# FastFlows: Flow-Based Models for Molecular Graph Generation

**Authors:** Nathan C. Frey, Vijay Gadepally, Bharath Ramsundar  
**arXiv:** [2201.12419](https://arxiv.org/abs/2201.12419)  
**Date:** 2022-01-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2201.12419v1/

## Key points

- FastFlows is a framework for generating small molecules using normalizing-flow based models.
- It combines normalizing flows, SELFIES, and multi-objective optimization to efficiently generate chemically valid molecules.
- With just 100 small molecules in the initial training set, FastFlows can generate thousands of molecules in seconds.
- FastFlows has an efficient sampling process that allows for the application of substructure filters to eliminate compounds with unreasonable moieties.
- Easily computable and learned metrics for druglikeness, synthetic accessibility, and synthetic complexity are incorporated into the framework.
- FastFlows is significantly simpler and easier to train compared to autoregressive molecular generative models.
- It enables fast generation and identification of druglike and synthesizable molecules in high-throughput virtual screening.
- Flow-based models face challenges when dealing with higher-dimensional chemical spaces, but more expressive flows are needed that preserve fast sampling and training.

## Summary

The paper introduces FastFlows, a framework for generating small molecules using normalizing-flow based models. The authors propose a combination of normalizing flows, SELF-Referencing Embedded Strings (SELFIES), and multi-objective optimization to efficiently generate chemically valid molecules. With just 100 small molecules in the initial training set, FastFlows is able to generate thousands of molecules in seconds. One key advantage of FastFlows is its efficient sampling process, which allows for the application of substructure filters to eliminate compounds with unreasonable moieties. This ensures that only desirable molecules are generated. The authors also incorporate easily computable and learned metrics for druglikeness, synthetic accessibility, and synthetic complexity into their framework. To demonstrate the effectiveness of FastFlows in a high-throughput virtual screening context, the authors perform a multi-objective optimization using the computed metrics. They show that their model is significantly simpler and easier to train compared to autoregressive molecular generative models. Additionally, it enables fast generation and identification of druglike and synthesizable molecules. The paper acknowledges that while FastFlows can be extended to more relevant target distributions for drug discovery, flow-based models face challenges when dealing with higher-dimensional chemical spaces. However, the authors emphasize the need for more expressive flows that preserve fast sampling and training. In conclusion, FastFlows presents a novel approach to generative modeling of small molecules using normalizing flows. Its efficiency in generating chemically valid molecules makes it a valuable tool for high-throughput virtual screening in drug discovery. The framework's simplicity and ease of training make it an attractive alternative to autoregressive molecular generative models.