Advances in neural AI and applications to drug discovery
Sprache des Vortragstitels:
Sprache des Tagungstitel:
New methods of artificial intelligence, especially those based on deep neural networks, attracted much attention in drug discovery research after wining the Merck Molecular Activity Challenge in 2014 and the Tox21 Data Challenge in 2015. Since then, many areas of drug development (e.g., virtual screening, bioactivity prediction, toxicity prediction, QSAR modeling, but also synergy models, generative models, and chemical synthesis) completely changed the analysis of pharma-related microscopy data.
Neural AIs in the form of Deep Learning suffer from the fundamental problem of vanishing and exploding gradients, but despite this fundamental problem, yield overwhelming successes. Recently developed techniques, such as normalization techniques, residual networks, or recurrent networks with memory, have contributed to mitigate the vanishing gradient problem and enabled scientific progress in many areas, notably in computer vision and speech, but also in drug discovery.
We give an overview of recent developments in the area of neural AIs, the vanishing gradient problem, how to mitigate it, and successful application areas. We focus on drug discovery as a especially successful application area.