<p></p><p>TorsionNet: A Deep Neural Network to Rapidly Predict Small
Molecule Torsion Energy Profiles with the Accuracy of Quantum
Mechanics </p>
<p> </p>
<p>Brajesh K. Rai<sup>*,1</sup>,
Vishnu Sresht<sup>1</sup>, Qingyi Yang<sup>2</sup>, Ray Unwalla<sup>2</sup>,
Meihua Tu<sup>2</sup>, Alan M. Mathiowetz<sup>2</sup>, and Gregory A. Bakken<sup>3</sup></p>
<p><sup>1</sup>Simulation
and Modeling Sciences and <sup>2</sup>Medicine Design, Pfizer Worldwide
Research Development and Medical, 610 Main Street, Cambridge, Massachusetts
02139, United States</p>
<p><sup>3</sup>Digital, Pfizer, Eastern Point Road,
Groton, Connecticut 06340, United States</p>
<p> </p>
<p> </p>
<p><b>ABSTRACT</b><b> </b><b></b></p>
<p>Fast and accurate assessment of small
molecule dihedral energetics is crucial for molecular design and optimization
in medicinal chemistry. Yet, accurate prediction of torsion energy
profiles remains a challenging task as current molecular mechanics methods are
limited by insufficient coverage of druglike chemical space and accurate quantum
mechanical (QM) methods are too expensive. To address this limitation,
we introduce TorsionNet, a deep neural network (DNN) model
specifically developed to predict small molecule torsion energy profiles
with QM-level accuracy. We applied active learning to identify nearly 50k
fragments (with elements H, C, N, O, F, S, and Cl) that maximized the coverage
of our corporate library and leveraged massively parallel cloud computing
resources to perform DFT torsion scan of these fragments, generating a training
dataset of 1.2 million DFT energies. By training TorsionNet on this dataset, we
obtain a model that can rapidly predict the torsion energy profile of typical
druglike fragments with DFT-level accuracy. Importantly, our method also provides
a direct estimate of the uncertainty in the predicted profiles without any
additional calculations. In this report, we show that TorsionNet can reliably identify
the preferred dihedral geometries observed in crystal structures. We also present
practical applications of TorsionNet that demonstrate how consideration of DNN-based
strain energy leads to substantial improvement in existing lead discovery and
design workflows. A benchmark dataset (TorsionNet500) comprising 500 chemically
diverse fragments with DFT torsion profiles (12k DFT-optimized geometries and
energies) has been created and is made freely available.</p><br><p></p>