Deep learning is a type of machine learning that uses deep neural networks: networks with multiple computational layers that learn representations and complex patterns from data. In drug discovery, these patterns can help identify promising molecules among many candidates. section 1.5 section 1.0
A model first learns from a training set, then makes predictions on data it has not seen during training. Molecules must be represented computationally before a model can use them. Two representations are SMILES, a character string encoding molecular structure, and molecular graphs, in which atoms are nodes and bonds are edges. section 1.2.2 section 1.5
Screening compounds for kinase targets
Protein kinases are an important family of drug targets. A ligand-based virtual-screening workflow can use a neural network to score compounds and rank likely active compounds near the top, so they can be prioritized for screening. section 8.1 section 8.3
A basic PyTorch workflow follows these steps:
- Convert inputs to numerical representations and encode them as tensors, multidimensional arrays used for model computations. section 8.3
- Specify the network architecture and fit the model to the data. section 8.3
- Evaluate predictions against actual test values, review possible architecture and training modifications, and save the final model for reuse. section 8.3
Other applications include molecular property prediction, generating molecules, predicting chemical reactions and synthesis pathways, and protein folding. In retrosynthesis, deep learning offers a way to explore a large space of possible transformations when planning how to make a molecule. section 1.3 section 1.3.3



















