Since we aimed at developing a CP classifier tailored for C. reinhardtii, we collected 1,262 protein sequences, of which 841 were revealed to be localized in the chloroplast (i.e., CP) and 421 were revealed to be localized in organelles other than the chloroplast (i.e., Non-CP) (Figure 1A; Supplementary Dataset S1). To furnish the developing program with data whose localizations are empirically corroborated, we manually mined gene identifications from 6 literature sources, including: 5 reports that experimentally investigated the proteomics of the model green microalga using fluorescence-tagged proteins (Wang et al., 2023) and mass spectrometry identifying semi-tryptic peptides (Terashima et al., 2010; Bienvenut et al., 2011; Zhan et al., 2018; Ramundo et al., 2020); and one reference that thoroughly scrutinized the subcellular destinations of proteins reported in various publications using its own stringent criteria (Tardif et al., 2012). Protein sequences corresponding to the gene identifications were then scrapped from the Chlamydomonas reinhardtii v5.6 database on the Phytozome (Merchant et al., 2007). Only primary transcripts were used for building this dataset. During the compilation of amino acid sequences from the references, redundant sequences (i.e., the identical sequence found in multiple sources) were treated as follows: (i) Sequences designated with the same localization were merged (i.e., one was selected); and (ii) those with controversial localizations were removed to avoid any reasonable doubt. After assembly, 420 CPs were excluded from the overall dataset to balance the data distribution. This resulted in the “balanced labeled dataset” consisting of an equal number of CPs and Non-CPs (421 each; Figure 1A), given that imbalanced input data may induce systematic bias in the model (Belhaouari et al., 2024). Subsequently, we split the resulting dataset into “train (Supplementary Dataset S2),” “validation (Supplementary Dataset S3),” and “test (Supplementary Dataset S4)” datasets with a ratio of 3:1:1, while maintaining the label distribution of each subset (Figure 1B).

We designed a deep learning-based classifier framework that integrates a stacked recurrent neural network and attention-based mechanisms to forecast CPs from ProtBERT-BFD embeddings (Figure 2). Prior to the model organization, the datasets were labeled in a binary fashion as 0 and 1 for Non-CPs and CPs, respectively, to proceed with supervised learning. After labeling, all protein sequences subjected to the train, validation, and test processes were fixed at a maximum length of 1,000 amino acids (AAs) from the N-terminus, with longer sequences truncated and shorter ones padded. The standard length of the N-terminal sequences likely containing cTPs (i.e., 1,000 AAs) was determined by considering that cTPs are typically no longer than this limitation (Tardif et al., 2012) and that >95% of sequences in the overall dataset fall within 1,000-AA coverage (Figure 1B), making the model framework lightweight while covering the entire length of the most sequences.

As an initial step in constructing the framework, a pre-trained ProtBERT-BFD model was used to obtain embeddings (i.e., 1,024-dimensional vectors) from datasets, enriched with biochemical and structural context pre-learned from large protein corpora. We used the token-level embeddings derived from the ProtBERT-BFD-based feature extractor as inputs of downstream. A positional encoding layer was applied to the input to preserve positional information lost during embedding. Using a layer normalization, the position-encoded sequences were then normalized. Subsequently, the normalized outputs were processed by a stack of triple bidirectional LSTM (BiLSTM) layers, which retain hidden dimensions of 128, 64, and 32 units, respectively, in order to capture contextual dependencies along the protein sequence in both the forward and backward directions. Next, we added a multi-head self-attention mechanism with 4 heads and 64 key dimensions to the framework to improve the representation with long-range inter-residue context-aware dependencies. The resulting output was incorporated back into the BiLSTM through a residual connection, followed by layer normalization and a Transformer feed-forward subnetwork. This attention block enables the model to dynamically weight sequence positions and understand the relationships (i.e., global dependencies) among all tokens in parallel. Following the contextualization step, we applied another sublayer of the attention mechanism, an attentive pooling layer, which calculates a weighted average of sequence features. This effectively prioritizes biologically informative sequence regions (e.g., cTP) and simultaneously attenuates less relevant residues for classification. Finally, the pooled representation was regularized using dropout with a probability of 0.45 and then passed through a fully connected dense layer of 256 units with rectified linear unit (ReLu) activation and L2 kernel regularization (λ = 1 × 10⁻⁴), followed by the final dropout step consisting of a single unit with a sigmoid activation function that prints the probability of the sequence being localized to the chloroplast. For binary classification, a probability of ≥0.5 was used as the threshold for determining localization.

