Multiple Instance Fine-mapping (MIFM)

"Multiple instance fine-mapping: predicting causal regulatory variants with a deep sequence model" preprint

This repository contains:

• a script for training an MIFM model: src/train.py
• weights for an already trained model used for the experiments in the paper: checkpoints/a7a0q7ar.ckpt
• a script for making predictions using a pretrained model: src/load_and_predict.py
• example files for the prediction script: example_data

Predicting using a pretrained model:

To predict based on genomic coordinates (chromosome and bp location):

python src/load_and_predict.py \
  --out_file=out.txt \
  --coordinates_file=example_data/variant_coordinates.tsv

To predict directly from DNA sequences (as strings):

python src/load_and_predict.py \
  --out_file=out.txt \
  --sequences_file=example_data/sequences.csv

Reproducing experimental results:

Experiments from the paper can be reproduced using the appropriate Snakemake rules, e.g., the rule all_evaluate_scores_secondary_signals will rerun the Polygenic Risk Scores experiments.

Experimental configuration (e.g., model ID of the trained MIFM model, UKB trait IDs) need to be specified in src/snakemake_config.py