PharmaNet: Pharmaceutical discovery with deep recurrent neural networks.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Pharmaceutical development is costly, with high failure rates and long timelines.
• Current in silico methods for drug discovery have limited predictive accuracy.
• Reducing investment and accelerating drug candidate identification is crucial.

Purpose of the Study:

• To introduce PharmaNet, a machine learning algorithm for accurate in silico prediction of pharmacological candidates.
• To leverage Recurrent Neural Networks (RNNs) for enhanced molecule-target interaction prediction.
• To improve the efficiency and reduce the cost of novel drug discovery.

Main Methods:

• PharmaNet converts molecular SMILES strings into images for analysis.
• A convolutional encoder generates molecular fingerprints, processed by an RNN.
• The algorithm was tested on the DUD-E database across 102 targets, predicting ligands for active molecules.

Main Results:

• PharmaNet achieved 97.7% ROC-AUC and 65.5% NAP, surpassing previous state-of-the-art methods.
• Perfect prediction performance was observed for human farnesyl pyrophosphate synthase (FPPS).
• Three potential FPPS inhibitors were identified in the CHEMBL dataset, with one (CHEMBL2007613) showing potential antiviral activity.

Conclusions:

• PharmaNet offers a highly accurate and efficient machine learning approach for identifying novel drug candidates.
• The algorithm demonstrates significant potential in predicting molecule-target interactions and discovering new therapeutic agents.
• Further validation of identified compounds, like CHEMBL2007613, could lead to new antiviral treatments.