We used Google Colaboratory (Colab) environment (Python 3, T4 GPU, and high-RAM mode) for running the deep learning-based program. The model was trained using the Adam optimizer with a learning rate of 1 × 10⁻⁴, minimizing the binary cross-entropy loss with prediction accuracy as the evaluation metric. The early stopping callback was used to prevent the neural network from overtraining by monitoring the validation loss with a patience of 3. We saved the best version of the trained model with the best validation accuracy for further evaluation using the model checkpoint callback.

When evaluating the performance of models developed in this study based on the input data described in Section 2.1, the performance metrics including the accuracy, precision, recall—also known as sensitivity or true positive rate (TPR), and F1 score were calculated and compared as follows (Yue et al., 2022):

Where TP is true positive (counted if an instance of CP is correctly predicted as CP), TN is true negative (counted if an instance of Non-CP is correctly predicted as Non-CP), FP is false positive (counted if an instance of Non-CP is wrongly predicted as CP), and FN is false negative (counted if an instance of CP is wrongly predicted as Non-CP). Accuracy measures the percentage of correct predictions across all samples, while precision calculates the proportion of positive classifications that were actually correct and improves as the number of false positive decreases. Recall is defined as the proportion of actual positives that were correctly classified as such and improves when the number of false negative decreases. F1 score is the harmonic mean of precision and recall, which balances the importance of both, indicating the reliability of a model.

Meanwhile, to evaluate the model’s performance (i.e., effectiveness of the binary classification model), we plotted the ROC (receiver operating characteristic) curve and estimated the AUC (area under the ROC curve). The ROC curve visually represents the performance of a model that is graphed by calculating the TPR (Equation 3) against the false positive rate (FPR) at various threshold values. The FPR, which measures how often a model incorrectly predicts negative samples as positive, is calculated as follows (Adabor et al., 2025):

The AUC is one of the most useful single indicators for comparing the overall performance of models, particularly when the dataset is balanced. An AUC score of 1 means the model achieves perfect classification.

As one of the performance measurements of the developed model, the entire C. reinhardtii proteome (i.e., 19,526 proteins provided in Chlamydomonas reinhardtii genome version 5.6 in the Phytozome database) (Merchant et al., 2007) was also prepared and tested as described above.

The performance of the proposed model was compared with that of the TargetP 2.0 software (Almagro Armenteros et al., 2019), which was selected as the primary benchmark in this study due to its ease of access, multi-species and cross-proteome versatility, and established authority in the field of protein localization prediction. When running the software, two parameters were selected: “Plant” for the organism group and “Long output” for the output format. Input dataset for the benchmark was identically pre-conditioned as described in Sections 2.1 and 2.2. For the TargetP 2.0 results, a protein was designated as a chloroplast protein when the “Chloroplast transfer peptide” score was the highest among other scores.

To more comprehensively position Chlamy_ChloroPred among currently available models, we examined the predicted localization of proteins in the test dataset (Supplementary Dataset S4) using broader predictors, including TargetP 1.1 (the previous version of TargetP 2.0) (Almagro Armenteros et al., 2019), PredAlgo (Tardif et al., 2012) (based on the prediction results publicly archived in the Phytozome Chlamydomonas reinhardtii v5.6 database, as real-time, up-to-date services for both tools are not currently available, as confirmed by server status checks and personnel communication), and PB-Chlamy (Wang et al., 2023)—a SOTA model capable of predicting multiple subcellular localizations, such as mitochondrial, secretory, and other cellular compartments (integratively classified as “other”), in addition to chloroplast localization, as well as assigning potential mitochondrial/chloroplast and secretory/chloroplast dual-targeted protein candidates. Given its ability to distinguish dual-targeted proteins, PB-Chlamy was further used to extract putative dual-targeted protein candidates (Supplementary Dataset S5).

The versatility of the Chlamy_ChloroPred program was evaluated using a chloroplast proteome dataset from the model terrestrial plant, A. thaliana. The chloroplast protein dataset was basically obtained from the report by Ferro et al. (2010). Of the 1,323 protein identifications (IDs) listed in the study, the amino acid sequences of 1,263 protein IDs (Supplementary Dataset S6), which are now available on the TAIR (The Arabidopsis Information Resource) portal were tested (Reiser et al., 2017). The identical dataset was also applied to the primary benchmark in this study, TargetP 2.0, and the resulting predictions were compared. For the predicted results from TargetP 2.0, proteins that were predicted to be localized in other organelles (e.g., the mitochondria or the extracellular region) were treated as non-CPs. Those that were predicted to be localized in the chloroplast or the thylakoid lumen—as a component of the chloroplast—were treated as CPs.

