PINS: A Perturbation Clustering Approach for Data Integration and Disease Subtyping

Abstract

Disease subtyping is accomplished by a computer-implemented algorithm that manipulates a first genetic dataset to construct a set of first connectivity matrices. To this set of matrices Gaussian noise is introduced to generate a perturbed dataset. The computer-implemented algorithm assesses which of the set of first connectivity matrices was least affected by introduction of noise and that matric is used to define the optimal clustering. Once the optimal clustering is determined, computer-implemented supervised classification is performed to determine, for a particular patient, with which disease subtype cluster that person's genetic data most closely aligns. Armed with this knowledge, the treatment regimen is specified with much higher likelihood of success.

1 - 18 . (canceled) 
     
     
         19 . A disease subtyping/patient subgrouping computerized method based on the molecular profile of a patient, the computerized method comprising:
 receiving assay results of samples collected from a population of human subjects;   generating a population dataset from the assay results that defines a molecular profile for the population;   receiving an assay result of a patient sample;   generating a patient dataset from the assay result that defines a molecular profile of the patient; and   analyzing the population dataset by:
 a) acquiring the population dataset from the population of human subjects; 
 b) performing an unsupervised cluster analysis on the population dataset to define clusters of the population, wherein each cluster represents a disease subtype or a patient subgroup; 
 c) using machine learning techniques to act predetermined features of each disease subtype or patient subgroup; and 
 d) finding a disease subtype or patient subgroup that represents the closest match to the molecular profile of the patient. 
   
     
     
         20 . The computerized method according to  claim 19 , further comprising:
 analyzing the patient dataset and selecting a disease treatment regimen.   
     
     
         21 . The computerized method according to  claim 19 , wherein the population of human subjects comprises a plurality of human subjects that are stricken with the same disease. 
     
     
         22 . The computerized method according to  claim 21 , wherein the patient has the same disease as the plurality of human subjects and the method comprises finding the disease subtype of the patient's disease. 
     
     
         23 . The computerized method according to  claim 22 , further comprising treating the patient based on the disease subtype. 
     
     
         24 . he computerized method according to  claim 19 , wherein the population of human subjects comprises a plurality of human subjects that have been treated with the same drug. 
     
     
         25 . The method according to  claim 24 , wherein the patient has been treated with the same drug as the plurality of human subjects and the method comprises subgrouping the plurality of human subjects based on responses to the drug. 
     
     
         26 . The computerized method according to  claim 25 , further comprising finding the subgroup of human subjects to which the patient belongs based on the patient's response to the drug. 
     
     
         27 . The computerized method according to  claim 19 ,
 wherein the population of human subjects comprises a plurality of human subjects who have the same disease, wherein the disease manifests differently in groups of the plurality of subjects, and the method further comprises:
 storing the clusters of the population in a data store comprising a non-transitory computer-readable memory; 
 processing the patient dataset using a computer-implemented classifier, the classifier ingesting the clusters in the data store and using the clusters to find which of the clusters represents a closest match to the patient dataset; 
   storing the cluster found to represent the closest match to the patient dataset as patient classification data in a non-transitory computer-readable memory; and   using the patient classification data in the selection of a disease treatment regimen,   wherein the unsupervised cluster analysis comprises:   (a) applying a computer-implemented algorithm to the population dataset to construct a set of first connectivity matrices which are then stored in non-transitory computer-readable memory;   (b) using a computer-implemented algorithm to construct and store in non-transitory computer-readable memory a perturbed dataset by introducing noise into the population dataset;   (c) applying a computer-implemented algorithm to the perturbed dataset to construct a set of perturbed connectivity matrices which are then stored in non-transitory computer-readable memory;   (d) using a computer-implemented algorithm to perform a stability assessment that reads from memory and compares the first connectivity matrices with the perturbed connectivity matrices;   (e) using a computer-implemented algorithm to select from among the first set of connectivity matrices the one matrix whose corresponding perturbed matrix was least affected by the introduction of noise; and   (f) storing the selected one matrix in non-transit computer-readable memory and using a computer-implemented algorithm to construct the clusters.   
     
     
         28 . The computerized method of  claim 27  wherein the assay results provide quantitative biological data selected from the group consisting of mRNA expression, DNA methylation, miRNA expression, protein abundance (proteomics), and metabolic concentrations (metabolomics) and genetic data. 
     
