Personalized genomes for DL models supporting data

Archive of models and data associated with our manuscript "Training deep learning models on personalized genomic sequences improves variant effect prediction".

Code for training and benchmarking LCL models is available at https://github.com/Danko-Lab/clipnet_ablation, whereas code for training and benchmarking K562 models is available at https://github.com/Danko-Lab/clipnet_k562/.

Model files & metadata:

• n{i}_run{j}.tar
  • CLIPNET LCL models trained on i individuals
• subsample_individuals_ids.tar
  • text files containing lists of the individuals used to train the above models.
• reference_models.tar
  • CLIPNET LCL model trained on data from 67 PRO-cap libraries, but using hg38 sequences instead of personal genomes.
• clipnet_k562_reference.tar
  
  • hg38-trained model described above transfer learned to K562.

Benchmark data:

• across_loci_metrics.tar
  • benchmarks of LCL models at predicting transcription initiation at individual CREs within the genome
• qtl_metrics.tar
  • benchmarks of LCL models at predicting differences in transcription initiation between individuals at initiation QTLs
• k562_data.tar
  
  • benchmarks of the reference-trained K562 model and one transferred over from the personalized CLIPNET model on MPRA data from https://www.biorxiv.org/content/10.1101/2024.05.05.592437v1