Neural Network Predicting DNA Accessibility Across Cell Types

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current neural network models for predicting DNA accessibility are limited by their cell-type-specific nature, requiring separate models for each cell type or multi-task predictions, which hinders their broad applicability and integration into broader pathway models.

Innovation Solution

A method involving a neural network configured to process DNA sequence data and gene expression levels from RNA-seq data, using paired data files to generate a DNA accessibility prediction output, allowing for predictions in new cell types without the need for distinct trained models or unique outputs per cell type.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If separate neural network models are trained for each cell type, then prediction accuracy for specific cell types is improved, but device complexity and difficulty of application to new data increase

Engineering Contradiction:
Improveprediction accuracyVSAvoidmodel complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies universality by designing a single neural network model that can predict DNA accessibility across multiple cell types simultaneously. Instead of creating separate specialized models for each cell type, the invention uses one universal model that processes DNA sequence data and predicts accessibility profiles for any target cell type, making the system multi-functional and broadly applicable without sacrificing accuracy

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent applies segmentation by dividing the prediction task into distinct computational components within the neural network architecture. The model segments the input DNA sequence into fixed-size windows and processes each window independently through convolutional layers, allowing the system to handle complex multi-cell type predictions by breaking down the problem into manageable computational units

Inventive Principle:
Principle #1Segmentation

2Adaptability or versatility

If multi-task predictions are implemented in a single model, then broad applicability is improved, but prediction accuracy for specific cell types deteriorates

Engineering Contradiction:
Improvebroad applicabilityVSAvoidprediction accuracy
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent applies local quality by allowing different parts of the neural network to specialize in different aspects of the prediction task. The convolutional layers learn cell-type-specific patterns locally, while the fully connected layers integrate information across cell types. This enables the model to maintain high accuracy for specific cell types while preserving broad applicability across multiple cell types simultaneously

Inventive Principle:
Principle #3Local quality

3Loss of information

If cell-type-specific models are used, then understanding of cell-specific phenomena is improved, but integration into broader pathway models becomes difficult

Engineering Contradiction:
Improvecell-specific informationVSAvoidintegration complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent resolves the integration problem by creating a universal neural network model that outputs predictions for multiple cell types simultaneously. This single model can be integrated into broader pathway models without requiring separate model instances for each cell type, simplifying the integration architecture while preserving cell-specific information through the model's multi-output capability

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3542296B1Methods and systems for predicting DNA accessibility in the pan-cancer genome
Publication Date: 2021.04.14 NANTOMICS LLC
  • EP3542296B1 patent drawingFigure 1
  • EP3542296B1 patent drawingFigure 2
  • EP3542296B1 patent drawingFigure 3

AI summary

Techniques are provided for predicting DNA accessibility. DNase-seq data files and RNA-seq data files for a plurality of cell types are paired by assigning DNase-seq data files to RNA-seq data files that are at least within a same biotype. A neural network is configured to be trained using batches of the paired data files, where configuring the neural network comprises configuring convolutional layers to process a first input comprising DNA sequence data from a paired data file to generate a convolved output, and fully connected layers following the convolutional layers to concatenate the convolved output with a second input comprising gene expression levels derived from RNA-seq data from the paired data file and process the concatenation to generate a DNA accessibility prediction output. The trained neural network is used to predict DNA accessibility in a genomic sample input comprising RNA-seq data and whole genome sequencing for a new cell type.