Abstract
Comparing promoter activity across tissues requires a consistent way to evaluate sequences in defined biological contexts. We present VirDLA, a zero-shot virtual assay harness for multi-omics foundation models that assembles reporter constructs, executes model-specific adapters and converts model signals into relative reporter activity. The zero-shot configuration fixes model and calibration parameters before candidate evaluation. Using AlphaGenome-backed VirDLA, we constructed APromoter Atlas, comprising 20,018 promoter records across 285 tissue and cellular contexts and more than 5.7 million virtual activity values. Atlas profiles reveal variation in activity breadth and similarity among related biological contexts. In matched-window benchmarks, the trunk activation readout configuration achieved its highest Pearson correlation and shortest evaluation time at the smallest tested window. In the primary experimental endpoint, AlphaGenome-backed H5 readouts reached a Pearson correlation of 0.837 with SHARPR activity in HepG2; correlations across the four tested backend–cell-line combinations ranged from 0.512 to 0.837. Across 78 literature-curated promoters, 72 (92.3%) recovered either the predefined target-panel direction or a tissue pattern concordant with the cited source in the complete spectrum. The accompanying web service combines Atlas queries with on-demand VirDLA prediction for user-submitted promoter sequences. VirDLA turns multi-omics foundation models into executable virtual reporter assays, while APromoter Atlas demonstrates this capability at the scale of 20,018 promoters and 285 biological contexts. VirDLA connects foundation-model capabilities to executable virtual reporter assays, and APromoter Atlas provides the resulting cross-context activity profiles for promoter comparison and selection.
Keywords: VirDLA; APromoter; Promoter; Virtual dual-luciferase assay; Foundation model; Gene regulation
1. Introduction
Promoter selection requires knowing how strongly a sequence drives expression and how that activity changes across cellular backgrounds. Massively parallel reporter assays (MPRAs) have made these questions accessible for large sequence collections. SHARPR-MPRA resolved activating and repressive contributions within regulatory regions through densely overlapping reporter measurements (Ernst et al., 2016). More recently, lentiviral MPRAs characterized more than 680,000 sequences across three cell types and examined sequence features underlying promoter activity and cell-type specificity (Agarwal et al., 2025). These studies establish functional measurements as a basis for comparing regulatory sequences. Extending such comparisons across a broad range of tissues calls for a way to evaluate the same candidate under consistently defined assay conditions.
Sequence-to-function models provide a computational basis for this expansion. Enformer integrates long-range sequence information to predict gene expression and other genomic signals (Avsec et al., 2021). AlphaGenome extends multimodal prediction to transcription, splicing and chromatin properties across tissue and cellular contexts (Avsec et al., 2026). Sei showed how predicted chromatin profiles could be organized into sequence classes that support a global map of regulatory activity and the interpretation of genetic variants (Chen et al., 2022). Reporter measurements also support task-specific use of these model outputs: the lentiMPRA study fitted predictors of measured activity using Enformer and Sei features (Agarwal et al., 2025). Here, multi-omics foundation models (MFMs) refer to pretrained sequence-to-function models that predict molecular readouts across multiple omics modalities. An MFM alone does not produce an assay-level quantity that can be compared with reporter activity; it returns position-resolved predictions across molecular tracks. Turning those predictions into a virtual reporter assay requires a defined candidate construct, matched reference construct, biological context and assay-level readout.
Here we introduce VirDLA, a zero-shot virtual assay harness for MFMs. VirDLA supplies this assay layer by assembling paired reporter constructs, invoking a model-specific adapter in the selected biological context and converting the returned model signals into relative reporter activity. The experimental specification remains fixed when the MFM backend is exchanged. VirDLA fixes the model and calibration parameters before evaluating new candidates. Prediction-based readout (PB) and Trunk Activation Readout (TAR) provide two ways to obtain the assay signals. Repeating the workflow across sequences and contexts produces APromoter Atlas, a cross-context resource for promoter comparison and selection.
