Аспирантский семинар: Peptide-spectrum matching using generative modeling with regularization for identification in shotgun proteomics
Мероприятие завершено
Where: Faculty of Computer Science, Kochnovskii proezd, 3, room 400.
When: 18:10 – 19:30 on November 26.
Speaker: Pavel Sulimov.
Score functions lie at the heart of software tools for analyzing tandem mass spectrometry data in shotgun proteomics. They essentially count the common ions in the experimental and theoretical spectra and can consider, upon user-request, other satellite ions related to secondary fragmentation products resulting from, for instance, carbon monoxide, ammonia, water, or phosphoric acid losses. Standard score functions, such as XCorr in SEQUEST or HyperScore in X!Tandem, take these ions into account with different, but manually calibrated weights. We first argue that XCorr is analogous to a hand-crafted, manually weighted graphical model, a Restricted Boltzmann Machine (RBM). Then we utilize full RBMs as score functions and train its parameters directly from observed spectrum data. The training is fully unsupervised and uses different regularization heuristics, so it does not require a good quality of annotated training data contrary to other learned scoring functions.
When: 18:10 – 19:30 on November 26.
Speaker: Pavel Sulimov.
Score functions lie at the heart of software tools for analyzing tandem mass spectrometry data in shotgun proteomics. They essentially count the common ions in the experimental and theoretical spectra and can consider, upon user-request, other satellite ions related to secondary fragmentation products resulting from, for instance, carbon monoxide, ammonia, water, or phosphoric acid losses. Standard score functions, such as XCorr in SEQUEST or HyperScore in X!Tandem, take these ions into account with different, but manually calibrated weights. We first argue that XCorr is analogous to a hand-crafted, manually weighted graphical model, a Restricted Boltzmann Machine (RBM). Then we utilize full RBMs as score functions and train its parameters directly from observed spectrum data. The training is fully unsupervised and uses different regularization heuristics, so it does not require a good quality of annotated training data contrary to other learned scoring functions.
Дата
26 ноября, 2018 г. 18:10
В статье упомянуты