     
         29 . The computerized method according to  claim 19 , further comprising:
 using the population dataset as input, wherein the population of human subjects have the same disease, the population dataset being developed by high-throughput assays selected from the group consisting of mRNA expression, DNA methylation, miRNA expression, proteomic expression, and metabolic concentration,   forming the molecular profile for the population, and performing the following steps:   a) applying k-means to the population dataset with different settings of numbers of clusters, each setting resulting in one clustering of patients;   b) constructing an original connectivity matrix for each clustering of patients;   c) introducing Gaussian noise to the dataset and constructing the perturbed connectivity matrices;   d) performing stability assessment between the original and perturbed connectivity, matrices;   e) selecting one original connectivity matrix that is the least affected by noise; and   f) choosing a cluster of patients that corresponds to the selected connectivity matrix as the optimal subtyping.   
     
     
         30 . The computerized method according to  claim 19 , further comprising:
 generating the population dataset from a plurality of types of molecular data obtained from the population of human subjects each human subject in the population having the same disease,   defining T molecular profiles for the population of human subjects and performing the following steps:   a) applying the algorithm defined in (1)-(6) below to define subtypes of the disease for each data type:   (1) applying a computer-implemented algorithm to the first quantitative biological dataset to construct a set of first connectivity matrices which are then stored in non-transitory computer-readable memory;   (2) using a computer-implemented algorithm to construct and store in non-transitory computer-readable memory a perturbed dataset by introducing noise into the first quantitative biological dataset;   (3) applying a computer-implemented algorithm to the perturbed dataset to construct a set of perturbed connectivity matrices which are then stored in non-transitory computer-readable memory;   (4) using a computer-implemented algorithm to perform a stability assessment that reads from memory and compares the first connectivity matrices with the perturbed connectivity matrices;   (5) using a computer-implemented algorithm to select from among the first set of connectivity matrices the one matrix whose corresponding perturbed matrix is least affected by the introduction of noise; and   (6) storing the selected one matrix in non-transitory computer-readable memory and using a computer-implemented algorithm to construct the plurality of clusters,   b) constructing the connectivity matrix for each data type;   c) constructing the similarity matrix from T connectivity matrices to gain a holistic view of the integrated data; and   d) partitioning the integrated similarity matrix to define subtypes of the disease using the integrated data.   
     
     
         31 . The computerized method according to  claim 19 , comprising:
 generating the population dataset from the population of human subjects, each human subject of the population having the disease;   processing the population dataset using the unsupervised cluster analysis to define the clusters;   storing the clusters in a data store comprising a non-transitory computer-readable memory;   processing the patient dataset using a computer-implemented classifier, the classifier ingesting the clusters in the data store and using the clusters to find which of the clusters represents a closest match to the patient dataset;   storing the cluster found to represent the closest match as patient classification data in a non-transitory computer-readable memory; and   using the patient classification data in the selection of a disease treatment regimen,   wherein the unsupervised cluster analysis comprises:   (a) applying a computer-implemented algorithm to the population dataset construct a set of first connectivity matrices which are then stored in non-transitory computer-readable memory;   (b) using a computer-implemented algorithm to construct and store in non-transitory computer-readable memory a perturbed dataset by introducing noise into the population dataset;   (c) applying a computer-implemented algorithm to the perturbed dataset to construct a set of perturbed connectivity matrices which are then stored in non-transitory computer-readable memory;   (d) using a computer-implemented algorithm to perform a stability assessment that reads from memory and compares the first connectivity matrices with the perturbed connectivity matrices;   (e) using a computer-implemented algorithm to select from among the first set of connectivity matrices the one matrix whose corresponding perturbed matrix is least affected by the introduction of noise; and   (f) storing the selected one matrix in non-transitory computer-readable memory and using a computer-implemented algorithm to construct the clusters.   
     
     
         32 . A system comprising:
 at least one processor and   a memory coupled to the at least one processor; wherein the memory stores instructions for execution by the at least one processor, and wherein the instructions include:
 receiving assay results of samples collected from a population of human subjects; 
 generating a population dataset from the assay results that defines a molecular profile for the population; 
 receiving an assay result of a patient sample; 
 generating a patient dataset from the assay result that defines a molecular profile of the patient; and 
 analyzing the population dataset by:
 a) acquiring the population dataset from the population of human subjects; 
 b) performing an unsupervised cluster analysis on the population dataset to define clusters of the population, wherein each cluster represents a disease subtype or a patient subgroup; 
 c) using machine learning techniques to extract predetermined features of each disease subtype or patient subgroup; and 
 d) finding a disease subtype or patient subgroup that represents the closest match to the molecular profile of the patient.