Given the heavy dependence of model performance on the quality of the input data, the preparation of refined input data is imperative in implementing a high-performance predictive program. Considering this, we first balanced the initially collected dataset consisting of 1,262 proteins that include 841 CP-labeled and 421 non-CP-labeled proteins by excluding 420 CP-labeled proteins, leading to a balanced labeled dataset with 421 CP-labeled and 421 non-CP-labeled protein sequences (842 proteins in total). Subsequently, approximately 80% (673 proteins) of the balanced labeled dataset was grouped into the model-building dataset while ca. 20% (169 proteins) of the dataset was grouped into the test dataset. The model-building dataset was further split into train and validation datasets, which contain about 75% (504 proteins) and 25% (169 proteins) of the original dataset, respectively. All splitting and grouping processes were conducted while maintaining balance in label distributions (i.e., CP vs. non-CP) because it reduces bias toward the majority class, ensures fair training, and improves the performance and interpretability of a binary classification model (Friedberg et al., 2013). As a result, we successfully secured the total dataset, consisting of train, validation, and test subsets in a 3:1:1 ratio (Figure 1A).

Next, we examined the length distribution of the datasets to confirm that they all have a similar dispersion of amino acid lengths. Similar length distributions were observed in the histograms of each split dataset, and it was found that the subsets demonstrate analogous summary statistic values (e.g., mean and minimum lengths) (Figure 1B). On the contrary, the maximum lengths of each subset are disparate because the total dataset (i.e., the balanced labeled dataset) contains a sparse number of proteins over 1,000 AAs (2.49%)—only 21 out of 842 proteins. The maximum length of sequence data in a dataset is of interest and importance since it would establish a standard for length normalization, which is used to uniform input dimensions. Given that using a long input length (e.g., 4,661 AAs in this study) can lead to a high computational burden and rather adversely affect the model’s performance, the input length was standardized to 1,000 AAs, which has a high chance of containing the localization determinant sequence, cTP, without sacrificing the sequence information of more than 95% of protein sequences in any dataset that are shorter than this coverage length (Figure 1B).

As shown in Figure 2, a predictive model, named Chlamy_ChloroPred, was constructed in this study that is primarily based on a triple stack of RNN-BiLSTM and an attentive pooling layer. Prior to data input, the length-normalized input sequences were tokenized using special tokens, such as the classification token (CLS), the padding token (PAD), and the separation token (SEP), and then embedded by the ProtBERT-BFD protein language model (Elnaggar et al., 2022), which provides enriched biochemical and structural features pre-trained on a large number of protein sequences. As a result, this model outputs the probability of a protein sequence being localized to the chloroplast with a threshold of 0.5, which implements binary classification (Lee et al., 2022). The detailed parameters are presented in the Materials and Methods section.

Under the Google Colab environment, the constructed code was executed. Running was terminated at epoch 11 to prevent overfitting, triggered by the early stopping callback. Since the loss value measures how much the model’s predictions deviate from the true labels, epoch 11 seemed adequate as the stopping point. This is evident from the result that the stopping point occurred before the train and validation losses largely diverged, while the train and validation accuracies increased steadily (Figures 3A,B). Based on the trained program, we graphed the ROC curve and calculated the AUC (Figure 3C) to examine the overall performance of the model. Our classifier showed an AUC of 0.9090, surpassing the 0.5 AUC of a random guess model. This demonstrates that our model is much better than a dummy classifier at distinguishing between positive (i.e., CP) and negative (i.e., non-CP) classes.

More specifically, the model’s inference performance was evaluated with regard to accuracy, precision, recall, and F1 score using the train, validation, and test datasets. The results were presented in the form of confusion matrices (Figures 3D–F). The accuracy (Equation 1), precision (Equation 2), recall (Equation 3), and F1 score (Equation 4) were not lower than 0.8462 (with a loss of 0.4052), 0.8718, 0.8000, and 0.8395, respectively, most of which were derived from the test dataset. This reveals the model’s balanced and reliable performance in predicting whether given proteins in C. reinhardtii are localized to the chloroplast or not.