We use VirDLA to map 20,018 promoter records across 285 tissue and cellular contexts with AlphaGenome. The resulting Atlas combines genome-derived fragments with literature-curated and evolved promoters. We characterize activity breadth and relationships among contexts, evaluate the H5 assay against experimental measurements for 1,055 SHARPR fragments in two cell lines, and compare the spectra of 78 literature-curated promoters with documented tissue-associated patterns. Matched-window experiments examine how input length affects experimental agreement and evaluation cost in the PB and TAR configurations.
2. Results
2.1 Establishing a virtual reporter assay with compact-context evaluation
VirDLA is a zero-shot virtual assay harness for multi-omics foundation models. The name derives from Virtual Dual-Luciferase Assay: paired test and reference reporter cassettes provide internal-reference normalization to correct scaling effects in MFM outputs. It defines the reporter construct, reference configuration and biological context for a candidate sequence, assembles the model input and executes the selected backend through a model-specific adapter (Fig. 1a). An assay readout rule converts the returned model signals into relative reporter activity. Model and routing-calibration parameters are fixed before candidate evaluation. Repeating the assay across promoter sequences and biological contexts produces the activity matrix of APromoter Atlas.
VirDLA supports two readout approaches: prediction-based readout (PB), which calculates assay activity from model output predictions, and Trunk Activation Readout (TAR), which extracts assay-relevant signals from internal trunk activations. Biological context is mapped to model tracks, linking trunk activations to the corresponding pretrained readout weights in TAR and selecting context-specific predictions in PB. We compared the window dependence and computational cost of these two VirDLA configurations.
Compact TAR inputs improved agreement and reduced computation. We compared TAR with a prediction-based readout calculated from model output predictions, using 100 K562 SHARPR sequences at five matched windows on one GPU. Expanding the input from 16K to 256K bp increased prediction-based Pearson r from 0.335 to 0.596 and evaluation time from 77.8 to 916.2 ms per sequence (Fig. 1b,c). Over the same range, TAR r declined from 0.630 to 0.588 (Fig. 1e). Shortening its window improved Pearson agreement and accelerated evaluation 9.52-fold, from 401.8 to 42.2 ms per sequence (Fig. 1f).
The PB–TAR benchmark tests the computational scalability required for large-scale Atlas precomputation. Correlation thresholds establish matched experimental agreement, and evaluation time at each threshold is the comparison endpoint (Fig. 1h). At Pearson r ≥ 0.55 or, separately, Spearman ρ ≥ 0.74, the fastest qualifying prediction-based configuration used a 256K window (916.2 ms per sequence), whereas the compact TAR configuration used a 16K window (42.2 ms), a 21.7-fold reduction in evaluation time. The 16K TAR configuration also reached r ≥ 0.60 and ρ ≥ 0.80. At matched experimental-agreement thresholds, the compact TAR configuration reduced evaluation time by 21.7-fold relative to the tested prediction-based configuration. This efficiency makes Atlas-scale mapping of 285-context promoter spectra computationally practical. We therefore used TAR-based VirDLA with AlphaGenome to construct APromoter Atlas.
2.2 Mapping cross-context promoter activity with APromoter Atlas
APromoter Atlas maps human promoter activity across 285 tissue and cellular contexts using AlphaGenome-backed VirDLA. The current collection contains 20,018 promoter records and more than 5.7 million virtual reporter activity values. It combines 19,915 genome-derived promoters with approximately 80 literature-curated promoters and a small set of evolved sequences. The contexts include 124 tissues, 85 primary-cell contexts, 52 cell lines and 24 contexts of in vitro differentiated cells. The online resource presents these precomputed profiles and accepts user-submitted promoter sequences for on-demand VirDLA prediction in selected contexts.
Named promoter profiles connect the resource to individual genes (Fig. 2a). The genome-derived ALB_2500 record reaches half of its maximum activity in six contexts, compared with 84 for TTR_2500 and 33 for SYN1_2500. Their highest readouts occur in the right liver lobe, neural progenitor cells and endocrine pancreas, respectively. These profiles illustrate the variation in breadth and peak context among fixed TSS-derived sequences.
