Source-linked AI summary
Genetic algorithms are strong baselines for molecule generation
Austin Tripp, José Miguel Hernández-Lobato
TL;DR
Molecular generation is important in drug discovery, yet the paper argues that simple genetic algorithms are strong baselines for these tasks. It evaluates GAs on molecular-generation benchmarks and finds that they often match or outperform more complicated methods, motivating a criterion that new methods show an advantage over GAs.
Problem
Method
Results
GAs perform at least as well as many more complicated methods in unconditional and single-objective molecular generation settings.
Source: §4 · p. 3Takeaways & Limitations
The authors propose the GA criterion: new molecular-generation methods should demonstrate an empirical or conceptual advantage over GAs.
Source: §4 · p. 4The paper leaves open whether newer generative methods produce variants of their known-molecule training data, identifying this as a direction for future work.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Generating molecules, both in a directed and undirected fashion, is a huge part of the drug discovery pipeline. Genetic algorithms (GAs) generate molecules by randomly modifying known molecules. In this paper we show that GAs are very strong algorithms for such tasks, outperforming many complicated machine learning methods: a result which many researchers may find surprising. We therefore propose insisting during peer review that new algorithms must have some clear advantage over GAs, which we call the GA criterion. Ultimately our work suggests that a lot of research in molecule generation should be re-assessed.
1 Introduction
Molecular generation is an important but challenging part of drug discovery, and many machine-learning approaches target it. The paper argues that simple genetic algorithms are strong baselines and proposes that new methods demonstrate an advantage over them.
1 Introduction
- Drug discovery can be framed as molecular generation with multiple objectives and constraints, including generating molecules for virtual screening or desired properties.The paper distinguishes molecular graph generation from generating 3D conformations for a given graph.Source: §1 · p. 1
- Molecules can be generated by adding, removing, or substituting fragments of known molecules, with repeated modifications forming a genetic algorithm.Molecular validity rules are described as straightforward to check with software such as RDKit.Source: §1 · p. 1
- GAs often outperform complicated deep-learning methods on molecular generation tasks and perform better than previously reported results.The authors propose the GA criterion: new methods should offer some advantage over GAs.Source: §1 · p. 1
2 Background on genetic algorithms
Genetic algorithms are a broad class of iterative methods that generate and select molecules across successive populations. Their implementations differ most in how they generate candidates, while sampling and selection also shape search.
2 Background on genetic algorithms
- A genetic algorithm starts with a molecular population, samples from it, generates new molecules, selects the next population from old and new candidates, and repeats.Sampling may favor better population members, and selection may retain the best molecules.Source: §2 · pp. 1–2
- The generation step is the main factor distinguishing GAs, with approaches modifying molecular strings or graphs.Sampling and population selection also matter; for example, they can be chosen to avoid a homogeneous population.Source: §2 · p. 2
3 Experiments and Results
The experiments evaluate a default MOL_GA configuration on unconditional generation and molecule optimization benchmarks. MOL_GA performs similarly to other methods on unconditional-generation metrics and exceeds the methods reported in the PMO comparison, though the authors suspect baseline tuning explains the latter result.
3 Experiments and Results
- The experiments use MOL_GA’s default configuration, with top-quantile sampling, mutation and crossover based on Jensen (2019), and greedy population selection.The mutation and crossover implementation is based on the GuacaMol benchmark.Source: §3 · p. 2
3.1 Unconditional molecule generation
- The unconditional-generation benchmark measures validity, novelty, and uniqueness of generated molecules relative to a reference set.The experiments use ZINC 250K as the reference set.Source: §3.1 · p. 2
- On the ZINC 250k unconditional-generation benchmark, virtually all methods perform similarly, near 100% on all metrics, and more complex algorithms do not generally outperform simpler ones.Table 1 omits generation speed, which the authors say would determine how many molecules could be generated under a computational budget.Source: §3.1 · p. 2, Table 1 · p. 3
3.2 Molecule optimization
- Molecule optimization is motivated by the observation that structurally similar molecules often have similar activities, making variants of strong known molecules a plausible search strategy.The paper notes that no-free-lunch results preclude one general algorithm for all objective functions.Source: §3.2 · p. 3
- The PMO comparison limits each method to 10,000 evaluations of the objective function, and MOL_GA generates five molecules per iteration to allow about 2,000 improvement opportunities.The authors contrast this with earlier GAs using generation sizes around 100.Source: §3.2 · p. 3
- MOL_GA outperforms the best GA and every other method reported by Gao et al. (2022) on the PMO benchmark.The authors say this may be an artifact of baseline tuning rather than evidence that MOL_GA is an especially good method.Source: §3.2 · p. 3
4 Discussion
The authors conclude that GAs are strong baselines in unconditional and single-objective molecular generation, while arguing that poor empirical practices may explain limited progress. They propose the GA criterion and flag indirect generation of training-molecule variants as a future question.
4 Discussion
- GAs perform at least as well as many more complicated methods in unconditional and single-objective molecular generation.Source: §4 · p. 3
- The authors attribute limited progress partly to experiments favoring novel algorithms and treating comprehensive baseline evaluation as cumbersome.Source: §4 · p. 3
- The GA criterion asks new molecular-generation methods to show an empirical or conceptual advantage over GAs.Source: §4 · p. 4
- The authors suggest testing whether newer models generate variants of their training molecules indirectly as a future research direction.Source: §4 · p. 4
A Quantile-based sampling
MOL_GA uses rank-based quantile sampling to favor high-objective molecules without sampling only them, implemented with a deterministic grid of samples.
A Quantile-based sampling
- Rank-based sampling is intended to favor high-objective molecules without relying exclusively on them or on objective values themselves.Source: A Quantile-based sampling · p. 6
- Molecules are drawn uniformly from the top ε fraction of the population to form the sample S.Source: A Quantile-based sampling · p. 6
- The implementation creates a deterministic uniform grid of u samples as a quasi Monte-Carlo version of the procedure.Source: A Quantile-based sampling · p. 6