RiSPICE

Genome-wide association studies in rice have mapped loci for yield, drought tolerance, and dozens of other agronomic traits, but the loci themselves are rarely actionable: each associated interval typically holds many variants in linkage disequilibrium, most of them outside coding sequence, with no protein-level consequence to reason about. RiSPICE — Rice SNP Prioritization Integrating Chromatin Effects — instead asks which variant in an interval is most likely to perturb regulatory chromatin. It was developed by the Bioinformatics Lab at De La Salle University in Manila.

The method adapts an existing DNA language model rather than pretraining a new one. RiSPICE fine-tunes DNABERT-2 into a rice-specific multi-label predictor of chromatin state, then derives allele-level effect scores by comparing its predictions for a reference sequence against the same sequence carrying the alternate allele. That places it in a different niche from plant regulatory-genomics backbones such as Deep-Plant and OneGenome-Rice, which pretrain from scratch: RiSPICE is a parameter-efficient adaptation plus a scoring and ranking pipeline, aimed squarely at post-GWAS interpretation.

A single fixed fine-tuned model is applied across three unrelated case studies in the preprint, without per-locus retraining. The work is a bioRxiv preprint and has not yet been peer reviewed. The code is MIT-licensed, and the trained weights are released on HuggingFace as LoRA adapters over the DNABERT-2 checkpoint rather than as a standalone genome model.

#Key Features

• Multi-label chromatin prediction: One forward pass predicts 12 chromatin features — ATAC-seq accessibility plus 11 histone marks — resolving a variant's predicted consequences mark by mark instead of collapsing them into a single activity score.
• Allele-level effect scoring: Per-Feature Scores are the log-odds difference between reference and alternate allele probabilities for each feature; they aggregate into an Overall Score via the Euclidean norm, which is what variants within a locus are ranked by.
• In silico saturation mutagenesis: Every possible substitution across a chosen window can be scored to map local regulatory sensitivity — the mode used to locate a high-impact sub-region of the DEP1 promoter.
• Parameter-efficient adaptation: Fine-tuning uses LoRA, with adapters distributed for 500 bp, 750 bp, and 1000 bp contexts and the variant centered in the window, keeping training and distribution costs low.
• Reproducible pipeline: The repository runs from a VCF through sequence generation, chromatin prediction, scoring, ranking, and visualization, with a walkthrough that reproduces the OsHAK1 case study.

#Technical Details

The backbone is DNABERT-2-117M, a byte-pair-encoded multi-species transformer encoder, extended with a multi-label sequence-classification head and adapted with LoRA on rice histone-mark ChIP-seq and chromatin-accessibility ATAC-seq data. The 12 predicted features are ATAC-seq accessibility, H3K4ac, H3K4me2, H3K4me3, H3K9ac, H3K9me1, H3K23ac, H3K27ac, H3K27me3, H3K36me3, H4K12ac, and H4K16ac. Sequences are drawn against the Nipponbare reference genome, IRGSP-1.0 assembly (GenBank accession GCA_001433935.1). The released artifacts are a base classification repository plus three sequence-length-specific LoRA adapters, so DNABERT-2 remains a required dependency rather than being absorbed into a self-contained checkpoint.

Validation rests on three case studies rather than a leaderboard. In a post-GWAS analysis of heat tolerance, prioritized variants overlapped promoters of candidate genes independently identified through combined GWAS and transcriptomic analyses. For the high-yield gene DEP1, promoter variant prioritization combined with saturation mutagenesis identified a localized region enriched for high-impact mutations that overlapped predicted transcription factor binding sites. For the drought-associated potassium transporter OsHAK1, the top-ranked variant was predicted to affect chromatin features in a pattern consistent with the reported co-occurrence of H2B ubiquitination and H3K4 methylation in plants.

#Applications

The primary audience is rice functional genomics and molecular breeding. RiSPICE narrows a linkage-disequilibrium block to a small set of candidate regulatory variants worth taking into the lab, whether for CRISPR editing, reporter assays, or marker development. The saturation mutagenesis mode also supports promoter dissection independently of any GWAS, mapping which positions in a regulatory region the model treats as sensitive. Because the pipeline consumes a standard VCF plus the reference genome, it slots into existing rice variant-calling and post-GWAS workflows without bespoke data preparation.

#Impact

RiSPICE shows that a released multi-species DNA language model can be turned into a crop-specific regulatory-variant tool with modest compute — relevant to the many agriculturally important species whose epigenomic data are too sparse to justify pretraining a dedicated foundation model. Its concrete contribution is an openly licensed pipeline that converts chromatin predictions into ranked variant lists, the step most sequence-to-function models leave to the user. The limitations are equally clear: training and evaluation used the Nipponbare cultivar, so transfer to the indica varieties and landraces carrying much agronomic variation is unestablished; predicted effects are correlative, and no prioritized variant has been experimentally validated, with support instead drawn from previously published GWAS and transcriptomic analyses; and the work awaits peer review.