Deep Learning to Separate Somatic and Artificial Mutations in FFPE DNA
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for analyzing FFPE samples in DNA sequencing face challenges in distinguishing between somatic mutations and artificial mutations, leading to inaccuracies in cancer genome analysis, which hinders precise cancer treatment predictions.
Innovation Solution
A deep-learning neural network model is employed to differentiate between somatic and artificial mutations by analyzing FFPE samples, using cosine similarity and various data features such as read counts and sequence depths, to improve prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional methods are used to analyze FFPE samples, then the analysis can be performed on stored samples, but the accuracy of mutation detection deteriorates due to artificial mutations induced by formalin fixation
Solution Approach 1:
The patent introduces an intermediary classification system that mediates between the FFPE sample analysis process and the mutation detection process. This system uses machine learning models to classify mutations as either somatic or artificial, allowing accurate detection to proceed despite the presence of formalin-induced artifacts. The intermediary classifier processes mutation calls and annotates them with reliability scores, enabling downstream analysis to distinguish true somatic mutations from artifacts.
Solution Approach 2:
The patent changes the parameter space for mutation analysis by introducing multiple classification dimensions: somatic vs. artificial classification, confidence score thresholds, and mutation type categorization. By transforming the binary mutation detection problem into a multi-parameter classification problem, the system can adjust sensitivity and specificity parameters to optimize detection accuracy while maintaining compatibility with FFPE samples.
2Object-generated harmful factors
If existing mutation removal methods are applied to FFPE data, then artificial mutations can be reduced, but most actual somatic mutations are also removed
Solution Approach 1:
The patent applies local quality by treating different mutation characteristics differently through location-specific classification. The machine learning model analyzes local features around each mutation site (such as read depth patterns, strand bias, and neighboring base characteristics) to determine whether a mutation is somatic or artificial. This localized analysis allows the system to preserve true somatic mutations while removing artifacts, as each mutation is evaluated on its specific local properties rather than applying a uniform filter.
Solution Approach 2:
The system implements feedback through an iterative classification process where the machine learning model continuously refines its predictions based on training data and performance metrics. The classifier receives mutation calls, predicts their origin (somatic or artificial), and provides feedback in the form of confidence scores and classifications. This feedback loop allows the system to learn from previous analyses and improve its ability to distinguish true mutations from artifacts over time.
3Measurement precision
If deep learning models are used to classify mutations, then the ability to distinguish somatic and artificial mutations improves, but the computational complexity and resource requirements increase
Solution Approach 1:
The patent segments the computational task into distinct modular components: a machine learning classification model for mutation origin determination, a separate confidence scoring mechanism, and downstream analysis tools. This segmentation allows the deep learning model to be trained independently on curated datasets and then applied efficiently to new samples. The modular architecture reduces computational complexity by separating the heavy lifting (training) from the inference phase, enabling deployment on standard computing hardware.
Data Source
Figure 1~2
Figure 3~4
Figure 5a
AI summary
The present invention provides a method for providing information on the prediction of artificial or somatic mutations, and a device for providing information on the prediction of artificial or somatic mutations, wherein the method for providing information on the prediction of artificial or somatic mutations, which is implemented by a processor, comprises the steps of: receiving medical data or data obtained from biological samples derived from individuals; generating data necessary for distinguishing between artificial and somatic mutations on the basis of the received data; and predicting artificial or somatic mutations in the samples obtained from the individuals based on the generated data by using a neural network model configured to predict artificial or somatic mutations, with the necessary data for somatic mutation distinction inputted thereto.