Methods and systems for determining proportions of distinct cell subsets
Inventors
Newman, Aaron M. • Alizadeh, Arash Ash
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-12584180-B2
Publication Date
2026-03-24
Expiration Date
Abstract
Methods of deconvolving a feature profile of a physical system are provided herein. The present method may include: optimizing a regression between a) a feature profile of a first plurality of distinct components and b) a reference matrix of feature signatures for a second plurality of distinct components, wherein the feature profile is modeled as a linear combination of the reference matrix, and wherein the optimizing includes solving a set of regression coefficients of the regression, wherein the solution minimizes 1) a linear loss function and 2) an L2-norm penalty function; and estimating the fractional representation of one or more distinct components among the second plurality of distinct components present in the sample based on the set of regression coefficients. Systems and computer readable media for performing the subject methods are also provided.
Core Innovation
The invention relates to obtaining a biological sample from a subject in which the biological sample comprises a tumor sample with a first plurality of distinct cell subsets, extracting ribonucleic acid (RNA) molecules, and assaying the RNA molecules to generate a feature profile m. The feature profile m represents an RNA transcriptome of cells in the biological sample and comprises combinations of features associated with the first plurality of distinct cell subsets.
The invention further optimizes, by a computer processor, a support vector regression between the feature profile m and a reference matrix B of feature signatures for a second plurality of distinct cell subsets. The feature profile m is modeled as a linear combination of the reference matrix B, and the optimizing solves for a set of regression coefficients f of the support vector regression by minimizing a linear loss function and an L2-norm penalty function.
In a specific reference-matrix setting, the reference matrix B comprises an LM22 signature matrix of relative gene expression values across 22 leukocyte subsets. The LM22 signature matrix is configured to represent feature signatures across 22 leukocyte subsets, and optimization against this reference yields support vector regression coefficients f that correspond to the second plurality of distinct cell subsets.
The invention also includes determining an empirical significance value by generating a random feature profile m* and preserving the Euclidean norm of the feature profile used for fitting. A second support vector regression is performed to obtain regression coefficients, a reconstituted feature profile is generated, and a difference measurement is computed between the random and reconstituted profiles.
Claims Coverage
The identified claim coverage includes four inventive features. The core coverage centers on SVR-based optimization linking an RNA feature profile to a reference signature matrix, with LM22 usage, non-negative coefficient post-processing, and empirical significance determination.
Support vector regression deconvolution from RNA feature profile to reference signatures
A method including extracting RNA from a biological tumor sample, assaying to generate a feature profile m representing an RNA transcriptome, and optimizing by a computer processor a support vector regression between the feature profile m and a reference matrix B of feature signatures, wherein the feature profile m is modeled as a linear combination of the reference matrix B and optimizing solves for regression coefficients f by minimizing a linear loss function and an L2-norm penalty function.
LM22 reference matrix for leukocyte subset feature signatures
The method wherein the reference matrix B comprises an LM22 signature matrix of relative gene expression values across 22 leukocyte subsets.
Significance value via random Euclidean-norm-preserving feature profiles
The method further comprising generating a random feature profile with preserved Euclidean norm, fitting a second support vector regression to the reference matrix B, computing a reconstituted feature profile from regression coefficients, measuring a difference between the random and reconstituted profiles, and deriving a significance value from a distribution of difference measurements across plural iterations.
Non-negative coefficient post-processing to estimate relative proportions
The method further comprising setting negative regression coefficients to zero and normalizing remaining non-negative regression coefficients to generate estimated relative proportion values for one or more distinct cell subsets.
The claims cover modeling an RNA-derived feature profile as a linear combination of a reference signature matrix using support vector regression with a linear loss and an L2-norm penalty. Additional features include LM22 reference usage, post-processing of coefficients to estimate relative proportions, and empirical significance determination using randomization with Euclidean norm preservation and iterative difference measurements.
Stated Advantages
Not explicitly described in patent.
Documented Applications
Not explicitly described in patent.
Interested in licensing this patent?