Structure-based prediction of the interactome

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Linked publications, trials & patents

Abstract

Supplement Summary ?Structure-based prediction of the interactome? NIH R01GM081871-10 PI: Bonnie Berger We have designed and implemented a system for privacy-preserving and scalable sharing of drug-target interaction data (Aim 1, under review at Science), where we required GPUs to run our protocol, discover and experimentally validate novel drug-target interactions and will make our software publicly-available for academic and non-profit use (Aim 3). At the same time, we have presented a novel loss function for training classifiers from positive and unlabeled data and developed a software pipeline, Topaz, which uses convolutional neural networks trained with few positive examples for protein detection (Aim 2, RECOMB 2018). We are now developing new deep learning models for protein structure embedding and extending the Topaz framework to learn a general deep learning model of protein images from multiple cryo-EM micrograph datasets. Our continued progress on these projects is significantly jeopardized by our lack of GPU compute power. While in the last year we have purchased a compute node with four GPUs, we are continually frustrated by wait-times and inability to try different models and hyperparameters. Thus, we are requesting an additional node with eight GPUs to enable us to reach the broader goals of our grant.