model version, reporter configuration and resource identifiers accompany the computational result.

Extract the signal.
Accelerate the assay.

VirDLA brings AlphaGenome, APenformer and Borzoi into one virtual assay workflow, connecting model capability to promoter comparison and sequence design.

Breaking the oracle bottleneck.

The full-readout route gains accuracy by expanding the input window, making repeated evaluation expensive. V3c extracts trunk activation signals directly: compact inputs deliver strong agreement at a fraction of the evaluation time.

27 ms per V3c evaluation

Evaluation time ms / sequence · lower is better

Full readout
1615
V3c
27
01615 ms

0.666 Pearson r

Agreement with measured activity Pearson r · higher is better

Full readout
0.655
V3c
0.666
01.0
Recorded ttset T8 / T9 benchmarks · 100 K562 SHARPR sequences · full-readout largest tested window vs. V3c smallest tested window.
Cassette encodingTrunk tensor activation extractionTissue-specific projectionV3c readout

Trunk activation extraction makes repeated virtual assays practical for sequence search. v11 carries this approach forward, powering VirDLA as the evaluation oracle in the VirEvo design loop.

VirDLA / APromoter

Define the assay.
Explore the sequence.

Open Design Studio →

Computation requires an authorized API key.
Please contact the administrator at contact@apromoter.bio from your institutional email for access.