Prediction of human pathogenic start loss variants based on iterative feature representation learning

Start loss variants can disrupt normal translation initiation, leading to proteins deletion or the production of alternative protein products. Distinguishing pathogenic variants from the large number of benign start loss variants remains a challenge in human genetics. Here, we present StartUp, a computational framework specifically designed for the pathogenicity prediction of start loss variants. Unlike traditional feature-based approaches that often fail due to missing or incomplete biological annotations, StartUp derives features directly from the local sequence context of the variant. By leveraging the positional characteristics of start loss variants, StartUp integrates deep features automatically extracted by a text convolutional neural network (TextCNN) with epigenetic modification profiles to capture informative patterns within the surrounding sequence context. In addition, StartUp employs an iterative feature representation strategy to progressively optimize input representations and enhance the extraction of deep feature signals. On the independent test set, StartUp achieved an area under the receiver operating characteristic curve (AUC) of 0.851 and an area under the precision-recall curve (AUPR) of 0.897. Overall, StartUp provides a start loss specific prediction framework with broad variant coverage and robust prediction performance, thereby facilitating the interpretation of start loss variants in clinical and research contexts.