PFP-WGAN: Protein function prediction by discovering Gene Ontology term correlations with generative adversarial networks

Abstract

Understanding the functionality of proteins has emerged as a critical problem in recent years due to significant roles of these macro-molecules in biological mechanisms. However, in-laboratory techniques for protein function prediction are not as efficient as methods developed and processed for protein sequencing. While more than 70 million protein sequences are available today, only the functionality of around one percent of them are known. These facts have encouraged researchers to develop computational methods to infer protein functionalities from their sequences. Gene Ontology is the most well-known database for protein functions which has a hierarchical structure, where deeper terms are more determinative and specific. However, the lack of experimentally approved annotations for these specific terms limits the performance of computational methods applied on them. In this work, we propose a method to improve protein function prediction using their sequences by deeply extracting relationships between Gene Ontology terms. To this end, we construct a conditional generative adversarial network which helps to effectively discover and incorporate term correlations in the annotation process. In addition to the baseline algorithms, we compare our method with two recently proposed deep techniques that attempt to utilize Gene Ontology term correlations. Our results confirm the superiority of the proposed method compared to the previous works. Moreover, we demonstrate how our model can effectively help to assign more specific terms to sequences.

Fig 1. The proposed method for protein function prediction.

Given the input sequence of amino acids, the generator has an embedding as its first layer which converts one-hot vectors to more compact representations. Then, one dimensional convolution filters are employed to explore meaningful sequential patterns. Then biochemical and biophysical features of the input are extracted from obtained activation maps. These features are used to predict GO annotations by the last fully connected layer. A discriminator judges about the validity of the obtained annotation for this sequence by observing pairs of protein sequences and their experimentally approved annotations in SwissProt. By observing proteins which are annotated by a common set of GO terms, discriminator could extract correlations between GO terms.

Fig 2. Comparison of BLAST, DeepGO-Seq and PFP-WGAN on dataset 1.

The F_max measure shows the superiority of PFP-WGAN in all three parts of the GO.

Fig 3. Differences between average F1 obtained for PFP-WGAN and DeepGO-Seq for the GO terms in each height (F¯P and F¯D).

In the BP branch (as the most important part of GO with a large number of terms) differences are increased when moving through the deeper terms. In the most and half parts of the charts for CC and MF branches we can observe this pattern too.

Fig 4. F1 obtained for PFP-WGAN and DeepGO-Seq for GO terms as a function of number of available positive training samples.

The improvement which is obtained by PFP-WGAN for rare terms is more considerable comparing to terms with large numbers of positive samples.

Fig 5. Comparison of BLAST, FFPRED, CSSAG, STDNN, MTDNN and PFP-WGAN on dataset 2.

The F_max measure shows the superiority of PFP-WGAN in all three parts of the GO.

Fig 6. Sensitivity of the PFP-WGAN on parameter λ₁.