Background

Coiled-coil domains (CCDs) are structural motifs where α‐helices pack together in an arrangement called knobs-into-holes [1], by which residues from one helix (the knobs) pack into holes formed by side chains in the other helices participating in the domain. CCDs have been observed in different kinds of proteins sequenced from all the kingdoms of life [2] and perform a great number of diverse functions.

Canonical CCDs include the interaction of two or more α‐helices, each characterized by the repetition of a seven-residue motif called heptad repeat. The positions of the heptad repeat are referred to as registers and are labeled with the letters a–g. CCDs can be classified into different oligomerization states, depending on the number (dimers, trimers, tetramers, and higher orders) and orientation (parallel or antiparallel) of the involved α‐helices.

Methods such as SOCKET [3] or SamCC-Turbo [4] can annotate CCDs starting from the experimental 3D structure of a protein. In the absence of structural information, several tools have been proposed over the years to perform automatic annotations on protein sequences, each addressing different tasks of CCD prediction (i.e., segment localization, heptad repeat annotation, oligomerization state classification).

Recently, the development of protein language models (PLMs) introduced a novel way of generating embeddings to encode protein sequences for downstream predictive tasks. We proposed CoCoNat [5], a deep learning–based approach that exploits two different and complementary PLMs, ProtT5 [6] and ESM2 [7], to produce a predictive pipeline for the complete ab initio annotation of CCDs.

CoCoNat processes input sequences with three cascading networks, each trained independently to solve a specific task. The first step adopts a deep architecture based on convolutional and recurrent layers to identify the presence of coil-coiled segments along the sequence. The second step adopts a probabilistic graphical model to assign registers to each residue in the segment. Finally, the third step adopts a neural network to predict the oligomerization state of each segment. Each prediction is also complemented with the probabilities computed by the network, allowing users to assess their reliability.

CoCoNat is trained on a dataset comprising 2,198 proteins annotated with 4,342 helices and 9,062 proteins without CCD. When tested on a non-redundant benchmark dataset, comprising 400 proteins annotated with 863 helices and 318 proteins without CCD, CoCoNat outperforms other methods on all three predictive tasks included in the pipeline. Specifically, it achieves a 0.54 per-residue F1 score and a 0.49 per-segment F1 score on the identification of segment boundaries over the sequence (first step in the Graphical overview), a Matthew's Correlation Coefficient (MCC) between 0.83 and 0.84 for each type of register in the annotation of the heptad repeats (second step in the Graphical overview), and an average MCC of 0.58 for the 4-class classification of the oligomerization state (third step in the Graphical overview).

Moreover, the adoption of PLMs to encode the input allows CoCoNat to be extremely time efficient. When tested on the virtual machine hosting the web server (AMD EPYC 7301 12-Core Processor, 48 GB RAM, no GPU), CoCoNat requires an average running time of 330 s (5.5 min) to predict 100 sequences of length comprised between 100 and 200 residues. The same computation takes approximately 2.5 h with CoCoPRED [8], a similar tool based on canonical multiple sequence alignments.

Here, we illustrate in detail how CoCoNat can be adopted as a web server or as a standalone tool, allowing for easy integration into any computational pipeline. As a test case, we select one of the proteins belonging to our benchmark dataset, the Mating-type switching protein swi5 from the organism Schizosaccharomyces pombe (UniProt accession: Q9UUB7). This protein presents two coiled-coil segments organized as a parallel dimer that CoCoNat identifies. Both the registers and the oligomerization classes are correctly assigned. The only difference between the putative and the real annotations is the length of the segments.

Supplementary File S1 reports a schema of the web server compliant with the Minimum Information About Bioinformatics investigation (MIABi) guidelines [9].