The currently generated bootstrap dataset can be described as a statistically simulated dataset based on real deep mutational scanning (DMS) protein data (i.e., a resampled experimental-like dataset). The data type is protein sequence regression data, where each row contains a SARS-CoV-2 RBD mutant protein sequence of length 201 amino acids (sequence) along with the corresponding experimental phenotype values (such as label, mut_escape, or KD_ratio). The dataset is stored in a CSV tabular format, which is consistent with the structure of the original training data used in E2VD.

The data generation process uses bootstrap resampling, meaning that samples are randomly drawn with replacement from the original dataset. In addition, Gaussian noise is added to the resampled labels to simulate experimental measurement variability. This approach preserves the original mutation distribution and protein sequence structure while expanding the dataset size, making it suitable for subsequent deep learning model training, validation, and generalization experiments.

Adding Gaussian noise to the labels after bootstrap resampling serves two main purposes: to better mimic the measurement uncertainty in biological experiments and to improve the generalization ability of machine learning models.

In real deep mutational scanning (DMS) experiments, the phenotype values of the same mutant (such as binding affinity, expression level, or antibody escape score) are not perfectly fixed. They often fluctuate due to experimental conditions, sequencing errors, or variability across experimental replicates. Therefore, the recorded measurements inherently contain random noise. If only bootstrap resampling were applied without adding noise, the generated dataset would consist merely of repeated copies of the original samples, which could lead to model overfitting to the existing data distribution. By introducing Gaussian noise with mean 0 and a small variance, this method simulates experimental measurement errors while maintaining the overall statistical distribution (e.g., mean and standard deviation) of the data. As a result, it increases data diversity, reduces the risk of overfitting, and improves the model’s ability to generalize to unseen mutations or emerging viral variants.