International Business Machines Corporation

Aldo Guzman Saenz

Laxmi Priya Parida

Filippo Utro

Kahn Rhrissorrakrai

Chaya Levovitz

Methods and systems for training a machine learning model are described. A processor can transform single cell data in a first space into projection data in a second space having a dimensionality lower than or equal to the first space. The processor can produce a cover having a plurality of sets of the projection data. The processor can determine a plurality of transition paths among the plurality of sets. A transition path can represent a transition from one cell state to another cell state. The processor can translate the transition paths from the second dimensional space to the first dimensional space. The processor can extract features from the transition paths in the first dimensional space. The processor can generate training data using the features, and use the training data to train a machine learning model for classifying cell state transitions.

Cell state transition features from single cell data

A computer-implemented method includes generating, by a processor, a set of training data for each phenotype in a database including a set of subjects. The set of training data is generated by dividing genomic information of N subjects selected with or without repetition into windows, computing a distribution of genomic events in the windows for each of N subjects, and extracting, for each window, a tensor that represents the distribution of genomic events for each of N subjects. A set of test data is generated for each phenotype in the database, a distribution of genomic events in windows for each phenotype is computed, and a tensor is extracted for each window that represents a distribution of genomic events for each phenotype. The method includes classifying each phenotype of the test data with a classifier, and assigning a phenotype to a patient.

Single sample genetic classification via tensor motifs

A computer-implemented method includes inputting, to a processor, an N&#xd7;K SSV frequency matrix M and an error tolerance &#x3b4;&#x2265;0, wherein N is a number of SSVs and K is a number of time points, wherein matrix M comprises a plurality of time-resolved mutation frequencies for each SSV; clustering, by the processor, matrix rows in M that satisfy the &#x3b4; to provide a plurality of SSV clusters; assigning, by the processor, a mean cluster frequency to each SSV within each SSV cluster; calculating errors for removing low frequency rows, for rounding rows to 1 or 0; assigning a root node for all SSV clusters of frequency 1; and calculating, by the processor, a &#x3b4;-compliant time-series evolution tree with error &#x2264;&#x3b4; comprising the root node and a plurality time-stratified nodes, wherein calculating includes assigning a clonal configuration, optionally re-configuring the clonal configuration, and calculating error for re-configuring.

Growing community of inventors

White Plains, NY, United States of America

Aldo Guzman Saenz