PharmaNet: Pharmaceutical discovery with deep recurrent neural networks

Abstract

The discovery and development of novel pharmaceuticals is an area of active research mainly due to the large investments required and long payback times. As of 2016, the development of a novel drug candidate required up to $ USD 2.6 billion in investment for only 10% rate of approval by the FDA. To help decreasing the costs associated with the process, a number of in silico approaches have been developed with relatively low success due to limited predicting performance. Here, we introduced a machine learning-based algorithm as an alternative for a more accurate search of new pharmacological candidates, which takes advantage of Recurrent Neural Networks (RNN) for active molecule prediction within large databases. Our approach, termed PharmaNet was implemented here to search for ligands against specific cell receptors within 102 targets of the DUD-E database, which contains 22886 active molecules. PharmaNet comprises three main phases. First, a SMILES representation of the molecule is converted into a raw molecular image. Second, a convolutional encoder processes the data to obtain a fingerprint molecular image that is finally analyzed by a Recurrent Neural Network (RNN). This approach enables precise predictions of the molecules' target on the basis of the feature extraction, the sequence analysis and the relevant information filtered out throughout the process. Molecule Target prediction is a highly unbalanced detection problem and therefore, we propose that an adequate evaluation metric of performance is the area under the Normalized Average Precision (NAP) curve. PharmaNet largely surpasses the previous state-of-the-art method with 97.7% in the Receiver Operating Characteristic curve (ROC-AUC) and 65.5% in the NAP curve. We obtained a perfect performance for human farnesyl pyrophosphate synthase (FPPS), which is a potential target for antimicrobial and anticancer treatments. We decided to test PharmaNet for activity prediction against FPPS by searching in the CHEMBL data set. We obtained three (3) potential inhibitors that were further validated through both molecular docking and in silico toxicity prediction. Most importantly, one of this candidates, CHEMBL2007613, was predicted as a potential antiviral due to its involvement on the PCDH17 pathway, which has been reported to be related to viral infections.

Fig 1. PharmaNet workflow.

For each molecule we compute a raw molecular image with nxm dimensions, where n is the number of unique atoms and bonds and m is the molecules’ maximum length. A convolutional encoder produces a fingerprint molecular image that is then analyzed globally by an RNN to predict scores for each of the targets in the AD Dataset.

Fig 2. Performance of molecular docking as a classifier.

Precision Recall Curve over the predictions computed with the binding energy of the molecular docking in the selected molecules against one of the 102 targets.

Fig 3. Comparison of PharmaNet against the state-of-the-art method of [24].

The curves correspond to the frequency in the 102 targets for each metric score: (A) Area Under the ROC Curve, (B) Average Precision, (C) Normalized Average Precision. We report area under the frequency curve. Best viewed in color.

Fig 4. Main ablation experiments.

(A) NAP curves evaluating the three main phases of our architecture. (B) NAP curves evaluating different RNNs. Best viewed in color.

Fig 5. Ablation study of convolutional architecture.

(A) Number of convolutional layers. (B) Kernel size for the convolution. (C) Type of normalization layers. (D) Type of residual connection. Best viewed in color.

Fig 6. Ablation study of GRU’s architecture.

(A) Unidirectional vs. Bidirectional GRU. (B) Hidden State Size. (C) GRU’s depth. Best viewed in color.

Fig 7. Correlation between normalized AP and tanimoto Similarity per protein.

Performance per protein in PharmaNet is shown as a function of the averaged Tanimoto Coefficient in molecules of the same class. R² is reported for estimating correlation between these two values.

Fig 8. Top-10 predictions of PharmaNet towards FPPS on the CHEMBL subset data.

CHEMBL ID, IUPAC name, SMILE and the molecule structure is given for the best 10 performing molecules in CHEMBL subset data towards FPPS target when predicting with PharmaNet.

Fig 9. Molecular docking.

(A) Zoledronate’s interaction with active site. Hydrogen bonds with Arg126 and Gln254. Binding Energy = -5.92. (B) CHEMBL250434’s interaction with active site. Hydrogen bonds with Arg126 and Thr215. Binding Energy = -10.04. (C) CHEMBL2007613’s interaction with active site. Hydrogen bonds with Arg126. Binding Energy = -9.56. (D) CHEMBL222102’s interaction with active site. Hydrogen bonds with Arg126. Binding Energy = -8.48.