Deep learning for functional microRNA target prediction and breast cancer prediction

 Deep learning for functional microRNA target prediction and breast cancer prediction

TEAM:

Ana Coporan

Alexandru Rusmir

MicroRNA target prediction

In the paper, TargetNet is introduced, a novel deep learning-based algorithm for functional miRNA target prediction. It has 3 key components:

    1. Selection
TargetNet uses relaxed candidate target sites(CTS) - generally conservative CTS selection criteria is used. Here a sliding window is used to scan the mRNA.

    2. Input encoding
TargetNet uses a novel encoding scheme for miRNA-CTS sequences to incorporate extended seed alignment information. Novel encoding refers to the process of converting a text document into a numerical representation.

    3. Residual Network

Let fk,n be a one-dimensional convolution layer, where k and n denote the filter lengths and the number of filters. Each convolution layer is followed by a rectified linear unit (ReLU) activation function and dropout regularization.Firstly we have an input that goes through a convolution layer with 16 filters of length 5 => The output is feed into the two residual blocks each consisting of two convolution layers(f3,16 and f3,32) => The output is fed into a max-pooling layer and a dense layer. We use a sigmoid function as an activation function for the dense layer.

For training of the ResNet prediction model, we use binary crossentropy objective function defined as:

Conclusion: The experimental results support that TargetNet not only demonstrates significant performance improvements in functional miRNA target classification

Breast cancer prediction

This study analyzes how deep learning can be applied to the prognosis of breast cancer. The data used is about 116 women, from which 64 women were diagnosed with breast cancer.

The solution:

Experiments using k-fold Monte Carlo cross-validation and feature selection were conducted to evaluate the ability that machine learning and deep learning classifiers have to differentiate between patients with and without cancer.  For feature selection a recursive minimum redundancy maximum relevance (MRMR) algorithm of SULOV-gradient boosting was applied.

The performance of the machine learning and deep learning classifiers was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC). The performance of the deep learning algorithm was compared against the AUC of seven machine learning algorithms: XGBoost, stochastic gradient, support vector machines, random forests, neural networks, naive Bayes and logistic regression.

The best predictors selected by the SULOV algorithm were used in the deep learning and machine learning algorithms. In the deep learning algorithm, a sigmoid activation function was used for the output layer of the neural networks, due to the binary nature of the target variable. The highest predictive power was obtained with a rectified linear activation function (ReLU) in the input layer, four hidden layers with 1x10^2 nodes, a L1-L2 regularizer in the kernel function and a L2 regularizer for the bias and the activation functions of the hidden units.

Conclusion:

Sources: 

2107.11381v2.pdf (arxiv.org)

2302.02406.pdf (arxiv.org)