Finding a new medicine can demand enormous effort and resources. This paper explores whether a machine that reads molecular strings can create promising new drug structures instead of searching only through existing ones. Drugs play an increasingly important role in the long struggle between people and disease, but drug discovery and design face obstacles because they require large human, material, and financial resources.

The model learns the distribution of molecules in a training set and then generates molecules similar to, but different from, those in the training set. The paper focuses on a text sequence called SMILES, a way to represent molecular structures with characters, because this molecular generative model is the most widely used among the forms described here.

A recurrent neural network can accept sequence data as input, and it was used in natural language processing before being used to generate new compound structures. When the molecules are expressed as SMILES, the network can process them as a string sequence and learn how molecular sequence grammar and chemical space are represented.

The everyday comparison is a person learning the rules of a language by reading many sentences: after learning the patterns, the network can generate new SMILES strings, meaning new molecules, for virtual screening and other applications. A variational autoencoder has an encoder and a decoder: it maps a molecular structure into hidden variables, then recovers those hidden variables as the SMILES sequence for the original structure.

After training, different hidden variables can be sampled and decoded to obtain different molecular structures. This creates a low-dimensional feature space, also called a drug space, where another point can be decoded into a potential targeted new drug.

A generative adversarial network uses two neural networks: a generator that fits the data distribution and a discriminator that judges whether an input is true. For molecular generation, the generator produces random SMILES while the discriminator tries to distinguish them from SMILES of real molecules in the training set.

During training, the generator keeps learning until the discriminator cannot distinguish whether a SMILES sequence came from the generator or the training set, allowing the generator network to generate molecules. Some generative models have successfully generated new lead compounds with expected physical and chemical properties, but their application in drug design can still be explored further.

That result points to a useful possibility: generated structures may help expand the search for molecules with properties researchers want, rather than limiting the search to compounds already listed. Compared with a rule-based virtual compound library, a generative model can learn molecular characteristics together with molecular properties, allowing it to sample new molecules satisfying specific properties more effectively.

SMILES has a more rigorous syntax and uses a mapping algorithm from molecular structure to text, which makes it easier to process and more suitable for training machine-learning models. Compared with two-dimensional and three-dimensional molecular representations, SMILES-based models are described as lower cost computationally, easier to scale to larger molecules or datasets, and able to benefit from progress in natural-language-processing algorithms.

Despite the spread of SMILES-based models, limitations remain: molecular similarity is not specified explicitly, existing natural-language-processing models may not apply directly, and invalid SMILES must be removed. With continued refinement, these models are expected to help pharmaceutical chemists expedite drug discovery and design.

That could offer a possible new way to speed the search for drug molecules, by helping pharmaceutical chemists expedite drug discovery and design. But the approach still has limitations, including the need to remove invalid SMILES representations, so researchers must check generated structures before using them.

The central idea is to treat a molecule like a sentence: learn how valid molecular strings are formed, then generate new ones with useful properties. It could help researchers search more quickly, though the approach still needs careful correction and testing.