—Machine Learning (ML) methods have been used with varying degrees of success on protein prediction tasks, with two inherent limitations. First, prediction performance often depends upon the features extracted from the proteins. Second, experimental data may be insufficient to construct reliable ML models. Here we introduce MP3vec, a transferable representation for protein sequences that is designed to be used specifically for sequence-to-sequence learning tasks. We use transfer learning to generate the MP3vecs by training a deep neural network on the source problem of protein secondary structure prediction, and then extracting representations learned by the trained network for use in related downstream prediction tasks. ML methods using MP3vecs perform as well as the state-of-the-art (or better) on the target problems, while being orders of magnitude faster in terms of training time. We suggest that MP3vec can act as a strong baseline for comparative work on the use of ML in protein-prediction tasks; and for future extensions with domainspecific features.
Autorzy
- Sanket Rajan Gupte,
- Dharm Skandh Jain,
- Ashwin Srinivasan,
- Raviprasad Aduri
Informacje dodatkowe
- DOI
- Cyfrowy identyfikator dokumentu elektronicznego link otwiera się w nowej karcie 10.1109/bibm49941.2020.9313301
- Kategoria
- Aktywność konferencyjna
- Typ
- publikacja w wydawnictwie zbiorowym recenzowanym (także w materiałach konferencyjnych)
- Język
- angielski
- Rok wydania
- 2020
