Schrödinger is excited to be participating in the 6th Artificial Intelligence in Chemistry Symposium taking place on September 4th-5th at Churchill College in Cambridge, England. Join us for a presentation by Márton Vass, Principal Scientist I at Schrödinger, titled “Enhancing physics-based simulations using machine learning to accelerate molecular discovery”. 

Speaker: Márton Vass, Principal Scientist I 
Abstract: 
Machine learning strategies in drug discovery are becoming increasingly popular. In order to leverage both the proven accuracy and wide applicability domain of physics-based computational models, such as QM and FEP, together with the speed and scale of machine learning, we have combined our physics-based modeling technologies with an active learning framework. This framework can be used to iteratively select interesting compounds to be labeled by the more computationally expensive methods in AL-Glide, AL-FEP, and AL-ABFEP, or to iteratively select training samples to achieve higher accuracy of the machine learning models to predict pKa in Epik 7, or more recently potential energy surfaces in the QRNN force field. Such scalable and generalizable machine learned force fields have widespread applications in chemistry. QRNN accurately predicts torsion profiles for use in classical MD and FEP simulations, crystal lattice energies, and also paves the way towards QM-accuracy MD simulations. QRNN efficiently reduces the number of candidate polymorphs by ~80% before ranking by DFT in crystal structure prediction. A specialised QRNN model for polymers produces stable MD trajectories and captures the dynamics of polymer chains as indicated by the agreement with experimental transport properties. Thus applications of this framework have a high impact in drug discovery and materials science.
 

REGISTER