COMPLETED COMPUTATIONAL REANALYSIS · 19 SEPTEMBER 2026

RNA-count filtering makes plant promoter rankings sensitive to sequencing depth

A reproducible audit of public measurements used in plant promoter modelling.

Research-note draft for expert review. No new wet-lab experiments or model training. No manuscript has been submitted. Open the complete discovery journal.

The result: across 12 libraries, retaining and pooling DNA-supported barcodes gives quarter-depth rank correlations of 0.990–0.997, compared with 0.768–0.956 for the source RNA-filtered median. Agreement between biological RNA replicates improves in all six comparisons. The estimator is established; the contribution is this measured effect in a widely reused plant resource.

What this means in ordinary language

A promoter is a DNA region involved in starting transcription. These experiments attach short DNA tags, called barcodes, to promoter constructs. Counting each tag in DNA estimates how much of the construct was present; counting it in RNA estimates its output. Several tags provide several measurements of each promoter.

The source analysis keeps only tags with at least five RNA reads. When fewer reads are sampled, some tags fall below that threshold. This changes which measurements contribute to a promoter’s score. We tested that effect by randomly removing reads from the same libraries and comparing promoter rankings. Pooling counts across a set selected using DNA alone substantially reduces the instability.

Study design and provenance

We reanalyzed the public resource of Jores and colleagues (2021). Promoter origins are Arabidopsis (At), maize (Zm) and sorghum (Sb); the assay hosts are tobacco leaves and maize protoplasts. These labels must not be confused: an At leaf-library result is an Arabidopsis promoter sequence assayed in tobacco.

The main robustness experiment covers both biological replicates for all three promoter libraries in both hosts, using the enhancer-present, dark-condition measurements. It uses the deposited WT/full-length barcode assignments. At protoplast replicates share the same DNA-input table; their RNA samples are separate. The other comparisons also remain limited to two replicates per library/host combination.

An initial At leaf pilot was followed by decision gates recorded before validation: at quarter RNA depth, rank correlation below 0.95 for the source estimator in At replicate 2 and Zm replicate 1, and at least 0.05 improvement for all-barcode pooling. Both gates passed. The extension to the remaining eight libraries was specified before their results. All outcomes are retained in the journal.

How the test works

For every barcode, RNA counts were independently drawn from Binomial(observed count, fraction), at fractions 0.5 and 0.25, using five fixed seeds per library. DNA counts were unchanged. The primary DNA cutoff was five reads, with a predeclared cutoff of twenty as sensitivity analysis. This is an RNA-depth intervention, not a simulation of every experimental source of noise.

We compared four fixed estimators: the median of log2 RNA/DNA ratios after requiring RNA ≥5; a median retaining all DNA-supported tags with 0.5 added to each RNA and DNA count; pooled RNA/DNA after RNA filtering; and pooled RNA/DNA retaining all DNA-supported tags, including observed RNA zeros. Pooled estimates use a single 0.5 pseudocount on each aggregate. Each estimator is normalized using its corresponding estimate for the same internal control.

Each depth comparison uses identical promoter sets across all four methods: at least ten DNA-supported tags at full depth, a measurable full-depth score in all methods, and a measurable thinned score in all methods. Dropout is reported separately. Rank correlation is insensitive to global control offsets; activity-shift and dropout tables are also provided. Full observed counts are a comparison baseline, not biological ground truth.

Quarter-depth rank stability for all four estimators in all twelve libraries.
Figure 1. Points are means over five thinnings; bars show their minimum and maximum, not biological confidence intervals. At, Zm and Sb identify promoter origins. Every method is compared on the same promoters within a library/seed.
Assay Promoter origin Replicate RNA-filtered median All-barcode pooled
leaf At 1 0.7681 0.9929
leaf At 2 0.8648 0.9950
leaf Zm 1 0.9379 0.9971
leaf Zm 2 0.9360 0.9966
leaf Sb 1 0.9564 0.9974
leaf Sb 2 0.9497 0.9970
proto At 1 0.9525 0.9959
proto At 2 0.9541 0.9959
proto Zm 1 0.9160 0.9901
proto Zm 2 0.9213 0.9932
proto Sb 1 0.9490 0.9929
proto Sb 2 0.9045 0.9927

Does it also help independent replicate agreement?

