TorsionNet: A Deep Neural Network to Rapidly Predict Small-Molecule Torsional Energy Profiles with the Accuracy of Quantum Mechanics.

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 challenging as the current molecular mechanics (MM) methods are limited by insufficient coverage of drug-like 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 compound library and leveraged massively parallel cloud computing resources for density functional theory (DFT) torsion scans of these fragments, generating a training data set of 1.2 million DFT energies. After training TorsionNet on this data set, we obtain a model that can rapidly predict the torsion energy profile of typical drug-like fragments with DFT-level accuracy. Importantly, our method also provides an uncertainty estimate for the predicted profiles without any additional calculations. In this report, we show that TorsionNet can accurately identify the preferred dihedral geometries observed in crystal structures. Our TorsionNet-based analysis of a diverse set of protein-ligand complexes with measured binding affinity shows a strong association between high ligand strain and low potency. 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. TorsionNet500, a benchmark data set comprising 500 chemically diverse fragments with DFT torsion profiles (12k MM- and DFT-optimized geometries and energies), has been created and is made publicly available.

Comprehensive Assessment of Torsional Strain in Crystal Structures of Small Molecules and Protein-Ligand Complexes using ab Initio Calculations.

The energetics of rotation around single bonds (torsions) is a key determinant of the three-dimensional shape that druglike molecules adopt in solution, the solid state, and in different biological environments, which in turn defines their unique physical and pharmacological properties. Therefore, accurate characterization of torsion angle preference and energetics is essential for the success of computational drug discovery and design. Here, we analyze torsional strain in crystal structures of druglike molecules in Cambridge structure database (CSD) and bioactive ligand conformations in protein data bank (PDB), expressing the total strain energy as a sum of strain energy from constituent rotatable bonds. We utilized cloud computing to generate torsion scan profiles of a very large collection of chemically diverse neutral fragments at DFT(B3LYP)/6-31G*//6-31G** or DFT(B3LYP)/6-31+G*//6-31+G** (for sulfur-containing molecule). With the data generated from these ab initio calculations, we performed rigorous analysis of strain due to deviation of observed torsion angles relative to their ideal gas-phase geometries. Contrary to the previous studies based on molecular mechanics, we find that in the crystalline state, molecules generally adopt low-strain conformations, with median per-torsion strain energy in CSD and PDB under one-tenth and one-third of a kcal/mol, respectively. However, for a small fraction (<5%) of motifs, external effects such as steric hindrance and hydrogen bonds result in strain penalty exceeding 2.5 kcal/mol. We find that due to poor quality of PDB structures in general, bioactive structures tend to have higher torsional strain compared to small-molecule crystal conformations. However, in the absence of structural fitting artifacts in PDB structures, protein-induced strain in bioactive conformations is quantitatively similar to those due to the packing forces in small-molecule crystal structures. This analysis allows us to establish strain energy thresholds to help identify biologically relevant conformers in a given ensemble. The work presented here is the most comprehensive study to date that demonstrates the utility and feasibility of gas-phase quantum mechanics (QM) calculations to study conformational preference and energetics of drug-size molecules. Potential applications of this study in computational lead discovery and structure-based design are discussed.

Fast and accurate generation of ab initio quality atomic charges using nonparametric statistical regression.

We introduce a class of partial atomic charge assignment method that provides ab initio quality description of the electrostatics of bioorganic molecules. The method uses a set of models that neither have a fixed functional form nor require a fixed set of parameters, and therefore are capable of capturing the complexities of the charge distribution in great detail. Random Forest regression is used to build separate charge models for elements H, C, N, O, F, S, and Cl, using training data consisting of partial charges along with a description of their surrounding chemical environments; training set charges are generated by fitting to the b3lyp/6-31G* electrostatic potential (ESP) and are subsequently refined to improve consistency and transferability of the charge assignments. Using a set of 210 neutral, small organic molecules, the absolute hydration free energy calculated using these charges in conjunction with Generalized Born solvation model shows a low mean unsigned error, close to 1 kcal/mol, from the experimental data. Using another large and independent test set of chemically diverse organic molecules, the method is shown to accurately reproduce charge-dependent observables--ESP and dipole moment--from ab initio calculations. The method presented here automatically provides an estimate of potential errors in the charge assignment, enabling systematic improvement of these models using additional data. This work has implications not only for the future development of charge models but also in developing methods to describe many other chemical properties that require accurate representation of the electronic structure of the system.

