Synthetic Gene Expression Profiles via Fourier-Domain Sampling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for generating gene expression profile datasets are resource-intensive, time-consuming, and reliant on hyperparameter tuning, often failing to capture individual cell nuances and requiring substantial cell counts, especially in scenarios with limited sample sizes, particularly for rare diseases.

Innovation Solution

A method involving Discrete Fourier Transform (DFT) and Inverse Discrete Fourier Transform (IDFT) is used to generate synthetic gene expression profiles by modifying complex components of real-world datasets, preserving cell-specific variations and avoiding the need for extensive training, thus producing diverse and reliable datasets.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If manifold-based or neural network-based approaches are used to generate synthetic gene expression datasets, then data availability is improved, but computational resources and time consumption increase significantly

Engineering Contradiction:
Improvegene expression dataset sizeVSAvoidcomputational resource consumption
Core Design Contradiction:
Quantity of substanceVSUse of energy by moving object

Solution Approach 1:

The patent extracts only the essential statistical properties (mean, variance, covariance) from the original gene expression data to construct synthetic datasets. By separating and utilizing only these critical statistical features rather than the complete high-dimensional data structure, the method generates synthetic data with minimal computational overhead while preserving biological variability.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The invention employs simple statistical distributions (normal, Poisson, negative binomial) that are computationally inexpensive and can be rapidly sampled. These lightweight statistical models replace complex deep learning architectures, enabling quick generation of synthetic datasets without requiring extensive computational resources or hyperparameter tuning.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

2Adaptability or versatility

If generative models like GANs or VAEs are applied to overcome data scarcity, then dataset diversity is improved, but training time and resource requirements increase

Engineering Contradiction:
Improvedataset diversityVSAvoidmodel training time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The method uses statistical distributions that inherently capture the necessary data characteristics and variability without requiring external training processes. The statistical models self-adjust to match the input data's distribution properties, eliminating the need for separate training phases and hyperparameter optimization that characterize generative models like GANs and VAEs.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent changes the approach from learning complex data representations through training to directly parameterizing synthetic data generation using statistical distribution parameters (mean, variance, covariance). By transforming the problem from representation learning to statistical sampling, the method achieves dataset diversity without the time-consuming training process.

Inventive Principle:
Principle #35Parameter changes

3Device complexity

If dimensionality reduction strategies are used in manifold-based methods, then computational complexity is reduced, but individual cell variations are lost

Engineering Contradiction:
Improvecomputational complexityVSAvoidindividual cell variations
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The patent applies partial dimensionality reduction by focusing only on the essential statistical moments (first moment for mean, second moment for variance/covariance) needed to capture cell variations. Rather than reducing to low-dimensional manifolds that lose information, the method selectively processes only the critical statistical components that preserve individual cell characteristics.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The method preserves local variations in the data by ensuring that statistical parameters are estimated and applied at the appropriate granularity. By calculating mean, variance, and covariance for relevant gene sets and applying them locally during synthetic data generation, the approach maintains individual cell variations without requiring complex high-dimensional manifold structures.

Inventive Principle:
Principle #3Local quality

4Reliability

If extensive gene expression profile datasets are collected for AI/ML model training, then model performance is improved, but research costs and time requirements increase

Engineering Contradiction:
Improvemodel performanceVSAvoidresearch efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent creates synthetic copies of gene expression data that preserve the statistical properties and biological variability of real datasets. These synthetic copies can be used for AI/ML model training and validation, providing sufficient data volume and diversity to improve model performance without the need to collect and process extensive additional real patient samples.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4668278A1Method, system and apparatus for generating and/or augmenting gene expression profile datasets
Publication Date: 2025.12.24 SANOFI SA(FR)
  • EP4668278A1 patent drawingFigure 1A
  • EP4668278A1 patent drawingFigure 1B
  • EP4668278A1 patent drawingFigure 2

AI summary

Methods, systems and apparatus are described for generating synthetic gene expression profile datasets for downstream biological/pharmacological analysis, processing and/or applications. Generating synthetic gene expression profile datasets includes: receiving a gene expression profile dataset comprising data representative of at least one gene expression profile; computing a complex gene expression profile dataset based on applying a discrete Fourier transform to data representative of each gene expression profile of the received gene expression profile dataset; sampling the complex gene expression profile dataset based on using a statistical distribution to sample and modify a predetermined number of one or more components of each complex gene expression profile for generating a plurality of synthetic complex gene expression profiles; computing a real-valued synthetic gene expression profile dataset based on applying an inverse discrete Fourier transform to the synthetic complex gene expression profile dataset; and outputting data representative of the real-valued synthetic gene expression profile dataset for downstream biological/pharmacological analysis, processing and/or applications.