Antifreeze proteins (AFPs) are the sub-set of ice binding proteins indispensable for the species living in extreme cold weather. These proteins bind to the ice crystals, hindering their growth into large ice lattice that could cause physical damage. There are variety of AFPs found in numerous organisms and due to the heterogeneous sequence characteristics, AFPs are found to demonstrate a high degree of diversity, which makes their prediction a challenging task. Herein, we propose a machine learning framework to deal with this vigorous and diverse prediction problem using the manifolding learning through composition of k-spaced amino acid pairs. We propose to use the deep neural network with skipped connection and ReLU non-linearity to learn the non-linear mapping of protein sequence descriptor and class label. The proposed antifreeze protein prediction method called AFP-CKSAAP has shown to outperform the contemporary methods, achieving excellent prediction scores on standard dataset. The main evaluater for the performance of the proposed method in this study is Youden's index whose high value is dependent on both sensitivity and specificity. In particular, AFP-CKSAAP yields a Youden's index value of 0.82 on the independent dataset, which is better than previous methods.