It is known the AlphaFold2 predicts protein structures quite well, however, the structural quality by AlphaFold2 model is concern. I constantly observe sidechain clashes in many cases. It is also problematic in analyzing the protein-protein complex structure predict by AlphaFold2 as the in accurate sidechain placement may lead to wrong conclusion.

As this rotamer issue is well-known and people use Amber to “relax” the AlphaFold model a bit to place sidechain rotamers better. So I decided to generate structures of Rsp5 HECT domain -Ubiquitin complex using AlphaFold 3 and AlphaFold 2 (without/with relax) followed by quality check using Molprobity online server. The reference structure determined by X-ray crystallography can be found in RCSB with accession code 3OLM.

The structures done by AF2 without relax (left) and with relax (right) are identical in Rsp5-Ub interactions, pLDDT scores, and pAE plot.

Here is the summary. The structures are identical (RMSD < 0.3Å), however, the structural quality by AF2 without relaxation is quite bad.

		AlphaFold 3	AlphaFold2 without relax	AlphaFold2 with relax	Satisfied Goal
All atom contacts	Clash scores	5.57	20.02	3.58	<10
Protein geometry	Poor rotamer	0.48%	0.72%	0.48%	<0.3%
	Favored rotamer	96.88%	98.32%	97.84%	>98%
	Ramachandran outliers	0.44%	1.54%	0.22%	<0.05%
	Ramachandran favored	97.41%	96.92%	97.14%	>98%
	Z-score	0.39±0.38	0.18±0.37	0.04±0.37	|Z|< 2
	Molprobity score	1.46	1.98	1.3
	Cß deviations > 0.25 Å	0%	0%	0.23%	0
	Bad bonds	0%	3.12%	0%
	Bad angles	0.06%	0.44%	0.31%	<0.1%
Peptide Omega	Cis Prolines	0%	0%	0%	< 1%
	Twisted Peptides	0.66%	0.44%	0.66%	0
Low-resolution criteria	CaBLAM outliers	0%	0.7%	0.9%	<1.0%
	CA geometry outler	0.22%	0.44%	0.44%	<0.5%