Language models can learn complex molecular distributions.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Deep generative models are increasingly used for molecular discovery.
• Their effectiveness relies on learning molecular distributions from data.
• Language models, using string representations, are a simple yet powerful approach.

Purpose of the Study:

• To investigate the capacity of simple language models for learning complex molecular distributions.
• To evaluate language models on challenging generative tasks with large datasets.
• To compare language model performance against recent graph generative models.

Main Methods:

• Compilation of large, complex molecular distributions for generative modeling tasks.
• Training and evaluation of recurrent neural network-based language models.
• Benchmarking against state-of-the-art graph generative models.

Main Results:

• Language models adeptly learn complex molecular distributions.
• Accurate generation of high-scoring penalized LogP molecules from ZINC15.
• Successful generation of multi-modal distributions and large molecules from PubChem.
• Identified limitations of graph generative models in scaling to these distributions.

Conclusions:

• Simple language models are powerful generative tools for molecular discovery.
• These models show significant advantages in handling large and complex chemical spaces.
• Language models offer a scalable and effective alternative to graph-based methods for inverse molecular design.