Activity is concentrated in a minority of contexts for much of the genome-derived collection (Fig. 2c). The median relative breadth is 37 of 285 contexts, with an interquartile range of 13–88; 4,219 promoters reach half-maximal activity in ten contexts or fewer, whereas 1,470 reach it in at least half of the contexts. Higher mean activity accompanies broader profiles (Spearman ρ = 0.513) and lower relative variation across contexts (ρ = −0.442).
Contexts with related biological annotations have more similar promoter profiles (Fig. 2b,d). Median context-pair Spearman correlation is 0.876 within the 19 ontology categories and 0.790 between categories. Inferior and superior right ventricular myocardium are closely matched (ρ = 0.9996), as are the amygdala and Ammon’s horn (ρ = 0.9991). The context-similarity matrices calculated from two random halves of the promoter collection correlate at r = 0.998. These relationships persist across promoter subsets, providing a stable organization for cross-context comparisons in the Atlas.
2.3 Experimental sequence ranking and MFM backend portability
We evaluated the H5 virtual assay using 1,055 promoter-associated SHARPR fragments with measurements in HepG2 and K562 from 69 regions (Ernst et al., 2016). H5 incorporates selected sequences from the reporter plasmids used to generate the experimental benchmark. This places the virtual construct closer to the experimental configuration and reduces construct-related variation. H5 compares a candidate-driven test reporter with a fixed-promoter reference reporter. Each fragment occupies the candidate slot of the same H5 reporter construct. The evaluation tests sequence-activity agreement separately within each cell line (Fig. 3). AlphaGenome is the backend used to generate the current APromoter Atlas profiles; APenformer is included in this benchmark to test execution of the same assay through a second MFM adapter.
The primary endpoint was agreement between experimental and virtual sequence activity within each cellular background (Fig. 3a–e). AlphaGenome reached a Pearson correlation of 0.837 in HepG2 and 0.644 in K562, with Spearman correlations of 0.507 and 0.631. APenformer reached Pearson correlations of 0.512 and 0.639, with Spearman correlations of 0.659 and 0.771. All four locus-bootstrap intervals for Pearson correlation lie above zero. These results establish experimental agreement for ranking candidate sequences under a defined assay and cellular context. APromoter Atlas uses the AlphaGenome-backed implementation; the AlphaGenome H5 results therefore provide experimental support for the sequence-ranking capability underlying its precomputed activity spectra.
The same primary endpoint showed increasing concordance across successive MFM backends (Fig. 3f). In the matched T1 comparison on the same 1,055 fragments, mean within-cell Pearson correlation increased from 0.061 with Borzoi (Linder et al., 2025) to 0.594 with APenformer/Enformer and 0.755 with AlphaGenome. The H5 benchmark retained the APenformer-to-AlphaGenome ordering, with mean Pearson correlations of 0.576 and 0.741, respectively. Assay concordance increased from Borzoi to APenformer/Enformer and AlphaGenome. VirDLA therefore converts advances in MFM capability into higher virtual-assay fidelity through replaceable model-specific adapters.
2.4 Literature-guided analysis recovers tissue patterns across curated promoters
We compared the complete 285-context activity spectra of 78 literature-curated promoters across 14 classes with their documented tissue associations (Fig. 4a; Supplementary Tables S3 and S4). Seventy-two records (92.3%) recovered the predefined target-panel direction or a concordant tissue pattern in the complete spectrum. The recovered spectra included hepatic, cardiac, pulmonary, adipose and epithelial patterns.
