Efficient molecular design is one of the fundamental goals in chemical science and engineering. In recent years, significant progress has been made in solving challenging problems across various aspects with computational molecular optimization algorithms. However, it's still common for authors to experiment with trivial oracles like QED or self-designed oracles and report selective metrics without any resource-constraint. Further, non-determinism in most algorithms can make reported results hard to reproduce. Without more meaningful standardization of experimental reporting, it is difficult to determine whether improvements over the prior results are significant. In this paper, we investigate the performance of the molecular design algorithms thoroughly in a limited-budget setting. We illustrate that most algorithms cannot optimize molecules efficiently in existing benchmarks, and many recent deep learning-based methods cannot surpass some traditional molecular design algorithms. Our results also highlight the necessity of multiple independent runs and tuning hyper-parameters. We suggest standard experimental reporting to minimize the wasted effort caused by non-reproducible and easily misinterpreted results.

Sample efficiency matters: Benchmarking molecular optimization

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.