Browsing by Subject "Zero-Shot Learning"

Now showing 1 - 2 of 2

Open Access
Deepkinzero: zero-shot learning for predicting kinase phosphorylation sites
(2018-08) Deznabi, Iman
Protein kinases are a large family of enzymes that catalyze the phosphorylation of other proteins. By acting as molecular switches for protein activity, the phosphorylation events regulate intracellular signal transduction, thereby assuming a central role in a broad range of cellular activities. On the other hand, aberrant kinase function is implicated in many diseases. Understanding the normal and malfunctioning signaling in the cell entails the identification of phosphorylation sites and the characterization of their interactions with kinases. Recent advances in mass spectrometry enable rapid identification of phosphosites at the proteome level. Alternatively, there are many computational models that predict phosphosites in a given input protein sequence. Once a phosphosite is identified, either experimentally or computationally, knowing which kinase would catalyze the phosphorylation on this particular site becomes the next question. Although a subset of available computational methods provides kinase-specific predictions for phosphorylation sites, due to the need for training data in such supervised methods, these tools can provide predictions only for kinases for which a substantial number of the phosphosites are already known. A particular problem that has not received any attention is the prediction of new sites for kinases with few or no a priori known sites. None of the current computational methods which rely on the classical supervised learning settings can predict additional sites for this kinases. We present DeepKinZero, the first zero-shot learning approach, that can predict phosphosites for kinases with no known phosphosite information. DeepKinZero takes a peptide sequence centered at the phosphorylation site and learns the embeddings of these phosphosite sequences via a bi-directional recurrent neural network, whereas kinase embeddings are based on protein sequence vector representations and the taxonomy of kinases based on their functional properties. Through a compatibility function that associates the representations of the site sequences and the kinases, DeepKinZero transfers knowledge from kinases with many known sites to those kinases with no known sites. Our computational experiments show that DeepKinZero achieves a 30-fold increase in accuracy compared to baseline models. DeepKinZero complements existing approaches by expanding the knowledge of kinases through mapping of the phosphorylation sites pertaining to understudied kinases with no prior information, which are increasingly investigated as novel drug targets.
Open Access
Fine-grained object recognition in remote sensing imagery
(2018-06) Sümbül, Gencer
Fine-grained object recognition aims to determine the type of an object in domains with a large number of sub-categories. The steadily increase in spatial and spectral resolution entailing new details in remote sensing image data, and consequently more diversi ed target object classes having subtle di erences makes it an emerging application. For the approaches using images from a single domain, widespread fully supervised algorithms do not completely t into accomplishing this problem since target object classes tend to have low between-class variance and high within-class variance with small sample sizes. As an even more arduous task, a method for zero-shot learning (ZSL), in which identi cation of unseen sub-categories is tackled by associating them with previously learned seen subcategories when there is no training example for some of the classes, is proposed. More speci cally, our method learns a compatibility function between image representation obtained from a deep convolutional neural network and the semantics of target object sub-categories explained by auxiliary information gathered from complementary sources. Knowledge transfer for unseen classes is carried out by maximizing this function throughout the inference. Furthermore, bene tting from multiple image sensors can overcome the drawbacks of closely intertwined sub-categories that limits the object recognition performance. However, since multiple images may be acquired from di erent sensors under di erent conditions at di erent spatial and spectral resolutions, they may be geometrically unaligned correctly due to seasonal changes, di erent viewing geometry, acquisition noise, an imperfection of sensors, di erent atmospheric conditions etc. To address these challenges, a neural network model that aims to correctly align images acquired from di erent sources and to learn the classi cation rules in a uni ed framework simultaneously is proposed. In this network, one of the sources is used as the reference and the others are aligned with the reference image at representation level throughout a learned weighting mechanism. At the end, classi cation of sub-categories is carried out with a feature-level fusion of representations from the source region and estimated multiple target regions. Experimental analysis conducted on a newly proposed data set shows that both zero-shot learning algorithm and the multi-source ne-grained object recognition algorithm give promising results.