Researchers map the sixteen-part Commander complex using crystals, cryo-EM and AlphaFold2

The work

Technical · from the record

The authors determined the architecture of the sixteen-subunit endosomal Commander complex by combining X-ray crystallography, cryo-electron microscopy and AlphaFold2 Multimer structure prediction. AlphaFold2 models of the ten COMMD proteins together with the N-terminal regions of CCDC22 and CCDC93 were docked into cryo-EM maps of the twelve-subunit CCC complex, reported at 3.1 A in CryoSPARC and 3.5 A in RELION 4.0, and predictions supplied coordinates for parts the experimental data did not resolve, including the Retriever trimer, the CCDC coiled-coils that link CCC to Retriever, and the DENND10 binding site. A neural-network particle picker was trained on manually picked particles and then used to pick particles from the micrographs. The COMMD proteins were found to form a hetero-decameric ring of five specific heterodimers intercalated by CCDC22 and CCDC93 linkers, mutations designed from the predicted interfaces perturbed the corresponding interactions in cells and in vitro, and mutations causing X-linked intellectual disability and Ritscher-Schinzel syndrome were mapped onto the assembled model.

How AI was used

Two learned components were used. For cryo-EM, roughly 3,000 manually picked Retriever particles were used to train the Topaz picker implemented in RELION, and the trained network then auto-picked particles for the Retriever and CCC datasets; the picks were classified and refined with conventional CryoSPARC and RELION 4.0 processing. For structure determination, AlphaFold2 Multimer was run through the ColabFold interface on Google Colab to predict the COMMD1-10 plus CCDC22/CCDC93 N-terminal dodecamer, the Retriever VPS35L-VPS26C-VPS29 trimer, the DENND10-CCDC22-CCDC93 coiled-coil complex, and the VPS35L-CCDC interface, with iPTM scores, PAE plots and model-to-model alignment used to judge the predictions, and analogous predictions made for zebrafish and choanoflagellate sequences. The predicted dodecamer was docked into the sharpened cryo-EM map in ChimeraX and refined with PHENIX, COOT and ISOLDE; an AlphaFold2 prediction also served as the molecular-replacement template for the VPS29-VPS35L peptide crystal structure, and the machine-learning model-building program ModelAngelo was run to build a structure ab initio from the map for comparison. Predicted structures were merged into a single sixteen-subunit model, with experimentally determined regions substituted in and the result refined in PHENIX. Candidate CCC-Retriever assemblies were tested by prediction before the reported interface was selected, and point mutations designed from the models were then assayed in cells and in vitro.