ChloroGuide: prediction of subplastid localization using protein language model

This repository contains the official model weights and datasets for deep learning-based plastid sub-compartment predictor. The tool utilizes an ESM-2 (650M) protein language model combined with a 1D CNN and N-terminal feature extraction to predict protein localization into the Stroma, Envelope, Thylakoid Membrane, and Thylakoid Lumen. Repository Contents: 1. Model Weights (CD-HIT 0.7 - Default): cdhit07_ensemble_fp16.zip: 16-bit weights for the 5-model ensemble (Optimized for inference speed). cdhit07_ensemble_fp32.zip: 32-bit weights for exact paper reproducibility. cdhit07_best_model_fp16.pth / fp32.pth: Single best model for high-throughput fast scanning. 2. Ablation Study Models (CD-HIT 0.5 & 1.0): cdhit05_models_fp32.zip and cdhit10_models_fp32.zip: Additional models used for homology bias testing in the manuscript. 3. Datasets: datasets.zip: Contains the exact training, validation, and independent test FASTA sequences used in our study, clustered at different CD-HIT.