viRNAtrap Deep Learning Model for Divergent Virus Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for identifying novel and divergent viruses in cancer transcriptomes are limited by their reliance on sequence similarity to known viruses, making it challenging to detect new cancer-associated viruses, especially with short RNA sequencing reads.
Innovation Solution
The development of the viRNAtrap system, which employs a deep learning model to identify viral RNAseq reads and assemble viral contigs without alignment, allowing for the detection of divergent viruses by distinguishing viral sequences from human sequences and characterizing their expression in cancer tissues.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If sequence similarity-based methods are used to identify viruses, then known viruses can be detected, but novel and divergent viruses cannot be identified
Solution Approach 1:
The method segments the viral identification process into two distinct phases: (1) alignment-based detection for known viruses using BLAST against reference databases, and (2) alignment-free deep learning-based detection for novel/divergent viruses using the VIRAID model. This segmentation allows each method to be optimized for its specific purpose, resolving the contradiction between detecting known viruses and discovering novel viruses.
Solution Approach 2:
The patent introduces an intermediary approach by using deep learning models (VIRAID) that process raw sequencing reads without requiring alignment to reference genomes. This intermediary method acts as a bridge between traditional alignment-based detection and novel virus discovery, enabling the system to identify divergent viruses that lack sequence similarity to known viruses while maintaining the ability to detect known viruses through the alignment-based pipeline.
2Measurement precision
If alignment-based methods are used, then viral sequences can be identified with known references, but the process is computationally intensive and time-consuming
Solution Approach 1:
The patent applies preliminary action by pre-training the deep learning model (VIRAID) on a comprehensive dataset of viral sequences before deployment. This pre-training phase allows the model to learn viral sequence patterns in advance, so that during actual virus identification, the model can rapidly process sequencing reads without requiring computationally intensive alignment operations, thus reducing analysis time while maintaining accuracy.
Solution Approach 2:
The patent substitutes the mechanical alignment process (BLAST, BLAT, or other sequence alignment tools) with a deep learning-based classification system. Instead of mechanically comparing each read against reference genomes, the VIRAID model uses learned features and patterns to rapidly classify reads as viral or non-viral, dramatically reducing computational time and resource requirements while maintaining high identification accuracy.
3Productivity
If short RNA sequencing reads are used, then high throughput data can be generated, but viral detection sensitivity decreases for divergent viruses
Solution Approach 1:
The patent addresses the limitation of short reads by introducing a new dimension of analysis through deep learning feature extraction. Instead of relying solely on linear sequence alignment (one-dimensional), the VIRAID model processes reads in multiple dimensions by considering k-mer compositions, positional features, and learned sequence patterns, enabling accurate viral detection even from short reads that lack sufficient length for traditional alignment methods.
Solution Approach 2:
The patent changes the detection parameters by using deep learning models that can identify viral signals in short reads through learned features rather than requiring long contiguous matches. The model adjusts its sensitivity parameters by using probabilistic classification thresholds and can detect viral reads even when they represent only a small fraction of the total sequencing data, thereby maintaining high detection sensitivity with high-throughput short read sequencing.
Data Source
AI summary
Systems and methods for identifying viral sequences in a subject of a species are provided. Sequence reads not associated with a reference genome of the species are obtained from a biological sample from the subject. At least a portion of each respective sequence read is encoded into a corresponding vector representing all or a portion of the sequence of the respective sequence read, thereby obtaining a plurality of vectors. Each sequence read is assigned a corresponding scalar model score by inputting a vector, in the plurality of vectors, corresponding to the sequence read into a model. Those sequence reads having a corresponding scalar model score that satisfies a first threshold score are selected as contig seeds. The plurality of sequence reads are aligned to these contig seeds through common k-mer sequences thereby forming a plurality of contigs which, in turn, are used to identify viral sequences in the subject.