Abstract
AbstractThe mapping from codon to amino acid is surjective due to the high degeneracy of the codon alphabet, suggesting that codon space might harbor higher information content. Embeddings from the codon language model have recently demonstrated success in various downstream tasks. However, predictive models for phosphorylation sites, arguably the most studied Post-Translational Modification (PTM), and PTM sites in general, have predominantly relied on amino acid-level representations. This work introduces a novel approach for prediction of phosphorylation sites by incorporating codon-level information through embeddings from a recently developed codon language model trained exclusively on protein-coding DNA sequences. Protein sequences are first meticulously mapped to reliable coding sequences and encoded using this encoder to generate codon-aware embeddings. These embeddings are then integrated with amino acid-aware embeddings obtained from a protein language model through an early fusion strategy. Subsequently, a window-level representation of the site of interest is formed from the fused embeddings within a defined window frame. A ConvBiGRU network extracts features capturing spatiotemporal correlations between proximal residues within the window, followed by a Kolmogorov-Arnold Network (KAN) based on the Derivative of Gaussian (DoG) wavelet transform function to produce the prediction inference for the site. We dub the overall model integrating these elements as CaLMPhosKAN. On independent testing with Serine-Threonine (combined) and Tyrosine test sets, CaLMPhosKAN outperforms existing approaches. Furthermore, we demonstrate the model’s effectiveness in predicting sites within intrinsically disordered regions of proteins. Overall, CaLMPhosKAN emerges as a robust predictor of general phosphosites in proteins. CaLMPhosKAN will be released publicly soon.
Publisher
Cold Spring Harbor Laboratory