Identifying chemical substructures associated with adverse drug reactions

A computer-implemented method for generating a framework for analyzing adverse drug reactions comprising:

• receiving, to a processor, a plurality of drug chemical structures;
• receiving, to the processor, a plurality of known drug-adverse drug reaction associations;
• constructing, by the processor, a deep learning framework for each of a plurality of adverse drug reactions based at least in part upon the plurality of drug chemical structures and the plurality of known adverse-drug reaction associations, wherein constructing the deep learning framework comprises defining a plurality of neighborhood-based fingerprints for each of the plurality of drug chemical structures using a plurality of hidden layers and generating a convolutional feature map comprising a plurality of convolutional steps, wherein each convolutional step encodes a respective neighborhood-based fingerprint at a respective hidden layer, wherein each fingerprint is obtained by starting from multiple atoms belonging to said fingerprint and resultant redundancies are removed by mapping each fingerprint into a lower dimension using a single layer of the deep learning framework;
• analyzing the deep learning frameworks to determine a set of substructure-adverse drug reaction associations;
• identifying a plurality of top predictive fingerprints for the set of substructure-adverse drug reaction associations based upon learned weights from a final layer of the deep learning framework;
• for each of the plurality of top predictive fingerprints, investigating each layer of the deep learning framework to identify atoms having a highest activation for the respective fingerprint;
• reconstructing a set of substructures by starting from each identified atom and expanding the neighborhood up to the respective layer, wherein reconstructing from a first identified atom on a first layer of the deep learning framework results in expanding the respective neighborhood up to the first layer, and wherein reconstructing from a second identified atom on a second layer of the deep learning framework results in expanding the respective neighborhood up to the second layer;
• calculating, for each substructure, a p value using a chi-squared test to evaluate a relative association strength between the substructure and the respective adverse drug reaction association;
• and
• redesigning a candidate substructure of a candidate drug to avoid a determined substructure-adverse drug reaction association.