Gene functional annotation via matrix completion

We present a matrix completion based multi-label learning model to handle the Gene functional annotation problem in this article. We consider the labels of Gene expression data as missing values of a construction matrix composed of expression data and label matrices. By assuming that the construction matrix as low rank and manifold structure, we can recover all the missing labels with minimization of its rank under constraints of data observed. The theories behind the low rank and manifold assumptions are the data correlation and similar data points sharing similar labels. Besides, the constructed matrix contains heterogeneous entries (continuous Gene expression data and binary labels), which existing works failed to model. To seal the gap between labels and expression data during prediction, we regularize our model with the maximum likelihood of unknown labels. A modified fixed-point continuation method is proposed to optimize the resulting nuclear norm minimization problem. The promising results of the proposed method on functional label predicting of Yeast gene microarray data have been shown in experiments.