Yes, in these data. On shared promoters with at least ten DNA-supported barcodes in each replicate, all-barcode pooling improves Spearman correlation in all six assay/library combinations. For At promoters in leaves, the correlation increases from 0.9550 to 0.9847 across 5,700 promoters; for Zm promoters in leaves, from 0.9688 to 0.9893 across 8,869. The six comparisons span 0.9212–0.9814 before and 0.9534–0.9931 after pooling.

Between-replicate agreement under the four measurement estimators.
Figure 2. Each point compares the same two RNA biological replicates. Promoter counts do not represent independent biological replicates. At protoplast replicates share their DNA input.

The factorial comparison separates two choices: which barcodes survive the RNA filter, and how surviving counts are aggregated. Keeping all DNA-supported tags already improves depth robustness with the median estimator; pooling provides a further improvement. This supports a contribution from changing barcode membership, while not proving that pooled estimates are unbiased biological truth.

Consequences for evaluating a published model

We reused the original archived held-out CNN predictions without retraining. Their files omit gene IDs; only unique matches by promoter species and published target value within 10⁻¹⁰ were retained. Across both test files, 699 ambiguous predictions were excluded. The further requirement of adequate DNA support in both replicates and complete estimates at both depths leaves 1,860 leaf and 2,311 protoplast test promoters.

For the same 1,860 leaf promoters and fixed predictions, Spearman correlation changes from 0.8185 to 0.7789 after quarter-depth thinning under the source median estimator. With all-barcode pooling it remains 0.8225 versus 0.8231. Protoplast values are 0.8248 versus 0.8024, and 0.8211 versus 0.8197, respectively. The model itself has not changed: only its evaluation targets have. This predeclared example uses thinning seed zero.

Frozen CNN evaluation depends on how experimental targets are processed.
Figure 3. Same frozen predictions and same selected test promoters at each depth. These are sensitivity measurements on a supported subset, not replacement leaderboard scores. Raw RMSE and species-specific results are included in the tables.

A separate provenance finding, with an important limit

Reconstructing the source measurements exposed a historical parsing discrepancy. readr 1.4.0 inferred fixed-width count columns from the first 1,000 lines; later counts with more digits could lose leading digits. An actual R 4.0.3/readr 1.4.0 container reproduced our independent Python audit summaries for all 58 native-library count files. The legacy reconstruction matches all 869,289 archived per-replicate values.

Robust whitespace parsing changes 60,051 of 448,345 averaged condition/promoter values, mostly through small shared control-normalization offsets. Only 53 change by more than one log2 unit. We subsequently found that an author-posted attachment dated 28 August 2023 already contains all 869,289 correctly parsed per-replicate values. Our reconstruction matches it to floating-point precision. Thus we do not claim to have first produced these corrected measurements.

The July 2025 RData deposit, the 2023 attachment and the CNN exports must be distinguished. All 147,966 PGB promoter examples match the original CNN sequences and labels exactly. The CNN labels differ slightly from both reconstructed source snapshots, with some larger exceptions; their precise generation history remains unresolved. The audit does not establish that current PGB labels contain the historical parser error, and it does not replace them.

What is new, and what already existed?

Earlier workWhat it establishesScope of this contribution
mpralm, 2019Count aggregation and averaging can differ in bias and statistical performance.No new pooling estimator is claimed.
MPRAsnakeflow, 2025Barcode/count downsampling and replicate-quality assessment already exist for MPRA.No new downsampling method is claimed.
Zhang et al., 2025Human enhancer assay comparisons demonstrate processing-dependent results and discuss DNA-based coverage filtering.This is a different, plant-promoter resource and a controlled estimator comparison.
Source-author 2023 attachmentCorrectly parsed original-estimator measurements were already publicly available.Those values independently validate our reconstruction; their correction is not claimed as novel.
This reanalysisQuantified rank instability across twelve plant-promoter libraries, six replicate-pair comparisons and fixed CNN predictions, with a reproducible target-provenance audit.A potentially useful original empirical resource audit. Targeted literature and repository searches found no earlier report of this exact result; expert review can still identify missed work.

Limits and publication scope

Reproduce and inspect

Reproduction guide and input versions · All depth metrics · Replicate agreement · Frozen CNN metrics · Prediction-mapping exclusions · Independent attachment verification · Full discovery journal.

All raw downloads are versioned or checksummed. The executable workflow uses one CPU thread and does not need a GPU. Figures are also available as standalone SVG and PDF files. No external author contact or publication submission has been made.