AlphaFold Multimer v2.2.2 Incorrectly Predicts Homo-oligomer Despite PDB Structure

An interesting case of ~accurate protomer prediction but poor multimer prediction:

I predicted a homo-oligomer complex (ring) for a well-characterized protein with AlphaFold Multimer v2.2.2 (using the full database). The protein has a nearly complete PDB structure which AlphaFold uses as a template. The predicted structure has the expected general topology; however, the alpha-carbons of the predicted and PDB structures align very poorly (RMSD 26), and it is visually apparent that there is excessive space between the chains of the predicted structure resulting in an abnormally large ring diameter. The alignment object suggests that the chains of the predicted structure should be in a different order.

The individual chains of the predicted structure align to the individual chains of the PDB structure fairly well (RMSD ~2). When the chains of the predicted structure are aligned individually, the resulting 'reassembled' predicted structure globally aligns to the PDB structure with an RMSD <2.