The literature-curated ALB and TTR fragments each placed their three highest readouts in liver and its left and right lobes (Artemyev et al., 2026); F9 showed the same pattern (Salier et al., 1990). NPPA and NPPB peaked in cardiac contexts (Wu et al., 1989; LaPointe et al., 1996), SFTPC in lung contexts (Wambach et al., 2010) and ADIPOQ in adipose contexts (Kyriakou et al., 2008). CKM and ACTA1 showed positive skeletal-muscle panel enrichment, while VWF, CDH5 and NOS3 showed positive endothelial and vascular enrichment (Trask et al., 1988; Brennan and Hardeman, 1993; Jahroudi and Lynch, 1994; Prandini et al., 2005; Guillot et al., 1999). KRT18 showed high activity across liver, gallbladder mucosa, pancreas, stomach and lung, consistent with its reported simple-epithelial distribution (Abe and Oshima, 1990). PLP1 peaked in sciatic nerve, with tibial nerve ranked third, recovering the peripheral-nerve component described in its source study (Hamdan et al., 2018); brain ranked 110th among the 285 contexts.
The predefined panel analysis measured broad tissue-class enrichment across the curated cohort (Fig. 4b–d). Target mean activity exceeded reference mean activity in 68 of 78 records (87.2%), and the target-ranking probability exceeded 0.5 in 70 records (89.7%). The median enrichment ratio was 2.23 and the median ranking probability was 0.799. All nine liver records showed target enrichment, with a median ratio of 25.3; the cardiac and adipose groups had median ratios of 4.07 and 4.25.
Full-spectrum review added four records to the 68 panel-positive records, giving 72 of 78 recovered tissue patterns. KRT18 was added because its leading liver, gallbladder mucosa, pancreas, stomach and lung contexts matched the simple-epithelial distribution in the source study (Abe and Oshima, 1990). PLP1 was added for its sciatic- and tibial-nerve peaks (Hamdan et al., 2018), KDR for a fifth-ranked coronary-artery signal (Patterson et al., 1995), and SLC12A3 for a kidney-cortex signal at rank 27 (MacKenzie et al., 2001). Supplementary Table S4 reports the record-level full-spectrum review. The recovery of source-consistent patterns in 72 of 78 curated promoters demonstrates that APromoter Atlas preserves biologically interpretable tissue specificity across its cross-context activity spectra.
3. Discussion
VirDLA establishes an experimental interface for multi-omics foundation models, organizing their predictive capabilities around a defined reporter assay. The assay specification evaluates candidate sequences through model-specific adapters and reports their relative activities in a chosen biological context. APromoter Atlas applies the AlphaGenome-backed assay to 20,018 promoter records across 285 contexts.
The compact TAR configuration provides the computational efficiency required to execute this assay at Atlas scale. At shared Pearson and Spearman agreement thresholds, it reduced per-sequence evaluation time 21.7-fold relative to the tested prediction-based configuration. This gain supports millions of sequence–context evaluations, enabling 285-context activity spectra to be mapped and maintained as a precomputed database (Fig. 1).
The H5 benchmark establishes experimental agreement for the assay readout within defined cellular environments. AlphaGenome-backed VirDLA reached Pearson correlations of 0.837 in HepG2 and 0.644 in K562, providing experimental support for the sequence-ranking capability underlying the precomputed Atlas spectra (Fig. 3). APenformer reproduced positive within-cell agreement through a second adapter, confirming that the same assay specification can be executed across MFM backends.
VirDLA converts progress in MFM performance into higher assay concordance. Mean within-cell Pearson correlation increased from 0.061 with Borzoi to 0.594 with APenformer/Enformer and 0.755 with AlphaGenome in the matched T1 comparison. Model-specific adapters preserve the reporter construct, reference configuration, biological context and readout task while exchanging the MFM backend. This architecture gives VirDLA a direct path to higher fidelity as successive MFMs improve and supports future remapping with stronger backends.
APromoter Atlas extends the experimentally supported sequence-ranking assay into a cross-context promoter resource. Its profiles distinguish promoter strength from activity breadth and recover biologically related organization among tissue contexts (Fig. 2). Literature-guided analysis recovered source-consistent tissue patterns in 72 of 78 curated promoters, demonstrating that APromoter Atlas preserves biologically interpretable tissue specificity in its cross-context activity spectra (Fig. 4). The Atlas therefore supports promoter comparison and selection across tissues through both precomputed profiles and on-demand VirDLA evaluation.