Subsequently, we illustrated heatmaps of residue importance derived from the attentive pooling layer, which was employed to improve the model’s predictive accuracy by focusing attention on relevant sequence features and to offer interpretability through attention weight visualization. As shown in Figure 3G, our model focused selectively on the N-terminal regions of most CP-labeled proteins in the test dataset (upper right panel of Figure 3G), while paying relatively higher attention to the C-terminal regions—comparatively less related to chloroplast localization (Tardif et al., 2012)—of most non-CP-labeled proteins in the same dataset (lower right panel of Figure 3G). Concentrated attention at the N-terminal region of non-CPs likely reflects the presence of alternative targeting signals or the absence of canonical chloroplast transit peptide features, thereby providing strong negative evidence for chloroplast localization in a binary classification setting. Meanwhile, the minor attention peaks outside the N-terminal region of CPs are likely attributable to the internal sequence features correlated with CPs (Caspari, 2022) and background noise effects, rather than bona fide targeting signals. It is also noteworthy is that the approximate length of the sequences focused on in the CP-labeled proteins corresponds to the previously known average length of cTPs in C. reinhardtii (~50 AAs; arrowed in Figure 3G) (Tardif et al., 2012), strongly implying that, as intended, our model pays close attention to the N-terminal extensions (i.e., cTP regions) rather than treating every amino acid region equally when predicting the localization of putative chloroplast proteins.

To objectively validate the merit of Chlamy_ChloroPred, we evaluated the summary statistic values (e.g., accuracy, precision, recall, and F1 score) of the primary benchmark protein localization predictor, TargetP 2.0, using the train (n = 504; Figure 4A), validation (n = 169; Figure 4B), test (n = 169; Figure 4C), and total (n = 842; Figure 4D)—the aggregate total of the split datasets, which is the same as the balanced labeled dataset in Figure 1A—datasets in this study and then collated the performance indices of our model (Figures 3D–F) with those of TargetP 2.0. By and large, TargetP 2.0 exhibited higher precision values, but lower accuracy and recall values, across the datasets. Despite the higher precision values, the much lower recall values contributed to the inferior F1 scores of TargetP 2.0. The precision-recall trade-offs observed from TargetP 2.0—i.e., prioritizing precision over recall—implicate that the benchmark was likely programmed with strict, conservative classification criteria for chloroplast proteins. Given the increased accuracy (from 0.7396% to 0.8462%, a 14.41% improvement) and F1 score (from 0.6563% to 0.8395%, a 27.91% improvement) observed on the test dataset (Figures 3F,C), it is evident that Chlamy_ChloroPred outperforms the existing benchmark. To more comprehensively establish the position of our developed model among broader predictors, we further examined the predicted localizations of proteins in the test dataset (Supplementary Dataset S4) using TargetP 1.1 (Figure 4E), PredAlgo (Figure 4F), and PB-Chlamy (Figure 4G). As a result, Chlamy_ChloroPred demonstrated superior classification performance compared with the legacy predictors, including TargetP 1.1 and PredAlgo, in terms of accuracy (70.26% and 9.36% improvements, respectively), precision (81.63% and 17.07% improvements, respectively), and F1 score (281.24% and 6.74% improvements, respectively). When compared with the current SOTA model, PB-Chlamy, Chlamy_ChloroPred exhibited comparable and competitive binary classification performance, showing a modest improvement in recall (3.03%), accompanied by marginal reductions in accuracy (0.69%), precision (3.57%), and F1 score (0.15%). Collectively, these results suggest that Chlamy_ChloroPred represents a viable alternative to existing models, particularly in cases where reliable binary chloroplast protein classification in C. reinhardtii is required.

Although our program was constructed as a binary chloroplast protein classifier solely based on the C. reinhardtii proteome, we proceeded to use Chlamy_ChloroPred to predict protein localization for the chloroplast proteome of the model terrestrial plant, A. thaliana, in order to examine its potential compatibility with other photosynthetic organisms, which could lead to broader applications. In this versatility test, the TargetP 2.0 program was again employed as the comparative benchmark, given its validated multi-species and cross-proteome versatility. For the publicly available A. thaliana chloroplast proteome collected from the previous report and the TAIR database (Ferro et al., 2010; Reiser et al., 2017), Chlamy_ChloroPred classified 924 proteins as CPs out of 1,263 CP-labeled proteins in A. thaliana, whereas TargetP 2.0 predicted only 821 as CPs (Table 1). As a result, the accuracy and F1 score were 0.7316 and 0.8450, respectively, for our model and 0.6500 and 0.7879, respectively, for the benchmark. Thus, improvements of 12.6% and 7.25% were observed in the accuracy and F1 score of Chlamy_ChloroPred, respectively, compared to TargetP 2.0—although the precision and recall values are less meaningful in this evaluation due to the systematic absence of TN and FP cases, since the input dataset is composed entirely of CP-labeled data, which is fully biased. On balance, we confirmed that this model has the potential to be versatile for various photosynthetic organisms with the core photosynthetic organelle, the chloroplast, as evidenced by its superior prediction performance for the A. thaliana chloroplast proteome compared to the commonly referenced protein localization predictor for plants, TargetP 2.0 (Björnsdotter et al., 2021; Sanaboyana and Elcock, 2024).