Variant Quality Classification for Non-Target Reference Calling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing variant calling methods struggle when desired reference genomes are unavailable, leading to numerous false positive calls, especially when using non-homologous or closely-related species as references.
Innovation Solution
Developed methods include random forest classifiers, linear regression models, neural network models, and a unique-mapper score to reduce false positives by comparing variant sets from non-target and pseudo-target reference genomes, followed by a quality classifier using machine learning to identify high-quality variants.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If variant calling is performed using non-target or closely-related species reference genomes, then variant detection can proceed in the absence of a target reference genome, but the number of false positive calls increases significantly
Solution Approach 1:
The patent introduces an intermediary classification system that mediates between the variant calling process and the final variant set. A machine learning classifier (random forest, neural network, or logistic regression) is trained to distinguish true positive variants from false positives by evaluating multiple features including unique-mapper scores, depth metrics, and quality scores. This intermediary classifier resolves the contradiction by filtering out false positives generated when using non-target reference genomes, thereby maintaining reliability while preserving adaptability.
Solution Approach 2:
The patent changes multiple parameters simultaneously to resolve the contradiction: (1) introduces unique-mapper scores as a new parameter to quantify mapping ambiguity, (2) adjusts depth thresholds dynamically based on local genomic context, (3) modifies quality score calculations to account for reference genome divergence, and (4) uses machine learning to optimize the combination of these parameters. These parameter changes enable the system to adapt to different reference genome scenarios while maintaining high accuracy.
2Reliability
If multiple features and machine learning models are used to classify variant quality, then false positive reduction improves, but computational complexity and processing time increase
Solution Approach 1:
The patent segments the variant classification process into distinct modular components: (1) feature extraction module that computes unique-mapper scores, depth metrics, and quality scores, (2) machine learning classification module that evaluates these features, and (3) filtering module that applies decision rules. Each module is independently optimized and can be processed in parallel, reducing overall system complexity while maintaining high false positive reduction capability.
Solution Approach 2:
The patent performs preliminary computations of key features (unique-mapper scores, depth metrics, quality scores) before the main classification step. These pre-computed features are stored and reused across different classification models, avoiding redundant calculations. This preliminary action reduces the computational burden of the machine learning classification while maintaining comprehensive feature evaluation for false positive reduction.
3Measurement precision
If comprehensive feature extraction and multiple classification models are applied, then variant quality assessment improves, but processing time and computational resources increase
Solution Approach 1:
The patent implements a tiered classification approach where not all features and models are applied uniformly to all variants. Instead, variants are first screened using a subset of key features (unique-mapper score and depth metrics), and only variants that pass this initial filter undergo comprehensive classification with all features and multiple models. This partial action approach maintains high precision for critical variants while reducing overall processing time.
Solution Approach 2:
The patent replaces traditional rule-based filtering mechanisms with machine learning models (random forest, neural networks, logistic regression) that can process multiple features simultaneously and make predictions more efficiently. These models learn optimal feature combinations and thresholds from training data, substituting complex mechanical filtering rules with adaptive statistical models that reduce processing time while maintaining or improving measurement precision.
Data Source
AI summary
The technology disclosed relates to variant calling of sequenced reads of a sample of a target species against a reference genome of a pseudo-target species. Low-quality variants are identified as false positive variants that are present in the second set of variants but absent from the first set of variants.