A Machine Learning-based Accurate Approach for Inferring Potential LncRNA-disease Associations

The identification of underlying disease-associated long non-coding RNA (LncRNAs) helps to understand the underlying molecular mechanisms of disease and thus contribute to its treatment, diagnosis, and prognosis. Therefore, developing effective computational strategies to discover novel associations between LncRNA and disease is of great significance for further experimental verification. Traditional algorithms based on biological networks often have limitations of graph scale. Using machine learning techniques may avoid this issue. This paper proposes a machine learning-based approach, which utilizes Multilayer Perceptrons along with Feature Reduction to predict LncRNA and Disease Associations (MPFRLDA). First of all, a restart random walk algorithm is implemented to construct the topological similarity, which fully considers the topological characteristics of a network. Then, the sample features are extracted through matrix factorization and statistical measures, and their dimension is reduced via principal component analysis. Besides, the problem of the lack of negative samples for training is solved by randomly sampling from the unlabeled associations. Finally, the outputs of the multilayer perceptron constitute the distribution of expectations for the final rank of the associated pairs. Consequently, the model performance of MPFRLDA is superior to other classic algorithms, and the label randomization test indicates the proposed algorithm avoids overfitting. In addition, the candidate LncRNAs of colorectal and breast cancers predicted by MPFRLDA in case studies have been experimentally verified. Therefore, MPFRLDA appears to be an effective computing resource in biomedical research.