ANC Workshop - Diego Oyarzun and Angus Chadwick

Speaker: Diego Oyarzun

Title: AI for drug discovery: it works

Abstract:

Drug discovery is a slow and increasingly expensive process. Since it can take years or even decades to bring a new drug to the clinic, AI has emerged as a way to radically speed up the discovery of new medicines.

Here I will describe the use of binary classification to discover new anticancer drug candidates. I will report the discovery and experimental validation of three chemical compounds that specifically kill senescent or "zombie" cells that are therapeutic targets in a range of malignancies. Our approach allowed to reduce the cost of drug screening by 250-fold, using machine learning models trained solely on published data and commercial patents. This paves the way for an open-science approach to drug discovery. This project is part of a growing collaboration between our team and the Institute of Genetics and Cancer to develop a wholly new approach to drug discovery for the benefit of cancer patients

Speaker: Angus Chadwick

Title: Learning and attention: distinct changes in neural processing over long and short timescales

Abstract:

Learning and attention enhance neural and perceptual responses to task-relevant stimuli, but each operates on a vastly different time scale. Whether learning and attention rely on shared or distinct neural circuit mechanisms has been debated. I will discuss recent work investigating the impact of learning and attention on neural population activity in the visual cortex of mice. In single neurons, we find that the improvements in sensory processing that occur with learning are largely uncorrelated with those that occur with attention switching. At the population level, a linear dynamical system model fit to neural activity suggests that interactions amongst simultaneously recorded neurons reorganise with learning but stay fixed with attention switching. These findings suggest that learning and attention are underpinned by distinct changes in neural processing for task-relevant stimuli over long and short timescales. This work is the result of a collaboration with experimental and computational research groups based at University College London, Imperial College London, and Kings College London, and Cambridge University.

A preprint is available here https://www.biorxiv.org/content/10.1101/2021.01.31.429053v1.full.