engleski

Transfer Learning for Improved Peptide Activity Prediction on Small Dataset

In recent years, deep neural networks have been successfully applied to peptide activity prediction as a part of in-silico virtual screening process. Such methods can guide the discovery of novel active peptides by providing prospective peptide sequences, and thus saving time and resources. However, such methods require significant amounts of collected data, which poses a huge obstacle for their application in the case of poorly researched peptide activities. In this work, we investigate how the predictive power of a deep learning model can be improved by transferring knowledge (i.e. transfer learning ; TL) from a larger category of peptides to the category of interest represented by a smaller dataset. The model is first pretrained on a large dataset and then fine-tuned by optimizing relevant hyperparameters (e.g. per layer learning rates, which weights to transfer and dropout rates) on a smaller target dataset, and consequently uses the knowledge gained during pretraining to solve the target task. In our case study, one antiviral and one antimicrobial dataset obtained from DRAMP 2.0 complemented with negative instances from Uniprot database are used for pretraining, while the target dataset is the antiviral AVPdb database. We represent the peptides by encoding each amino acid with 94 physico-chemical properties (e.g. hydrophobicity, Cruciani properties, etc.) A model based on convolutional and recurrent layers is used in which convolutional layers learn to detect short patterns while the recurrent layer learns long-term dependencies of short patterns in the input sequence. The convolutional filters estimated during pretraining are preserved, which creates a better starting point in comparison to training the entire model "from scratch". By using sequential feature selection, the dependency of ROC-AUC score on the number of selected features is investigated for baseline and TL models. Preliminary results demonstrate that TL-based models outperformed the baseline model, either by pretraining on a database of the same or different peptide activity. This suggests that some of the short patterns responsible for antimicrobial activity are also responsible for antiviral activity and they can be exploited to improve the predictive power.

Transfer learning ; Machine learning ; Peptide activity prediction

nije evidentirano