Cellular Activity Quantification Using Gene Module Activation Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-throughput sequencing data, particularly single-cell RNA sequencing, is plagued by noise such as batch effects and dropout events, leading to poor prediction accuracy for treatment responses, making effective therapy recommendations impossible.
Innovation Solution
A computational pipeline using graph convolutional neural networks and collaborative filtering to combine gene regulatory networks with gene annotations, creating gene modules and activation vectors to represent cellular activity, which is then used for patient stratification through machine learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high throughput sequencing data is used to analyze gene expression at single-cell resolution, then detailed insights on tumoral heterogeneity can be obtained, but the data is affected by noise such as batch effects and dropout events leading to poor prediction accuracy
Solution Approach 1:
The patent introduces an intermediary representation layer (activation vectors) that mediates between the noisy raw sequencing data and the prediction model. This intermediate representation is constructed by combining cell embeddings with gene module centroids, filtering out noise while preserving biologically relevant signals for accurate treatment response prediction
Solution Approach 2:
The patent replaces direct use of noisy measurement data with a computational reconstruction approach. Instead of using raw expression values directly for prediction, the system substitutes them with reconstructed activation vectors that represent underlying biological states, effectively replacing the mechanical measurement process with an information-theoretic reconstruction
2Productivity
If raw high throughput sequencing data is used directly for predictions, then the analysis process remains simple, but treatment response predictions are poor making effective therapy recommendations impossible
Solution Approach 1:
The patent performs preliminary actions by pre-computing gene module centroids from domain knowledge and pre-processing sequencing data into cell embeddings. These preliminary computations create a foundation that simplifies the final prediction process while dramatically improving therapy recommendation effectiveness
Solution Approach 2:
The patent segments the complex data processing pipeline into distinct functional modules: (1) gene module identification, (2) cell embedding generation, (3) activation vector construction, and (4) prediction modeling. This segmentation allows each component to be optimized independently while maintaining overall system effectiveness
3Reliability
If multiple modalities of domain knowledge are combined to create gene modules, then prediction accuracy improves, but the computational pipeline becomes more complex
Solution Approach 1:
The patent merges multiple modalities of domain knowledge (gene regulatory networks, gene annotations, and expression data) into a unified representation system. By combining these diverse data sources through the activation vector framework, the system achieves higher prediction reliability while managing computational complexity through integrated processing
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a computer-implemented method and to a processing system for quantifying cellular activity from high throughput sequencing data. The method comprises generating a multimodal knowledge graph by combining a gene regulatory network, GRN, with gene annotations coming from domain knowledge, wherein the nodes of the multimodal knowledge graph are genes and wherein the gene annotations enrich the relations among the genes; creating a number of gene modules, GMs, by clustering the embeddings of the genes of the GRN; embedding the samples of the sequencing data into the multimodal knowledge graph; and generating, for each sample of the sequencing data, an activation vector in which the respective sample is expressed as the distances between the embedding and the centroids of each of the number of GMs.