Software splits coiled-coil proteins into short windows for AlphaFold to model

What this research was about

Many proteins contain long rope-like stretches called coiled coils, in which two or more helices wind around each other. Helices are the corkscrew segments of a protein chain; in a coiled coil they pack together by slotting bumps on one helix into gaps on its neighbour, an arrangement known as knobs-into-holes. These structures can run for hundreds of residues, the individual building blocks of a protein chain, and they can be parallel, with the chains running the same way, or antiparallel, with one reversed. Working out where such a stretch starts and stops, how many chains it involves and in which direction they run is awkward, because the behaviour is spread over a long, repetitive sequence.

The authors set out to approach this piece by piece rather than all at once. Instead of asking a structure predictor to handle an entire fibrous protein, they wanted to scan along the sequence in short overlapping windows and record what each window looks like on its own. They applied this to long fibrous and non-canonical coiled coils, including EEA1, myosin, a MACH-family protein, tetrabrachion and the SARS coronavirus spike protein.

Where AI came in

AlphaFold, a program that predicts a protein's three-dimensional shape from its sequence, supplies every structural observation here. It was used as released, through ColabFold version 1.5.2 with the alphafold2_multimer_v3 model, default sequence-alignment generation and no templates. CCfrag itself does no predicting. A divider module chops the sequence according to a chosen window length, overlap and number of chains, and writes the query files; the predictions run separately. An integrator module then reads the resulting models and logs, for each position, the per-residue confidence score pLDDT, the averaged error estimate PAE, whether the helices run parallel or antiparallel, and whether the SOCKET program finds knobs-into-holes packing, averaging where windows overlap.

ESMFold is supported as an alternative predictor. For comparison the authors also ran a full-length AlphaFold model of the same protein and two sequence-based coiled-coil predictors, COILS and DeepCoil 2, and superposed selected fragment models on published crystal structures. For EEA1 the fragment models scored better on pLDDT than the full-length prediction, with 70-residue windows giving the best scores, while shorter windows often failed to encode a parallel orientation. In the spike protein, fragment 925-975 came out as a parallel trimer 0.4 Angstroms from the solved structure 1ZVB.

Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.