Two aspects of the current evaluation can be improved. Four of the 78 promoters produced source-consistent patterns only when the complete spectrum was examined because the predefined tissue panels did not include every informative context. Future panel definitions can incorporate these context-level patterns directly. The literature-guided evaluation covers promoters with established tissue associations; expanding this curated set will test additional promoter classes and biological contexts.
Together, these results establish VirDLA as a scalable virtual-assay harness: TAR enables large-scale computation, replaceable MFM adapters provide a direct path to higher fidelity, and APromoter Atlas converts these capabilities into biologically interpretable promoter spectra for comparison, screening and prioritization.
Methods
Virtual reporter assay specification
VirDLA takes a candidate sequence, a reporter-cassette configuration and a tissue or cellular context as inputs. The cassette places the candidate promoter before the test reporter, followed by an insulator, a fixed reference promoter and a reference reporter. Replacing the candidate sequence preserves the remaining cassette components. The resulting virtual activity is indexed by candidate, context and model execution configuration.
Atlas generation uses T1 (pGL4.10_DLAmonitor), selected because most of its construct architecture derives from the pGL4.10-based dual-luciferase system.
Experimental validation uses H5, which incorporates selected sequences from the reporter plasmids used in the study that generated the benchmark dataset. H5 reduces construct-related variation between the virtual and experimental assays. It contains hluc2 and rluc2 reporter segments with hGH polyadenylation sequences, the same insulator as T1 and a minimal reference promoter.
Model execution and context mapping
VirDLA assembles the reporter construct, executes the selected model and maps its signals to the requested biological context. AlphaGenome contexts resolve to ontology identifiers and their associated model tracks. APenformer uses Enformer outputs with context-specific track selection and routing calibration. HepG2 and K562 map to EFO:0001187 and EFO:0002067 in the context-to-track mapping.
The Atlas snapshot uses the AlphaGenome-based VirDLA v11 implementation with T1 and a 12,288-bp input window. The H5 evaluation uses the AlphaGenome adapter at 12,288 bp and the APenformer adapter at 196,608 bp. Both H5 adapters center the complete construct with N padding and evaluate human model outputs. The three-backend T1 comparison evaluates the same 1,055 fragments through the Borzoi, APenformer/Enformer and AlphaGenome adapters at their registered input windows of 524,288, 196,608 and 12,288 bp, respectively. Mean within-cell Pearson correlation is the arithmetic mean of the HepG2 and K562 correlations for each backend.
APenformer routing is calibrated using the predicted activity profiles in the database. H5 evaluation retains the T1 routing-calibration parameters and pretrained Enformer weights. Context-specific activity values and within-cell analyses were unchanged. Zero-shot evaluation fixes the model and calibration parameters before applying the configuration to experimental candidates.
Virtual assay readouts
Prediction-based readout (PB) calculates activity from model output predictions. Trunk activation readout (TAR) extracts reporter-region signals from internal trunk representations using context-linked readout weights. The AlphaGenome v11 adapter applies the local RNA-seq head to trunk representations and averages the resulting signals over the selected context tracks and each reporter region.
For candidate s in context c, let Atest(s, c) and Aref(s, c) denote the mean test- and reference-reporter signals. The AlphaGenome dual-reporter activity is
RAG(s, c) = Atest(s, c) / Aref(s, c).
The implementation uses a sign-preserving denominator floor of 10−12. APenformer combines candidate-to-reference promoter-track ratios with differential test- and reference-reporter representations. Its fixed context-specific calibration maps these features to relative activity. The single-reporter TAR configuration used for the window sweep is specified under Computational benchmark.
APromoter Atlas construction
The analyzed Atlas snapshot contains 20,018 records. The genome-derived subset contains 19,915 TSS-centered sequences from hg38 using GENCODE v39 gene annotation (Frankish et al., 2021). Sequences span 2,000 bp upstream and 500 bp downstream in the transcriptional orientation, with inclusive coordinates yielding 2,501 bp; negative-strand intervals are reverse-complemented. Each record retains its source annotation. Curated records link the stored fragment to the relevant promoter literature.
