In this study, we utilized the rice dataset constructed by Lv et al. (2020) for model training and evaluation using 5-fold cross-validation. To ensure the high quality of the data, sequences with greater than 80% similarity were removed via the CD-HIT program (Li and Godzik, 2006). The dataset is made of 154,000 sequences with 6mA sites and 154,000 sequences without 6mA sites. This is a widely adopted and balanced rice dataset. During model training, unbalanced datasets may lead to unreliable results. The majority class samples are dominant and the model will favor the majority class during training, thus ignoring the minority class. This may result in the model having high accuracy for the majority class but low recognition for the minority class during prediction. For ease of reference, we denote it as “rice-Lv” throughout this study. Both positive and negative sequences in the rice-Lv dataset are 41 base pairs in length. Positive sequences represent 6mA modifications at their centers, while negative sequences lack such modifications at theirs. By employing this well-established dataset, we enable a fair comparison between our method and those previously reported.

The architecture of iRice6A-LMXGB is presented in Figure 1 , comprising two main components: the pre-trained DNABERT-2 module and the XGBoost module. DNABERT-2 is a pre-trained BERT model specifically designed for encoding DNA sequences. It can efficiently identify complex long-range dependencies in these sequences (Zhou et al., 2023). And this module will undergo further fine-tuning in this study. XGBoost’s superior performance, particularly in terms of speed and accuracy when processing large-scale datasets, enables its extensive use in solving classification problems (Chen and Guestrin, 2016; Yang et al., 2021). It utilizes the feature vectors output from the DNABERT-2 model to generate final prediction results. A detailed explanation of the model follows.

In this study, we validate our approach using a traditional 5-fold cross-validation method and compare it to previous studies based on the benchmark dataset rice-Lv.

we will combine five metrics, including accuracy (ACC), sensitivity (Sn), specificity (Sp), Matthew’s correlation coefficient (MCC), and area under the curve (AUC), to comprehensively evaluate the prediction performance of our model (Zou et al., 2023; Zulfiqar et al., 2023; Guo et al., 2024; Huang et al., 2024; Zhu et al., 2024).

ACC indicates the overall correctness of the model prediction and is a basic benchmark used to evaluate the model performance, which can be expressed as:

The sensitivity Sn, also known as the true positive rate (TPR), is expressed as:

The specificity Sp, also known as the true negative rate (TNR), is expressed as:

MCC is a composite metric that assesses the overall quality of classification model predictions by examining the performance of the classification model in each of the four quadrants of the confusion matrix. The superior score reflects the balanced excellence between true positives (TP), true negatives (TN), false negatives (FN) and false positives (FP). It can be defined as:

The last performance metric we use is AUC, defined as the value of the area under the subject’s operating characteristic curve. AUC is also an important measure of the performance of a dichotomous model. The larger the value of AUC, the better the model performs. AUC is a floating-point number between 0 and 1. 1 indicates that the model predicts perfectly, whereas 0.5 indicates that the model is similar to a random prediction (Zhang et al., 2025).

In this study, we developed three models. For the first model, we directly used the pre-trained DNABERT-2 to extract 768-dimensional features from rice DNA and fed them into an XGBoost classifier for prediction tasks. The XGBoost classifier shows unique advantages in genomics data classification tasks, mainly due to its ability to efficiently handle high-dimensional sparse data and its built-in regularization mechanism. Our dataset, with more than 300,000 samples, is characterized by high feature dimensionality, and XGBoost is able to efficiently capture nonlinear interaction effects through the gradient boosting framework combined with second-order derivative optimization. Its regularization term can in turn suppress overfitting and enhance model generalization (Chen and Guestrin, 2016). Cross-validation results showed ACC=0.6259, Sn=0.6207, Sp=0.6312, MCC=0.2519, and auROC=0.6728 for this configuration. For the second model, we loaded the rice-Lv dataset into the DNABERT-2 module and conducted 200 iteration loops to develop a fine-tuned version of the model. The 5-fold cross-validation scores were ACC=0.9903, Sn=0.9898, Sp=0.9907, MCC=0.9805, auROC=0.9994 which are 58.22%, 59.47%, 56.96%, 289.24%, and 48.54%, respectively higher than those of the non-fine-tuned model. For the third model, we utilized LightGBM’s built-in function to assess and prioritize feature importance using features extracted from the fine-tuned DNABERT-2 model (Ke et al., 2017). The feature ranking principle of LightGBM is based on the Gradient Boosting Decision Tree (GBDT) framework, which evaluates feature importance by quantifying the contribution of features in the process of constructing the decision tree (Ke et al., 2017). Following this, we selected the top 300 features for modeling with XGBoost. The 5-fold cross-validation yielded ACC=0.9899, Sn=0.9890, Sp=0.9908, MCC=0.9799, and auROC=0.9994. As shown in Figure 2 , our cross-validation results indicate that: (1) Fine-tuned models outperformed non-fine-tuned counterparts significantly. (2) However, applying feature selection after fine-tuning caused minor performance degradation compared to models without feature selection, not much difference overall. These findings demonstrate the effectiveness of our fine-tuning strategy. The pre-training model is usually trained on multi species datasets, and may not be able to capture the 6mA distribution pattern unique to rice. Through the fine-tuning strategy, the model parameters are recalibrated, which can give priority to the local features in the rice genome, and the sensitivity of the model to the sequence context of rice 6mA is improved. Additionally, while XGBoost’s tree-based architecture excels at managing high-dimensional data through regularization techniques, our results suggest that applying LightGBM-based feature selection after fine-tuning may slightly reduce model performance due to fewer feature interactions. We selected the second model with the best performance, performing fine-tuning for 200 iterations without feature selection, to name iRice6mA-LMXGB.

