Embeddings and benchmarks results for fusionAI dataset

Dataset Description for Genomic Foundation Model Benchmarking This dataset accompanies the research article "Benchmarking Genomic Foundation Models for Binary Classification of Gene Fusion Breakpoints from DNA Sequences." It is designed to facilitate the reproduction and extension of the gene fusion detection experiments performed using various state-of-the-art genomic foundation models. Included Files and Content The dataset is organized into 4 archives: embeddings.zip, embeddings_augmented and two separate model archives (learned_models_svm.zip and learned_models_nn.zip). embeddings.zip This archive contains the calculated middle embeddings derived from raw DNA sequences published https://compbio.uth.edu/FusionGDB2/FusionAI/ . There are embeddings for two fusion sequences for training ( https://compbio.uth.edu/FusionGDB2/FusionAI/fg_newdata_train.txt ) and testing data ( https://compbio.uth.edu/FusionGDB2/FusionAI/fg_newdata_test.txt ) . These embeddings serve as the input features for the downstream classification models. Embedding Types: For every sequence and model, two distinct pooling strategies are provided: Middle Embedding: The hidden state extracted specifically from the central position of the sequence. Mean Embedding: The average of all embeddings across the sequence length. Data Splits: Includes embeddings for both training (fg_newdata_train.txt) and testing (fg_newdata_test.txt) datasets. Models Included: Nucleotide Transformer, EVO2, Hyena, and BERTDNA. embeddings_augmented.zip This archive contains data used to evaluate model robustness against sequence perturbations. embeddings_augmented/: Contains perturbed versions of both mean and middle embeddings for all four models. Methodology: A subset of 800 samples was selected and randomly augmented. Augmentation Parameters: Samples were modified by a random noise injection (1–32%) or a positional shift (1 to 64 units). embeddings_augmented_results/: Contains the classification results and performance metrics of these augmented samples when processed by the pre-trained models. learned_models_svm.zip This archive contains the saved Support Vector Machine (SVM) classification models trained on the foundation model embeddings. learned_models_nn.zip This archive contains the saved DenseNet-based Neural Network (NN) classification models trained on the foundation model embeddings and FusionAI classifier.