Enabling full-length evolutionary profiles based deep convolutional neural network for predicting DNA-binding proteins from sequence.

Sequence based DNA-binding protein (DBP) prediction is a widely studied biological problem. Sliding windows on position specific substitution matrices (PSSMs) rows predict DNA-binding residues well on known DBPs but the same models cannot be applied to unequally sized protein sequences. PSSM summaries representing column averages and their amino-acid wise versions have been effectively used for the task, but it remains unclear if these features carry all the PSSM's predictive power, traditionally harnessed for binding site predictions. Here we evaluate if PSSMs scaled up to a fixed size by zero-vector padding (pPSSM) could perform better than the summary based features on similar models. Using multilayer perceptron (MLP) and deep convolutional neural network (CNN), we found that (a) Summary features work well for single-genome (human-only) data but are outperformed by pPSSM for diverse PDB-derived data sets, suggesting greater summary-level redundancy in the former, (b) even when summary features work comparably well with pPSSM, a consensus on the two outperforms both of them (c) CNN models comprehensively outperform their corresponding MLP models and (d) actual predicted scores from different models depend on the choice of input feature sets used whereas overall performance levels are model-dependent in which CNN leads the accuracy.

ECG anomaly class identification using LSTM and error profile modeling.

Automatic diagnosis of cardiac events is a current problem of interest in which deep learning has shown promising success. We have earlier reported the use of Long Short Term Memory (LSTM) networks-trained on normal ECG patterns-to the detection of anomalies from the prediction errors for real-time diagnostic applications. In this work, we extend our anomaly detection algorithm by introducing a second stage predictor that can identify the actual anomaly class from the error outputs of the first stage model. Results from seven types of anomalies have been presented including Atrial Premature Contraction (APC), Paced Beat (PB), Premature Ventricular Contraction (PVC), Right Bundle Branch Block (RBBB), Ventricular Bigeminy (VB), Ventricular Couplets (VCs) and Ventricular Tachycardia (VT). To optimize anomaly class prediction performance, multiple choices of second stage models such as multilayer perceptron (MLP), support vector machine (SVM) and logistic regression have been employed. A featurization scheme for LSTM prediction errors in the form of overall summaries has been proposed and a successful predictor for the same was developed with good performance. Our results indicate that the error vectors represented by their summary features carry useful predictive information about actual ECG anomaly type. We discuss how the accuracy scores without attention to inherent class imbalances and paucity of data instances may produce misleading performance estimates and hence accurate background models are needed to estimate true predictive performance of multi-class predictors such as those presented in this work. The training data sets and related resources for this study are provided at http://ecg.sciwhylab.org.

A Comparison of Shallow and Deep Learning Methods for Predicting Cognitive Performance of Stroke Patients From MRI Lesion Images.

Stroke causes behavioral deficits in multiple cognitive domains and there is a growing interest in predicting patient performance from neuroimaging data using machine learning techniques. Here, we investigated a deep learning approach based on convolutional neural networks (CNNs) for predicting the severity of language disorder from 3D lesion images from magnetic resonance imaging (MRI) in a heterogeneous sample of stroke patients. CNN performance was compared to that of conventional (shallow) machine learning methods, including ridge regression (RR) on the images' principal components and support vector regression. We also devised a hybrid method based on re-using CNN's high-level features as additional input to the RR model. Predictive accuracy of the four different methods was further investigated in relation to the size of the training set and the level of redundancy across lesion images in the dataset, which was evaluated in terms of location and topological properties of the lesions. The Hybrid model achieved the best performance in most cases, thereby suggesting that the high-level features extracted by CNNs are complementary to principal component analysis features and improve the model's predictive accuracy. Moreover, our analyses indicate that both the size of training data and image redundancy are critical factors in determining the accuracy of a computational model in predicting behavioral outcome from the structural brain imaging data of stroke patients.