Back to results
US 12,597,524 B2Granted
Cancer evolution detection and diagnostic
AssigneeGuardant Health, Inc.
InventorsHelmy Eltoukhy, AmirAli Talasaz
Life of this patent
Estimating the term…
Abstract
The present disclosure provides methods for determining a probability that after any of a number of therapeutic interventions, an initial state of a subject, such as somatic cell mutational status of a subject with cancer, will develop a subsequent state. Such probabilities can be used to inform a health care provider as to particular courses of treatment to maximize probability of a desired outcome for the subject.
Claims20
A computer-implemented method, comprising:
- (a) obtaining information about a plurality of subjects with cancer at a first time point, wherein the information comprises, for each subject of the plurality of subjects, at least a genetic profile of a tumor obtained by genotyping nucleic acids from a cell-free bodily fluid and any treatment provided to the subject before the first time point, and determining a first state of each of the plurality of subjects based on the information at the first time point, to produce a set of first states;
- (b) obtaining the information about the plurality of subjects at one or more second time points subsequent to the first time point, and determining a second state of each of the plurality of subjects at each of the one or more second time points based on the information at a given one of the one or more second time points, to produce a set of subsequent states;
- and
- (c) training, with the set of first states from (a) and the set of subsequent states from (b), a predictive machine learning model that comprises a state-transition-probability matrix whose row indexed by each first state contains the probability of transitioning to every second state, and an emission-probability matrix that maps each hidden state to a corresponding observable state, each matrix being stored in a memory of a system, and wherein the predictive machine learning model is configured, upon receipt of a query first state, to determine a row of the state-transition-probability matrix indexed by the query first state, and weighted by the emission-probability matrix, thereby providing a probability distribution over candidate second states.
1 more element
A non-transitory computer readable medium storing processor executable instructions for performing a method, comprising:
- (a) obtaining information about a plurality of subjects with cancer at a first time point, wherein the information comprises, for each subject of the plurality of subjects, at least a genetic profile of a tumor obtained by genotyping nucleic acids from a cell-free bodily fluid and any treatment provided to the subject before the first time point, and determining a first state of each of the plurality of subjects based on the information at the first time point, to produce a set of first states;
- (b) obtaining the information about the plurality of subjects at one or more second time points subsequent to the first time point, and determining a second state of each of the plurality of subjects at each of the one or more second time points based on the information at a given one of the one or more second time points, to produce a set of subsequent states;
- and
- (c) training, with the set of first states from (a) and the set of subsequent states from (b), a predictive machine learning model that comprises a state-transition-probability matrix whose row indexed by each first state contains the probability of transitioning to every second state, and an emission-probability matrix that maps each hidden state to a corresponding observable state, each matrix being stored in a memory of a system, and wherein the predictive machine learning model is configured, upon receipt of a query first state, to determine a row of the state-transition-probability matrix indexed by the query first state, and weighted by the emission-probability matrix, thereby providing a probability distribution over candidate second states.
1 more element
Dependencies are inferred from the claim text.