Long shortterm memory networks successfully addressed these difficulties.Based on the protein sequence segment, we first extracted multi perspective nominal features from five kinds of descriptors, such as onehot encoding, positionspecific propensity matrix and informative physicochemical properties. The highranked selected features were then fed as input to train the stacked LSTM model.A middle subsequence with a probed phosphorylation site is considered positive; otherwise, the site is considered negative.To avoid biasness, CDHIT with a identity cutoff was applied to each of the positive and negative sets to remove surplus subsequences.Onehot encoding features strategy replicates the types and relative positions of amino acids around phosphorylation sites in protein sequences.Therefore, this study adopted the onehot encoding scheme to transform protein fragments into numeric abstraction.In doing so, each of the different amino acids was encoded into a dimensional vector, which contains only s and s.The amino acids were arranged in order of ARNDCQILKMFPSTWYVX, where amino acid A, tyrosine Y was represented by and the virtual residue X was denoted by the vector. The standard weight function was inducted to contemplate the grouping weight coding of protein sequences.This study implemented a similar concept to extract numerical descriptors from protein sequences of phosphorylation sites.The amino acid residue is partitioned on the basis of the following disjoint groups. The three sequence features obtained in the form of three vectors were integrated to form a L dimension vector, represented by V, where V encoding was based upon the grouped weight of protein sequence X.In the PSPM feature representation scheme, the dataset was first divided into positive and negative datasets, where the positive dataset possessed phosphorylation sites, whereas the negative dataset was comprised of nonphosphorylation sites.If a positive dataset consists oflnumber of sample fragments, and every sample fragment length is m, the value ofm will be determined empirically as there is no theoretical justification for it.In the current work, for a given protein fragment, we extracted the nominal descriptors based on physicochemical properties for the prediction of phosphorylation sites. Fscore is a simple but ample algorithm to evaluate the discriminative power of each feature in the feature set. Given the ith feature vector with N number of samples, where the total numbers of positive and negative samples are n andn k,ii x represents mean values of an ith feature of entire positive and negative samples, respectively, xi is the mean value of ith feature of total samples.Similarly,x k,i indicates the value of ith feature of kth sample in a positive and negative dataset, respectively.The numerator shows discriminator between positive and negative sample sets, and the denominator is the sum of deviation within each feature set.To find the underlying cause of this issue, LSTM equipped with incredible network architecture, which can learn longterm dependency information naturally, was implemented.The general architecture of LSTM consists of an input gate, forget gate, update gate, and a memory block.The primary difference between the LSTM network and other deep learning networks is in the formers usage of complex memory cells as an alternative to the usage of neurons of the public network.