Single Read SNV (GSI mod)

This repository contains modified code of Ultima Genomics single_read_ssnv workflow (https://github.com/Ultimagen/healthomics-workflows/tree/main/workflows/single_read_snv). This workflow runs on Ultima NGS data and is capable of calling small variants (SNVs) in either germline or somatic modes.

The Single Read SNV (SRSNV) pipeline is a read-centric de-noising framework, developed to overcome the limitations of traditional locus-centric variant calling, particularly in scenarios where rare mutations may be supported by only a single read. These rare mutations need to be distinguished from artefactual SNVs, which can derive from sequencing, library or alignment errors. To achieve this, we employed a supervised machine learning model trained to classify actual SNVs (labelled True or TP) from noise (False or FP). First, a comprehensive dataset capturing every candidate SNV is generated, along with a rich suite of annotations that describe sequencing quality, local sequence motifs, fragment-specific features, and locus-specific information. Randomly selected bases in the data matching the reference genome are collected and annotated as True, while low VAF (≤5%) SNVs in high-coverage (≥20×) regions (SNVs with low support, indicating they are likely to be artifacts) are annotated as False SNVs. Using these curated sets, we train an XGBoost classifier to robustly distinguish between true and artifactual SNVs. Once trained, the classifier assigns a calibrated quality score to each SNV in the input CRAM, providing a precise estimate of the residual error rate. To avoid overfitting, an ensemble of models (3) are trained on different sets of chromosomes and applied using a cross-validation scheme.

Please treat this repository as a work in progress project as not all of it's elements have been thoroughly tested in production environment.

Strusture of the workflow

The diagram below outlines the connections between tasks and resource modules, also see below

Outputs

Output	Type	Description
featuremap	File	Feature Map output file in VCF format
featuremap_index	File	Feature Map index file
featuremap_random_sample	File?	Feature Map random_sample file in VCF format
featuremap_random_sample_index	File?	Feature Map random_sample index
downsampling_rate	Float	Reported downsampling rate used by CreateFeatureMap.downsampling
snv_qualities_assigned	Boolean	Flag which shows if snv qualities can be assigned
used_self_trained_model	Boolean	Flag which shows if snv qualities assigned
raw_filtered_featuremap_parquet	File?	Filtered featuremap parquet file
random_sample_trinuc_freq_stats	File?	File with trinucleotide frequencies, made by CreateFeatureMap
featuremap_df	File?	featuremap_df_output
application_qc_h5	File?	application qc_h5 output
report_html	File?	HTML report
srsnv_metadata_json	File?	srsnv metadata json
model_files	Array[File]?	model files
md5_checksums_json	File?	Optional JSON file produced with MergeMd5sToJson

Setting up to run in production environment

Single Read SNV uses multiple resources which may be downloaded from Ultima Genomics website(s) and Amazon buckets. A simple bash script is included, although in the nearest future we may switch to using single_read_snv resources wrapped in modules (with Modulator).