APromoter Atlas is mapped with AlphaGenome. The analyzed precomputed profiles cover 285 contexts: 124 tissues, 85 primary-cell contexts, 52 cell lines and 24 in vitro differentiated contexts. Each promoter is represented by its activity profile across these contexts.
Atlas profile analysis
Genome-wide summaries use the 19,915 genome-derived records. Relative breadth counts contexts with activity at least half of a promoter’s maximum. Mean activity is the arithmetic mean across contexts, and the coefficient of variation is the population standard deviation divided by that mean. Context similarity is Spearman correlation across promoters. Within-category and between-category summaries use the upper triangle of the context-similarity matrix. Split-half stability is the Pearson correlation between the corresponding upper triangles of context-similarity matrices calculated from two randomly partitioned promoter subsets.
Experimental validation
SHARPR-MPRA data originate from GSE71279. The evaluation contains 1,055 fragments of 145 bp from 69 parent regions with measurements in HepG2 and K562. Published nucleotide-level log2 activity estimates are averaged within each fragment and exponentiated to obtain linear activity. Each fragment is inserted into H5, and predictions are matched to observations by fragment identity and cell line. Pearson and Spearman correlations measure within-cell agreement.
Agreement between experimental and virtual activity within each cellular context is the primary H5 endpoint. Pearson correlation measures linear agreement across the 1,055 fragments, and Spearman correlation measures agreement in their activity ordering. Pearson intervals are percentile intervals from 1,000 bootstrap samples of the 69 parent regions, retaining all fragments from each sampled region together. Each replicate samples 69 regions with replacement; its correlation uses the concatenated fragments, including repeated regions. Interval endpoints are the 2.5th and 97.5th percentiles. The same sampled groups are used across execution configurations.
Literature-guided promoter comparisons
We searched the complete database for promoter sequences curated from published sources. Records containing at least one specific literature citation entered the analysis; records marked as evolved designs and placeholder source entries were excluded. This database-wide rule yielded 78 promoters across 14 classes.
The curated sources include studies of the F9 regulatory region (Salier et al., 1990), muscle promoters CKM and ACTA1 (Trask et al., 1988; Brennan and Hardeman, 1993), neuronal SYN1 and astrocytic GFAP promoters (Kügler et al., 2003; Lee et al., 2008), and endothelial promoters VWF, CDH5 and NOS3 (Jahroudi and Lynch, 1994; Prandini et al., 2005; Guillot et al., 1999). Supplementary Data S3a retains the source associated with each fragment.
Literature-linked hTERT hotspot variants were retained as experimentally defined promoter constructs.
Each promoter class maps to Supplementary Data S3b in APromoter Atlas. We retrieved both sets and their activity values from the saved 285-context profile for each promoter. Mean enrichment equals mean target activity divided by mean reference activity. Target-ranking probability is the fraction of all target–reference pairs in which target activity is higher, with ties assigned a weight of 0.5. Values above one for mean enrichment and above 0.5 for ranking probability indicate the expected tissue-class direction.
For the spectrum heatmap, each record was divided by its maximum activity across the 285 contexts. Rows follow database promoter-class order; columns are ordered by ontology category and context name, using the same category definitions as Figure 2. All 78 records and all 285 contexts are displayed. Heatmap normalization affects the visualization; panel summaries use the original activity values.
For records with a mean ratio at or below one or a ranking probability at or below 0.5, we inspected the complete ranked spectrum against the promoter-specific tissue or cell-type description in the cited source. A record was classified as recovered when its target mean exceeded its reference mean or when the complete spectrum contained a tissue signal concordant with the cited source. Full-spectrum review added KRT18, PLP1, KDR and SLC12A3 based on their simple-epithelial, peripheral-nerve, vascular and renal signals. Supplementary Table S4 reports the peak context, the best-ranked predefined target and the record-level interpretation.
We calculated both measurements per promoter and summarized each class by its median. Supplementary Data S3a lists the complete cohort and activity statistics; Supplementary Table S4 gives the record-level review of promoters below the panel-mean threshold.
