Hanjo Kim

June 27, 2024

A few points about AI model validation

Oleg Trott, the developer of AutoDock Vina, one of the most popular molecular docking software, has expressed his skepticism about the comparison between AlphaFold3 and AutoDock Vina. Although I am not fully versed in the details of the AlphaFold3 paper, I can highlight some critical points from Oleg's post that are particularly relevant to chemists and molecular modelers—points that are often overlooked by data scientists.

  1. Data Split is Not Trivial: Understanding the dependencies within scientific data, such as protein-ligand pairs used in molecular docking algorithm tests, can be challenging for those without specialized expertise. Oleg explained the differences between sequence similarity and binding pocket similarity. While this distinction is clear to molecular modelers, it may not be as apparent to data scientists lacking extensive experience. Defining similarity between different entities (like small molecules and proteins) is a fundamental task for basic scientists.
  2. Exploring New Chemical Spaces: Drug discovery projects aim to produce patents, where novelty and superiority are crucial. Because of this, medicinal chemists strive to explore uncharted chemical spaces. Although a high hit rate in virtual screening campaigns seems beneficial, it can be misleading if it results in finding already known scaffolds. This highlights the difference between molecular similarity and scaffold similarity, another significant issue.
  3. Learning vs. Memorizing: Distinguishing whether an AI model has learned or merely memorized is challenging. Meaningful tests involve predicting novel data, which can sometimes be difficult or even impossible. Extrapolation will only be successful if the model selects the correct axis, aka a very complex combination of parameters. In drug discovery, the desired outcome is often a "black swan"—an exceptional molecule with superior activity and druggability compared to existing molecules. Combining physics-based functions with data models can be effective here, and there are many cases already. In fact, most commercial docking software are using this approach.

While AlphaFold3 simplifies the molecular docking process by eliminating the need for complex receptor preparation, a thorough understanding of the underlying algorithms, the implications of omitted tasks, and how to analyze results is essential for ensuring good outcomes.