RiSPICE (Rice SNP Prioritization Integrating Chromatin Effects) is a method for prioritizing non-coding rice variants by predicting their impact on chromatin features using a fine-tuned DNA language model.
This repository accompanies the paper:
Prioritizing Non-coding Variants in Rice GWAS Loci with a Chromatin-Informed DNA Language Model
It includes:
- Data preprocessing and dataset generation
- Fine-tuning DNABERT-2 with LoRA
- Model evaluation
- Variant effect scoring
- Downstream case study analyses
RiSPICE predicts chromatin feature changes caused by non-coding variants and ranks variants according to their predicted regulatory impact. The repository includes:
- Data preprocessing and dataset generation
- Fine-tuning DNABERT-2 with LoRA
- Model evaluation
- Variant effect scoring
- Downstream case study analyses
RiSPICE builds upon DNABERT-2-117M. We provide the RiSPICE Base model, which contains a trained multi-label chromatin feature prediction head, together with LoRA adapters optimized for different input sequence lengths.
- RiSPICE Base:
paolomanlapaz/rispice-base - 1000 bp Adapter:
paolomanlapaz/rispice-1000bp - 750 bp Adapter:
paolomanlapaz/rispice-750bp - 500 bp Adapter:
paolomanlapaz/rispice-500bp
Download the RiSPICE Base model together with the desired LoRA adapter before running inference.
hf download <repo_id> --local-dir <local_directory>If you use RiSPICE in your research, please cite the associated paper.
This project is released under the MIT License. See the LICENSE file for details.