Dataset and Models: MS2 Spectra Quality Prediction using Positive-Unlabeled (PU) Learning

This repository contains training data and pre-trained models for a pipeline that predicts MS2 spectrum quality. The model leverages MS1 spectra and instrument configuration parameters to assess the quality of consecutive MS2 scans. The approach uses Positive-Unlabeled (PU) Learning, training on positive examples (library-matched spectra) alongside unlabeled data to handle the challenge of incomplete ground truth labels. Repository Contents: Model Checkpoints (.ckpt): best_model_nnpu.ckpt: Final model trained with non-negative PU loss best_model_bce_positive.ckpt: Positive polarity model for prior estimation (BCE loss) best_model_bce_negative.ckpt: Negative polarity model for prior estimation (BCE loss) Datasets (Lance Format): lance_data_train_validation.tar.gz: Pre-processed training and validation data lance_data_test_set_1.tar.gz: Test set 1 (used for class prior approximation) lance_data_test_set_2.tar.gz: Test set 2 (used for decision threshold calculation) lance_data_test_set_3.tar.gz: Test set 3 (final evaluation) Analysis Results: sirius_results_combined.zip: Combined output folders generated by SIRIUS, containing result data in .tsv format.