Chlamy_ChloroPred: a deep learning-based, highly accurate binary classifier for chloroplast protein prediction in the model microalga, Chlamydomonas reinhardtii, with potential cross-proteome versatility

Introduction The chloroplast, a living relic of an ancient endosymbiotic interaction between a microalga and a microbe and the principal subcellular organelle responsible for biological CO 2 assimilation, is emerging as a key target for research to enhance photosynthetic efficiency beyond its current limitations. Given that accurate protein localization is a prerequisite for the in-depth scientific investigation and practical application of the membrane-compartmentalized photosynthetic organelle, numerous computational prediction tools have been proposed, yet their accuracy remains unsatisfactory. Methods To address the limitation, we herein present Chlamy_ChloroPred, a newly developed deep learning-based framework composed of multi-layered artificial neural networks, carefully designed to perform binary classification of chloroplast proteins in the model photosynthetic microorganism, Chlamydomonas reinhardtii . The model captures locality-aware features of determinant amino acid residues in the chloroplast transit peptide (cTP), generally located within the ~50-amino-acid N-terminal region of mature chloroplast proteins, through the integration of ProtBERT-BFD embeddings, stacked bidirectional long short-term memory (BiLSTM) networks, and an attentive pooling layer. Results and discussion Our model achieved an accuracy of 0.8462 for the C. reinhardtii proteome, outperforming widely used localization predictors, including TargetP 1.1 (0.4970), TargetP 2.0 (0.7396), and PredAlgo (0.7738) under a binary classification scheme. Comparative analyses further demonstrated that Chlamy_ChloroPred exhibits competitive performance relative to the current state-of-the-art model, PB-Chlamy (0.8521), under identical evaluation conditions. Notably, despite being trained solely on the algal proteome, Chlamy_ChloroPred showed substantial cross-species versatility when applied to the proteome of the terrestrial plant, Arabidopsis thaliana , achieving an accuracy of 0.7316 – representing a 12.6% improvement over TargetP 2.0, a predictor with previously demonstrated cross-proteome versatility. This likely stems from the model’s robust ability to capture conserved features of chloroplast proteins across proteomes from diverse photosynthetic lineages. Conclusion We developed a deep learning–based framework, Chlamy_ChloroPred, that integrates carefully designed neural layers with low computational complexity, achieving high predictive accuracy and interpretability. We believe that Chlamy_ChloroPred represents a compelling alternative to existing predictors, especially when accurate inference of chloroplast proteins is required.