Deep Learning Drug Toxicity and Cancer Drivers

Speaker: Dr. S. Joshua Swamidass MD PhD

Date: Wednesday, April 19, 2017

Time: 3:00 PM to 4:00 PM Note: all times are in the Eastern Time Zone

Public: Yes

Location: G882

Event Type:

Room Description:

Host: Bonnie Berger and Manolis Kellis

Contact: Geraldine McGowan, 617-253-3497, gmcgowan@csail.mit.edu

Relevant URL: http://swami.wustl.edu/.

Speaker URL: None

Speaker Photo:
None

Reminders to:

Reminder Subject: TALK: Deep Learning Drug Toxicity and Cancer Drivers

Our group has been using Deep Learning to solve problems at the intersection of biology, chemistry and medicine. Two projects that use Deep Learning will be covered here: (1) modeling the metabolism and subsequent reactivity of drugs, and (2) identifying cancer drivers in cancer sequencing data.

Adverse drug reactions (ADRs) are dangerous and expensive. Idiosyncratic ADRs, especially rare and severe hypersensitivity-driven ADRs, are the leading cause of medicine withdrawal and termination of clinical development. At the same time, a large proportion of drugs are not associated with hypersensitivity driven ADRs, offering hope that new medicines could avoid them entirely with reliable predictors of risk. Hypersensitivity driven ADRs are caused by the formation of chemically reactive metabolites by metabolic enzymes. These reactive metabolites covalently attach to proteins to become immunogenic and provoke an ADR. To predict ADRs, we have been building mathematical models of metabolism and reactivity. The models are constructed using Deep Learning: a machine-learning algorithm that quantitatively summarize the knowledge from thousands of published studies. Taken together, this approach is more accurately modeling the properties determining whether metabolism renders drugs toxic or safe.

Methods are needed to reliably prioritize biologically active driver mutations over inactive passengers in high-throughput cancer sequencing datasets. We present ParsSNP, an unsupervised functional impact predictor that trains a neural network with a parsimony-based constraint. We compare ParsSNP to five existing tools (CanDrA, CHASM, FATHMM Cancer, TransFIC, Condel) across five distinct benchmarks. ParsSNP outperformed existing tools in 24 out of 25 comparisons. This unsupervised approach outperforms all other supervised approaches in identifying cancer drivers.

Research Areas:

Impact Areas:

This event is not part of a series.

Created by Geraldine McGowan Email at Wednesday, April 05, 2017 at 1:18 PM.