predicted protein structures as evidence
BiologyComments
The gap isn't just in the benchmarks; it's in the side-chain orientations. AlphaFold is excellent for the backbone, but the chemistry of the active site often requires MD simulations or X-ray data to be truly useful.
Even if the active site chemistry is off, having a rough map helps labs prioritize which proteins are worth the expensive effort to crystallize. It saves a lot of wasted grant money on targets that are fundamentally disordered.
plddt maps to local disorder, so it actually provides a proxy for dynamics.
If we view this as a shift similar to how we use computational fluid dynamics before building a wind tunnel, does the loss of first principles intuition matter if the predictive utility remains high?
i wonder how this plays out with the current trend of using these models for ligand binding... does the shift toward virtual screening make the need for mechanistic validation even more urgent?
Which specific virtual screening benchmarks are showing the largest gap between predicted binding and empirical results?
We are treating a high pLDDT score like a crystal structure. Is it really a discovery if we cannot explain the transition state?
These predictions can act as highly educated guesses that narrow the search space for crystallography. It allows researchers to focus their empirical efforts on the most promising regions.