"Celebrities turning 40 this week"

We believe that incorporating interaction-aware metrics into evaluation pipelines is essential for advancing the field, particularly for drug discovery applications. This could involve developing PLIF or pharmacophore-sensitive loss functions for training new ML models.

- While cofolding models show promise in predicting the full complex structure, our findings suggest a significant need for improvement in ensuring the predicted poses form meaningful interactions with the target protein.

- Our study shows that classical docking algorithms, when benchmarked with proper structure preparation, generally outperform current machine learning-based methods in recovering interactions.

- We introduce the Protein-Ligand Interaction Fingerprint (PLIF) recovery as an additional metric to directly assess the recall of these vital interactions.

Key takeaways:
- Focusing solely on ligand placement accuracy can overestimate model performance. We found that many predicted poses, especially from recent protein-ligand cofolding models, fail to capture key interactions.

We explore a critical aspect often overlooked in recent protein-ligand pose prediction benchmarks: the accurate recapitulation of protein-ligand interactions. While low RMSD and physical plausibility are important, they don’t tell the whole story.

It's annoying that the `declare` command doesn't work like `set` in csh which allows `set var = value`, but with a little bit of dubious trickery:

"He's a little confused, but he's got the spirit"