Context-Free Cell Type Deconvolution Using Pseudobulk Mixtures
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for deconvolving cell types in spatial transcriptomics require reference profiles that are not always available or appropriate, leading to challenges in interpreting gene expression patterns due to varying cell type mixtures and transcriptional differences.
Innovation Solution
A context-free model is trained using pseudobulk training mixtures formed from aggregated abundance datasets across multiple data stores, allowing for the determination of cell type fractions without relying on contextualized reference data, utilizing a multiple layer fully connected neural network to calculate cell type fractions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If reference-based deconvolution methods are used, then deconvolution can be performed with available reference data, but accuracy deteriorates when reference profiles are not available or are confounded by experimental or disease-related factors
Solution Approach 1:
The invention extracts and removes the dependency on external reference profiles from the deconvolution process. By training a neural network model on pseudobulk mixtures generated from the training data itself, the method eliminates the need for separate reference datasets, thereby maintaining accuracy while improving adaptability to samples without appropriate references.
Solution Approach 2:
The invention creates synthetic copies of cell type compositions through pseudobulk mixtures. By randomly sampling and combining cells from different cell types to generate artificial bulk samples with known compositions, the method creates training data that mimics real spatial transcriptomics data without requiring external references, enabling the model to learn accurate deconvolution patterns.
2Measurement precision
If matched reference samples are collected for reference generation, then deconvolution accuracy improves, but the complexity and difficulty of the process increases
Solution Approach 1:
The invention enables the system to generate its own training data and references autonomously. By creating pseudobulk mixtures from the training data itself and using these to train the deconvolution model, the method eliminates the need for separate reference sample collection processes, thereby maintaining measurement precision while significantly reducing process complexity.
Solution Approach 2:
The invention performs preliminary generation of training data in the form of pseudobulk mixtures with known compositions. By pre-generating these synthetic reference profiles from the training data before actual deconvolution, the method eliminates the need for complex post-hoc reference matching while ensuring accurate cell type fraction determination.
3Adaptability or versatility
If general scRNA-Seq atlases are used as references, then deconvolution can be performed without matched samples, but accuracy deteriorates due to transcriptional differences from experimental or disease-related factors
Solution Approach 1:
The invention adapts the reference profiles to match the specific local characteristics of the training data. By generating pseudobulk mixtures from the actual training samples and using these to train the model, the method creates references that are locally optimized for the specific experimental conditions and disease states present in the data, thereby maintaining high accuracy while achieving versatility across different sample types.
Data Source
AI summary
Systems and methods for training a context-free model to determine cell type fractions are provided. A training set is obtained comprising, for each of a plurality of data stores, for respective each cell represented in the data store, a dataset comprising abundance values for cellular constituents associated with the respective cell. Pseudobulk training mixtures are formed from the training set. For each mixture, the abundance value for each cellular constituent is averaged across abundance datasets of the cells represented by the respective mixture thereby forming an averaged abundance dataset for the mixture. For each mixture, a corresponding averaged abundance dataset is inputted into the model thereby obtaining a respective plurality of calculated cell type fractions, each fraction for a different cell type. Model parameters are adjusted based on differences between calculated cell type fractions and mixture fraction ratios for each unique cell type in the respective mixture.