In order to perform an in-depth analysis of the interpretability of the iRice6mA-LMXGB model after integrating DNABERT-2 with XGBoost, we used the UMAP (Uniform Manifold Approximation and Projection) technique. This is a nonlinear dimensionality reduction and visualization algorithm for large-scale datasets. Umap assumes that the data is distributed on a low dimensional manifold. Firstly, the probability weight is defined in the high dimensional space using the neighborhood graph to reflect the similarity between points. Then the cross entropy loss function is used to optimize the embedding in the low dimensional space to align the low dimensional similarity with the high dimensional structure. Based on graph theory and flow learning methods, it is assumed that the available data samples are uniformly distributed in the topological space and can be approximated and mapped from these finite data samples to a lower-dimensional space for visualization and analysis (McInnes and Healy, 2018).

To be more specific, we will visualize the distribution of 6mA and non-6mA by projecting each feature vector onto a 2D view using the UMAP technique. Figure 3 shows the arrangement of 6mA and non-6mA samples in 2D space before and after fine-tuning, and the decision boundary drawn in black by the XGBoost algorithm. Blue markers denote non-6mA samples, and orange markers denote 6mA samples. The first subplot represents the UMAP results of the original features without fine-tuning, which can be interpreted as all the sample points not showing any representative clustering. In Figure 3A , poor separation indicates significant feature overlap between the 6mA sample points and the non-6mA sample points ( Figure 3A ), suggesting a high degree of overlap in their distributions. The second subfigure shows the results of projecting the high-dimensional feature space learned from the iRice6mA-LMXGB model into a 2D view, which shows much improved clustering, indicating a significant increase in separation and a decrease in overlap in the feature space ( Figure 3B ), resulting in improved performance. In summary, our approach allows for better learning of model decision boundaries. Through this visualization technique, we can more intuitively understand the impact of features on model predictions, further deepening our exploration of model interpretability.

To better evaluate the performance of our model, we compare it with the following state-of-the-art methods, including MM-6mAPred (Pian et al., 2019), iDNA6mA-Rice (Lv et al., 2019), SNNRice6mA (Yu and Dai, 2019), iRicem6A-CNN (Lv et al., 2021), ENet-6mA (Abbas et al., 2022), Deep6mA (Li Z. et al., 2021) and SpineNet-6mA (Abbas et al., 2020). Our model is evaluated using the same five-fold cross-validation protocol on the same dataset as previous studies, employing the identical metrics: ACC, MCC, Sn, Sp, and AUC. As shown in Table 1 , our iRice6mA-LMXGB model outperforms all previous predictors across all metrics and demonstrates more stable performance with less fluctuation in ACC, MCC, Sn, Sp, and AUC values. In ACC, MCC, Sn, and AUROC metrics, our model improves over the previous best predictor SpineNet-6mA by 5%, 11.42%, 3.42%, 6.62%, and 1.98%, respectively. Furthermore, it outperforms the previous best model, ENet-6mA, by 6.08% in Sp metric. To facilitate visualization of the comparison results, we created a box-and-whisker plot, as illustrated in Figure 4 . To sum up, our iRice6mA-LMXGB model demonstrates superior performance compared to both machine learning-based and CNN/LSTM-based deep learning models for 6mA prediction in rice, showcasing its robustness as a predictive tool.