PEAgent

Sequence-to-single-cell modelling of plant chromatin accessibility and its regulatory grammar across three crop species.

PEAgent trains a per-species convolutional neural network that predicts the single-cell chromatin-accessibility state of every accessible chromatin region (ACR) directly from 1,344 bp of DNA sequence, and uses the trained models to read out the regulatory grammar of plant genomes: a cross-species lexicon of transcription-factor contribution patterns, their spatial co-occurrence and synergy, cross-species transferability, ACR evolutionary constraint, and the effect of natural sequence variants on accessibility.

This repository contains the analysis code accompanying the manuscript. Scripts are organised as self-contained prototypes that document the method for each figure; they read a small set of per-species inputs (atlas, model, peaks, genome) and are pointed at a local data directory through $PEAGENT_DATA_ROOT.

Species

Species	Assembly	Nuclei	ACRs	Tissues	Chromosomes
Soybean (Glycine max)	Wm82.a4.v1	200,732	303,113	10	Gm1..Gm20
Maize (Zea mays)	B73 NAM-5.0	50,639	158,586	6	chr1..chr10
Rice (Oryza sativa)	Nipponbare MSU7	104,029	128,764	9	Chr1..Chr12

Repository layout

peagent-analysis/
├── config/species.py      per-species paths + genome/metadata conventions
├── 01_training/           preprocess ACRs, train the model
├── 02_evaluation/         per-peak AUROC, per-group PCC/AUROC/AUPR, latent NMI
├── 03_lexicon/            attribution maps, motif discovery, meta-pattern lexicon
├── 04_cooccurrence/       base-motif spatial co-occurrence
├── 05_synergy/            in-silico motif-pair synergy
├── 06_cross_species/      cross-species prediction and evaluation
├── 07_acr_evolution/      grass-conserved vs lineage-specific ACR grammar
└── 08_variant_effect/     REF/ALT variant scoring + caQTL benchmark

The analysis scripts import the peagent package (model architecture, sequence/attribution/cell-group helpers, plotting defaults), which is released as a separate package and installed alongside this repository (see Installation).

Analyses

Step	Script(s)	What it does
1. Training	01_training/01_preprocess_acrs.py, 02_train_model.py	Recentre ACRs on 1,344 bp windows, one-hot encode, split by ACR, and train the multitask CNN (Adam, binary cross-entropy, early stopping on validation AUROC).
2. Evaluation	02_evaluation/01_per_peak_auroc.py	AUROC / AUPR per held-out ACR across cell-type groups.
	02_evaluation/02_per_group_metrics.py	Per-cell-group profile PCC, AUROC and AUPR with a shuffled-label null.
	02_evaluation/03_latent_nmi.py	UMAP + Leiden on the learned cell embedding; NMI against annotated cell types.
3. Lexicon	03_lexicon/01_attribution_maps.py	Bulk and cell-type contribution maps from bottleneck ISM.
	03_lexicon/02_modisco_finemo.py	TF-MoDISco motif discovery + Fi-NeMo genome-wide instance calling.
	03_lexicon/03_metapattern_compendium.py	Cross-species meta-pattern clustering (CPM-Leiden).
	03_lexicon/04_classify_lexicon.py	Classify meta-patterns into base / composite / unresolved / noisy.
4. Co-occurrence	04_cooccurrence/01_seqlet_cooccurrence.py	Base-motif instance-pair co-occurrence and spacing summaries.
5. Synergy	05_synergy/01_insilico_synergy.py	In-silico marginalization test for motif-pair synergy.
6. Cross-species	06_cross_species/01_cross_species_predict.py	Score one species' ACRs with another species' model; semantic cell-type matching.
7. ACR evolution	07_acr_evolution/01_acr_conservation.py	Grass-conserved vs rice-specific ACR contribution + conservation/selection intersection.
8. Variant effect	08_variant_effect/01_score_variants.py	REF/ALT Δlogit scoring and fine-mapped-caQTL discrimination benchmark.

References

• Yuan H. & Kelley D. R. scBasset. Nat. Methods (2022).
• Kelley D. R. et al. Basenji. Genome Res. (2018); Kelley D. R. et al. Basset. Genome Res. (2016).
• Shrikumar A. et al. TF-MoDISco (2020); Fi-NeMo motif instance calling.
• Gupta S. et al. TomTom / MEME Suite. Genome Biol. (2007).
• Traag V. A. et al. Leiden clustering. Sci. Rep. (2019).

Citation

This code accompanies:

Yao J. et al. Sequence-based modeling of plant epigenomes reveals cell-type-specific cis-regulatory grammar. Under review (2026).

Software metadata is in CITATION.cff; a Zenodo DOI is minted for the archived release (see docs/RELEASE.md).

License

MIT — see LICENSE.