Nat commun ｜ language model can learn complex molecular distribution

2022 year 6 month 7 Japan , From the computer science department of the University of Toronto Daniel Flam-Shepherd Et al. Nat Commun Publish research work , The research introduces three complex modeling tasks for the deep generation model of molecules to test the ability of chemical language model , The results show that the language model is a very powerful generative model that can learn any complex molecular distribution .

Thesis link ：

https://www.nature.com/articles/s41467-022-30839-x

Abstract

The deep formation model of molecules has been widely used , These models are trained on relevant data sets , Used to search the entire chemical space . The value of generation model for reverse design of new functional compounds , Depends on its ability to learn molecular distribution . The simplest example is a language model , It takes the form of a recurrent neural network , And the generated molecules are characterized by strings .

In this work , We studied the ability of simple language models to learn more complex molecular distributions . So , We compiled a larger 、 More complex molecular distribution , Several challenging generative modeling tasks are introduced , The ability of the language model on each task is evaluated .

It turns out that , The language model is a powerful generative model , Be able to skillfully learn complex molecular distribution . The language model can accurately generate ：ZINC15 The highest scoring punitive LogP Molecular distribution , Multimodal molecular distribution and PubChem The largest molecule in . These results highlight the limitations of some of the most popular and recent graph generation models , Many of these models cannot be extended to these molecular distributions .