Shaping a screening file for maximal lead discovery efficiency and effectiveness: elimination of molecular redundancy.

High Throughput Screening (HTS) is a successful strategy for finding hits and leads that have the opportunity to be converted into drugs. In this paper we highlight novel computational methods used to select compounds to build a new screening file at Pfizer and the analytical methods we used to assess their quality. We also introduce the novel concept of molecular redundancy to help decide on the density of compounds required in any region of chemical space in order to be confident of running successful HTS campaigns.

Design of a multi-purpose fragment screening library using molecular complexity and orthogonal diversity metrics.

Fragment Based Drug Discovery (FBDD) continues to advance as an efficient and alternative screening paradigm for the identification and optimization of novel chemical matter. To enable FBDD across a wide range of pharmaceutical targets, a fragment screening library is required to be chemically diverse and synthetically expandable to enable critical decision making for chemical follow-up and assessing new target druggability. In this manuscript, the Pfizer fragment library design strategy which utilized multiple and orthogonal metrics to incorporate structure, pharmacophore and pharmacological space diversity is described. Appropriate measures of molecular complexity were also employed to maximize the probability of detection of fragment hits using a variety of biophysical and biochemical screening methods. In addition, structural integrity, purity, solubility, fragment and analog availability as well as cost were important considerations in the selection process. Preliminary analysis of primary screening results for 13 targets using NMR Saturation Transfer Difference (STD) indicates the identification of uM-mM hits and the uniqueness of hits at weak binding affinities for these targets.

Efficient exploration of large combinatorial chemistry spaces by monomer-based similarity searching.

In modern drug discovery, 2-D similarity searching is widely employed as a cost-effective way to screen large compound collections and select subsets of molecules that may have interesting biological activity prior to experimental screening. Nowadays, there is a growing interest in applying the existing 2-D similarity searching methods to combinatorial chemistry libraries to search for novel hits or to evolve lead series. A dilemma thus arises when many identical substructures recur in library products and they have to be considered repeatedly in descriptor calculations. The dilemma is exacerbated by the astronomical number of combinatorial products. This problem imposes a major barrier to similarity searching of large combinatorial chemistry spaces. An efficient approach, termed Monomer-based Similarity Searching (MoBSS), is proposed to remedy the problem. MoBSS calculates atom pair (AP) descriptors based on interatomic topological distances, which lend themselves to pair additivity. A fast algorithm is employed in MoBSS to rapidly compute product atom pairs from those of the constituent fragments. The details of the algorithm are presented along with a series of proof-of-concept studies, which demonstrate the speed, accuracy, and utility of the MoBSS approach.

A hybrid mixture discriminant analysis-random forest computational model for the prediction of volume of distribution of drugs in human.

A computational approach is described that can predict the VD(ss) of new compounds in humans, with an accuracy of within 2-fold of the actual value. A dataset of VD values for 384 drugs in humans was used to train a hybrid mixture discriminant analysis-random forest (MDA-RF) model using 31 computed descriptors. Descriptors included terms describing lipophilicity, ionization, molecular volume, and various molecular fragments. For a test set of 23 proprietary compounds not used in model construction, the geometric mean fold-error (GMFE) was 1.78-fold (+/-11.4%). The model was also tested using a leave-class out approach wherein subsets of drugs based on therapeutic class were removed from the training set of 384, the model was recast, and the VD(ss) values for each of the subsets were predicted. GMFE values ranged from 1.46 to 2.94-fold, depending on the subset. Finally, for an additional set of 74 compounds, VD(ss) predictions made using the computational model were compared to predictions made using previously described methods dependent on animal pharmacokinetic data. Computational VD(ss) predictions were, on average, 2.13-fold different from the VD(ss) predictions from animal data. The computational model described can predict human VD(ss) with an accuracy comparable to predictions requiring substantially greater effort and can be applied in place of animal experimentation.