Machine learning for predicting response scores for drugs

Abstract

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating a respective response score for each of a plurality of patient categories. In one aspect, a method comprises: generating a drug signature for a drug; generating an embedding of the drug signature in a latent space; and processing: (i) the embedding of the drug signature in the latent space, and (ii) data defining a plurality of patient categories, to generate a plurality of response scores, wherein each response score corresponds to a respective patient category and characterizes a predicted response of patients included in the patient category to the drug.

What is claimed is: 
     
         1 . A method performed by one or more computers, the method comprising:
 generating a drug signature for a drug, wherein: 
 the drug signature comprises a respective impact score for each of a plurality of features; and 
 the impact score for a feature characterizes an impact, caused by administering a drug to one or more entities, on a value of the feature measured for the one or more entities; 
   generating an embedding of the drug signature in a latent space, comprising: 
 generating a network input to an encoder neural network based on the drug signature; and 
 processing the network input generated based on the drug signature using the encoder neural network to generate the embedding of the drug signature in the latent space; and 
   processing: (i) the embedding of the drug signature in the latent space, and (ii) data defining a plurality of patient categories, to generate a plurality of response scores, 
 wherein each response score corresponds to a respective patient category and characterizes a predicted response of patients included in the patient category to the drug. 
   
     
     
         2 . The method of  claim 1 , wherein generating the drug signature comprises:
 obtaining, for each of the entities: 
 (i) a pre-treatment feature representation of the entity that comprises, for each of the plurality of features, a respective pre-treatment value of the feature that is measured for the entity prior to the drug being administered to the entity; and 
 (ii) a post-treatment feature representation of the entity that comprises, for each of the plurality of features, a respective post-treatment value of the feature that is measured for the entity after the drug is administered to the entity; and 
   generating the drug signature based on the pre-treatment and post-treatment feature representations of the entities.   
     
     
         3 . The method of  claim 2 , wherein generating the drug signature based on the pre-treatment and post-treatment feature representations of the entities comprises:
 generating, for each of the plurality of entities, a respective differential feature representation of the entity based on a difference between: (i) the pre-treatment feature representation of the entity, and (ii) the post-treatment feature representation of the entity; and   generating the drug signature based on the differential feature representations of the entities.   
     
     
         4 . The method of  claim 3 , wherein generating the drug signature based on the differential feature representations of the entities comprises:
 generating a respective entity-specific drug signature for each of the entities based on the differential feature representation of the entity; and   generating the drug signature by combining the entity-specific drug signatures.   
     
     
         5 . The method of  claim 4 , wherein for each of the entities, generating the entity-specific drug signature for the entity comprises:
 element-wise dividing the differential feature representation for the entity by the pre-treatment feature representation of the entity.   
     
     
         6 . The method of  claim 4 , wherein generating the drug signature by combining the entity-specific drug signatures comprises:
 averaging the entity-specific drug signatures.   
     
     
         7 . The method of  claim 1 , wherein drug signature comprises one or more impact scores that each characterize an impact, caused by administering the drug to the one or more entities, on a level of expression of a respective gene in the one or more entities. 
     
     
         8 . The method of  claim 1 , wherein the drug signature comprises one or more impact scores that each characterize an impact, caused by administering the drug to the one or more entities, on a level of expression of a respective protein in the one or more entities. 
     
     
         9 . The method of  claim 1 , wherein the network input to the encoder neural network includes the drug signature. 
     
     
         10 . The method of  claim 1 , wherein each of the plurality of patient categories is defined by a cluster of patient embeddings in the latent space, wherein each patient embedding corresponds to a respective patient and is generated by processing multi-modal data characterizing the patient using the encoder neural network. 
     
     
         11 . The method of  claim 10 , wherein for each of the plurality of patient categories, generating the response score for the patient category comprises:
 determining a respective similarity measure between: (i) the embedding of the drug signature, and (ii) each of one or more patient embeddings in the cluster of patient embeddings defining the patient category; and   determining the response score for the patient category based on the similarity measures.   
     
     
         12 . The method of  claim 1 , further comprising determining a ranking of the plurality of patient categories based on the response scores. 
     
     
         13 . The method of  claim 1 , further comprising:
 determining that a new patient is included in a patient category of the plurality of patient categories;   identifying the response score for the patient category of the new patient; and   automatically generating a recommendation for whether the new patient should receive the drug based at least in part on the response score for the patient category of the new patient.   
     
     
         14 . The method of  claim 1 , wherein each of the one or more entities comprises a cell. 
     
     
         15 . The method of  claim 1 , wherein each of the one or more entities comprises a collection of cells. 
     
     
         16 . The method of  claim 1 , wherein each of the one or more entities is a patient. 
     
     
         17 . The method of  claim 1 , wherein the encoder neural network has been trained to process multi-modal data characterizing patients. 
     
     
         18 . The method of  claim 17 , wherein the encoder neural network has been trained by operations comprising:
 obtaining a plurality of training examples, wherein each training example corresponds to a respective patient and includes multi-modal data that characterizes the patient;   jointly training the encoder neural network along with a decoder neural network on the plurality of training examples, comprising, for each training example: 
 processing the multi-modal data from the training example using the encoder neural network to generate an embedding of the multi-modal data from the training example; 
 processing the embedding of the multi-modal data from the training example using the decoder neural network to generate a reconstruction of the multi-modal data from the training example; and 
 updating current values of a set of encoder parameters and current values of a set of decoder parameters using gradients of a reconstruction loss function that measures an error in the reconstruction of the multi-modal data from the training example. 
   
     
     
         19 . A system comprising:
 one or more computers; and   one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations comprising: 
 generating a drug signature for a drug, wherein: 
 the drug signature comprises a respective impact score for each of a plurality of features; and 
 the impact score for a feature characterizes an impact, caused by administering a drug to one or more entities, on a value of the feature measured for the one or more entities; 
 
 generating an embedding of the drug signature in a latent space, comprising: 
 generating a network input to an encoder neural network based on the drug signature; and 
 processing the network input generated based on the drug signature using the encoder neural network to generate the embedding of the drug signature in the latent space; and 
 
 processing: (i) the embedding of the drug signature in the latent space, and (ii) data defining a plurality of patient categories, to generate a plurality of response scores, 
 wherein each response score corresponds to a respective patient category and characterizes a predicted response of patients included in the patient category to the drug. 
 
   
     
     
         20 . One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:
 generating a drug signature for a drug, wherein: 
 the drug signature comprises a respective impact score for each of a plurality of features; and 
 the impact score for a feature characterizes an impact, caused by administering a drug to one or more entities, on a value of the feature measured for the one or more entities; 
   generating an embedding of the drug signature in a latent space, comprising: 
 generating a network input to an encoder neural network based on the drug signature; and 
 processing the network input generated based on the drug signature using the encoder neural network to generate the embedding of the drug signature in the latent space; and 
   processing: (i) the embedding of the drug signature in the latent space, and (ii) data defining a plurality of patient categories, to generate a plurality of response scores, 
 wherein each response score corresponds to a respective patient category and characterizes a predicted response of patients included in the patient category to the drug.