Computational benchmark
Both routes evaluate the same 100 K562 SHARPR sequences at 16,384, 32,768, 65,536, 131,072 and 262,144 bp. Correlation thresholds provide the matched-performance criterion, and minimum evaluation time at each threshold is the comparison endpoint. This design measures the computational requirement of large-scale precomputation. Both configurations run on an NVIDIA GeForce RTX 4080 SUPER using FP32 and batch size one. Each configuration and window has three timed runs following three warmup sequences. CUDA-synchronized wall time includes construct assembly, encoding, device transfer, inference and scalar readout; model loading and warmup are excluded. Times are summarized as the mean and sample standard deviation of the three run-level means.
The prediction-based readout route uses a construct containing the reference promoter, centered N padding and an RNA-seq dual-reporter ratio. The TAR uses a construct without the reference promoter, left-aligned repeated-insulator padding and a test-reporter activation readout. These settings remain fixed within each window sweep. For each Pearson or Spearman threshold, we select the tested window with the lowest mean evaluation time among those meeting or exceeding that threshold. Pearson and Spearman thresholds are evaluated separately; a threshold is marked as not reached when none of the tested windows meets it. Pearson and Spearman correlations compare the 100 sequence predictions with their experimental activities; repeated execution measures timing variation on this fixed sample set.
Data availability
Supporting data are publicly available on Zenodo at doi:10.5281/zenodo.22763539 (Nie, 2026).
The analyzed Atlas data are available through the APromoter repository at apromoter.bio. The public interface supports queries of the precomputed profiles. On-demand VirDLA prediction accepts user-submitted promoter sequences and requires an API key obtained by contacting the administrator. The analyzed release, database_P_v1, contains 20,018 promoter records across 285 tissue and cellular contexts. Supplementary Data S1 and its sample table contain the window benchmark; S2a and S2b contain cell-line-specific H5 results for AlphaGenome and APenformer; S2c contains the matched three-backend T1 results for Fig. 3f; S3a and S3b contain the literature-promoter cohort and its context sets. Tables S1–S4 and data-file descriptions appear in the Supplementary information.
Code availability
The VirDLA harness and a reference implementation of the virtual-assay readout are released as the open-source virdla package at github.com/JiGuzhai/VirDLA. The framework can also be executed through the APromoter web service at apromoter.bio, which provides model access and the precomputed Atlas.
References
Abe, M. & Oshima, R. G. A single human keratin 18 gene is expressed in diverse epithelial cells of transgenic mice. Journal of Cell Biology 111, 1197–1206 (1990). doi:10.1083/jcb.111.3.1197.
Agarwal, V. et al. Massively parallel characterization of transcriptional regulatory elements. Nature 639, 411–420 (2025). doi:10.1038/s41586-024-08430-9.
Artemyev, V. et al. Engineering liver-specific promoters: A comprehensive review of design, mechanisms, and clinical applications in gene therapy. Cells 15, 14 (2026). doi:10.3390/cells15010014.
Avsec, Ž. et al. Effective gene expression prediction from sequence by integrating long-range interactions. Nature Methods 18, 1196–1203 (2021). doi:10.1038/s41592-021-01252-x.
Avsec, Ž. et al. Advancing regulatory variant effect prediction with AlphaGenome. Nature 649, 1206–1218 (2026). doi:10.1038/s41586-025-10014-0.
Brennan, K. J. & Hardeman, E. C. Quantitative analysis of the human alpha-skeletal actin gene in transgenic mice. Journal of Biological Chemistry 268, 719–725 (1993). doi:10.1016/S0021-9258(18)54211-3.
Chen, K. M. et al. A sequence-based global map of regulatory activity for deciphering human genetics. Nature Genetics 54, 940–949 (2022). doi:10.1038/s41588-022-01102-2.
Ernst, J. et al. Genome-scale high-resolution mapping of activating and repressive nucleotides in regulatory regions. Nature Biotechnology 34, 1180–1190 (2016). doi:10.1038/nbt.3678. Data: GSE71279.
