Two recent pre-prints about how well AlphaFold models agree with experimentally derived data and if they can be used to accurately predict drug interactions.
Both studies highlight the limitations of AF when it comes to drug discovery.
🧵
Conversation
1) Terwilliger et al. explored how well AF models fit into experimentally derived electron density maps.
So far, the accuracy of AF models has mainly been assessed by how well predictions match protein structures but not the actual electron density.
1
4
The electron density is the actual data one collects to create a 3D model (what we call the structure).
In some cases AF models fit well, in others, they were off by a large margin and overall performed much worse than the models of the same protein.
biorxiv.org/content/10.110
1
5
2) Karelina et al. examined how useful AF2 models are for predicting protein-drug interactions.
Molecular docking analyzes the interactions between ligands and protein structures, identifying potential drug candidates.
1
2
AF2 modeled binding pockets with much higher accuracy. But the accuracy when used for molecular docking was comparable to traditional models and fell significantly short when compared to experiments with experimentally determined structures.
biorxiv.org/content/10.110
1
7
I think the studies above are good reminders that predicted models are great for hypothesis creation but that it's still a very long way until we can use them to develop new drugs.
Experimental is still👑
2
2
19
