4D Nucleome Hackathon

Project 3: In silico variant prioritization using sequence-based predictive models

​

March 21st, 2024

Seattle, WA

Shu, Katie, Julia, Yang, Leo

SuPreMo generates mutated sequences for ISM

Gjoni and Pollard, bioRxiv (2023)

​

Akita predicts 3D genome folding from DNA sequence alone

Fudenberg et al, Nature Methods (2020)

Our goals for the Hackathon

1. Incorporate SuPreMo with additional models (Leo, Yang)
1. Enformer
2. ExPecto

​

• Add additional functionality to SuPreMo (Julia, Katie, Shu)
• Custom reference fasta
• P-values for disruption scores
• Additional cell types
• Weighted disruption scores
• Sequence mutagenesis

Implementing SuPreMo-Enformer

Testing SuPreMo-Enformer

Data: tumor SV called by Manta

Method: Enformer applied to investigate SVs’ influence on CTCF ChIP assay in HFF cells

*BND is forced to have a length of 30000 for the visualization purpose

Closer look of the representative case

González-Rico, Francisco J., et al. "Alu retrotransposons modulate Nanog expression through dynamic changes in regional chromatin conformation via aryl hydrocarbon receptor." Epigenetics & Chromatin 13 (2020): 1-13.

Implementing SuPreMo-ExPecto

• ExPecto (Zhou et al 2018)
• DNA sequence -> Gene expression
• Input
• SNP, small indel (paper)
• SNP, indels, SV (eg, BND)
• Variable shift size
• A JSON for customized parameters
• Output
• Predicted variant scores for all genes within +- 20kb of variants
• 20min for 115,452 prediction using 16 CPU threads

{

"beluga_model_file": "/home/yang/Project/SuPreMo/ExPecto_model/resources/deepsea.beluga.pth",

"is_cuda": "False",

"model_list_file": "/home/yang/Project/SuPreMo/ExPecto_model/resources/modellist",

"gene_tss_file": "/home/yang/Project/SuPreMo/ExPecto_model/resources/geneanno.hg38.sorted.bed.gz",

"model_dir": "/home/yang/Project/SuPreMo/ExPecto_model",

"model_name": ["Fibroblast of Lung", "Small Intestine", "Small Intestine Terminal Ileum"],

"threads":16,

"n_features": 2002,

"fixed_dist": 100,

"maxshift": 800,

"split_flag": 1,

"split_index": 0,

"split_fold": 10,

"verbose_level": 0

}

Implementing SuPreMo-ExPecto - Cont’d

• ExPecto (Zhou et al 2018)
• DNA sequence -> Gene expression
• Input
• SNP, small indel (paper)
• SNP, indels, SV
• Variable shift size
• A JSON for customized parameters
• Output
• Predicted variant scores for all genes within +- 20kb of variants
• 20min for 115,452 prediction using 16 CPU threads

Incorporating personalized genome into SuPreMo

• provide VCF with patient specific SNVs
• Incorporate patient SNVs into personalized reference genome
• Use personalized reference genome when scoring individual SVs

Adding P-values for disruption scores

• Take set of control scores generated from control variants
• Compare scores of target variants to control scores to determine significance

Calculating SuPreMo scores with functional weights

Weights examples

• Gene promoters
• ATACseq peaks
• H3K4me3 HiChIP

​

SuPreMo implementation

Regions of interest (--roi):

• ‘genes’
• Path to file with roi

Weights (--roi_weights):

• Values 0-1
• 0: no weights applied
• 0.5: half of the score is driven by roi
• 1: only roi are scored

Weighted scores highlight variants that disrupt relevant regions

1.283e-6

1.282e-6

0.052

0.087

0.065

0.049

Disruption track

ATACseq weights

Sequence mutagenesis options with SuPreMo

Parameters: (1) % GC deletion (2) mutate flanking regions of the variant

Future parameters: (1) Apply to ALT/ALT-mutated sequences

SuPreMo-Akita scores

Variant (CTCF site) GC and nucleotide shuffling mutagenesis

TFBS mutagenesis options with SuPreMo

Parameters: (1) INV/shuffling of TFBS (2) mutate flanking regions of the variant

Future parameters: (1) DEL/DUP of TFBS (2) Apply to ALT/ALT-mutated sequences

SuPreMo-Akita scores

TFBS INV and shuffling mutagenesis

Conclusions

• SuPreMo generates mutated sequences for ISM, which can be directly input into the following models:
• Previously: Akita, DeepSEA
• New: Enformer, ExPecto
• We can obtain disruption scores to quantify the effect of individual variants on genome folding, chromatin states, gene expression, etc

​

• SuPreMo is extremely customizable with new options to:
• Select cell types for predictions
• Add additional sequence perturbations
• Use a custom reference sequence
• Weight disruption scores based on regions of interest
• Generate p-values for disruption scores

THANK YOU!

Acknowledgements

Katie Pollard (UCSF)

Peter J. Park (HMS)

Alexey I. Nesvizhskii (UMich)

Jian Ma (CMU)

​