Frankish, A. et al. GENCODE 2021. Nucleic Acids Research 49, D916–D923 (2021). doi:10.1093/nar/gkaa1087.
Guillot, P. V. et al. A vascular bed–specific pathway regulates cardiac expression of endothelial nitric oxide synthase. Journal of Clinical Investigation 103, 799–805 (1999). doi:10.1172/JCI6017.
Hamdan, H., Patyal, P., Kockara, N. T. & Wight, P. A. The wmN1 enhancer region in intron 1 is required for expression of human PLP1. Glia 66, 1763–1774 (2018). doi:10.1002/glia.23339.
Jahroudi, N. & Lynch, D. C. Endothelial-cell-specific regulation of von Willebrand factor gene expression. Molecular and Cellular Biology 14, 999–1008 (1994). doi:10.1128/MCB.14.2.999-1008.1994.
Kyriakou, T. et al. Adiponectin gene ADIPOQ SNP associations with serum adiponectin in two female populations and effects of SNPs on promoter activity. Journal of Human Genetics 53, 718–727 (2008). doi:10.1007/s10038-008-0303-1.
Kügler, S., Kilic, E. & Bähr, M. Human synapsin 1 gene promoter confers highly neuron-specific long-term transgene expression from an adenoviral vector in the adult rat brain depending on the transduced area. Gene Therapy 10, 337–347 (2003). doi:10.1038/sj.gt.3301905.
LaPointe, M. C., Wu, G., Garami, M., Yang, X.-P. & Gardner, D. G. Tissue-specific expression of the human brain natriuretic peptide gene in cardiac myocytes. Hypertension 27, 715–722 (1996). doi:10.1161/01.HYP.27.3.715.
Lee, Y., Messing, A., Su, M. & Brenner, M. GFAP promoter elements required for region-specific and astrocyte-specific expression. Glia 56, 481–493 (2008). doi:10.1002/glia.20622.
Linder, J. et al. Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation. Nature Genetics 57, 949–961 (2025). doi:10.1038/s41588-024-02053-6.
MacKenzie, S., Vaitkevicius, H. & Lockette, W. Sequencing and characterization of the human thiazide-sensitive Na-Cl cotransporter (SLC12A3) gene promoter. Biochemical and Biophysical Research Communications 282, 991–1000 (2001). doi:10.1006/bbrc.2001.4673.
Nie, L. Supporting data for “VirDLA: A Zero-Shot Virtual Assay Harness for Multi-Omics Foundation Models”. Version v1. Zenodo [dataset] (2026). doi:10.5281/zenodo.22763539.
Patterson, C. et al. Cloning and functional analysis of the promoter for KDR/flk-1, a receptor for vascular endothelial growth factor. Journal of Biological Chemistry 270, 23111–23118 (1995). doi:10.1074/jbc.270.39.23111.
Prandini, M. H. et al. The human VE-cadherin promoter is subjected to organ-specific regulation and is activated in tumour angiogenesis. Oncogene 24, 2992–3001 (2005). doi:10.1038/sj.onc.1208483.
Salier, J. P., Hirosawa, S. & Kurachi, K. Functional characterization of the 5′-regulatory region of human factor IX gene. Journal of Biological Chemistry 265, 7062–7068 (1990). doi:10.1016/S0021-9258(19)39259-2.
Trask, R. V., Strauss, A. W. & Billadello, J. J. Developmental regulation and tissue-specific expression of the human muscle creatine kinase gene. Journal of Biological Chemistry 263, 17142–17149 (1988). doi:10.1016/S0021-9258(18)37510-0.
Wambach, J. A. et al. Surfactant protein-C promoter variants associated with neonatal respiratory distress syndrome reduce transcription. Pediatric Research 68, 216–220 (2010). doi:10.1203/PDR.0b013e3181eb5d68.
Wu, J., LaPointe, M. C., West, B. L. & Gardner, D. G. Tissue-specific determinants of human atrial natriuretic factor gene expression in cardiac tissue. Journal of Biological Chemistry 264, 6472–6479 (1989). doi:10.1016/S0021-9258(18)83372-5.