No machine is more sophisticated than the cell. This is because there are too many sophisticated mechanisms in the cell, including transcription, gene splicing, translation, and posttranslational modification (PTM). All constituted the sophisticated life. As a key mechanism, PTM increases not only diversities of protein structures and functions but also make regulation more sophisticated. Many studies indicated that aberrant of PTM was always implicated in many human diseases including cancer (Martin et al., 2011; Nakamura et al., 2015; Junqueira et al., 2019). Propionylation, one of more than 400 types of PTM, was firstly discovered in histone in 2007 (Chen et al., 2007), and later in nonhistone (Cheng et al., 2009). The propionylation was a dynamic process where propionyl group was conjugated by some acetyltransferases to substrate proteins which was thus propionylated and could be removed by Sirt1 and Sirt2 (Chen et al., 2007; Leemhuis et al., 2008; Zhang et al., 2008; Cheng et al., 2009). Although it was known that lysine propionylation played a regulating role in the metabolism (Yang et al., 2019) and was a mark of active chromatin (Kebede et al., 2017), many of its unknown functions were still not uncovered.

Identifying propionylation sites was crucial to further explore functions of propionylated proteins. The mass spectrometry has been developed to detect propionylation sites in the past decades and obtained vast achievements (Chen et al., 2007). However, this technique was time consuming and labor intensive. Another alternative was computational methods which learned a model from the known data and then gave the predictions for unknown data. The process was similar with learning of human. In the past 30 years, more than 100 computational methods or tools have been developed for predicting PTM sites (Huang and Zeng, 2016; Zhou et al., 2016; Ai et al., 2017; Wei et al., 2017, 2019; Xiang et al., 2017; Chen et al., 2018; de Brevern et al., 2018; Ning et al., 2018, 2019; Xie et al., 2018; Huang et al., 2019, 2020; Luo et al., 2019; Malebary et al., 2019; Wang et al., 2019; Lv et al., 2020; Qian et al., 2020; Thapa et al., 2020). For example, Malebary et al. (2019) proposed a computational model for lysine crotonylation prediction by integrating various position and composition relative features along with statistical moments, and reached the average accuracy of 0.9917 in the experimental dataset. Chen et al. (2018) presented a computational tool named ProAcePred to predict prokaryote lysine acetylation sites by extracting sequence-based, physicochemical property and evolutionary information features. Wei et al. (2017, 2019) used sequence-based information to build computational models for predicting phosphorylation sites and protein methylation sites, respectively. Although propionylation was a newly discovered PTM, there still were two computational methods developed to detect propoinylation sites. One was that the biased support vector machine (SVM) model (Ju and He, 2017) which incorporated four different sequence features into Chou’s pseudo-amino acid composition. Another was the PropSeek which was also a SVM model and which exploited evolutionary information, sequenced-derived information, predicted structural information, and feature annotations (Wang et al., 2017). Advance in deep learning techniques could accelerate development of propionylation prediction. A well-known example was that the AlphaFold, a deep-learning-based method, accurately determined a protein’s 3D shape from its amino-acid sequence (Callaway, 2020). The detection of protein structure especially in more than two dimensions was one of biology’s grandest challenges and to date no better technique can solve this issue. In this paper, we attempted to build a deep learning model to predict propionylation sites. However, the accumulated propionylation data was too small to better train deep learning model. Lysine propionylation has in situ crosstalk with lysine malonylation. Wang et al. (2017) statistically compared 1,471 propionylation sites in 605 proteins with the dataset of 1,745 malonylation sites in 595 proteins and found that 600 (40.8%) of 1,471 propionylation sites are overlapped with malonylation. What is more, the number of malonylation was much more than that of propionylation sites. Inspired by this, we proposed a transfer learning method for predicting propionylation sites. We firstly constructed a recurrent neural network (RNN)-based deep learning model, which was trained by the malonylation data. The model was then fine tuned by the propionylation data. The model served as feature extractor. Finally, the SVM-based classifier was trained to discriminate propionylation from nonpropionylation.

All lysine propionylation sites were both from the protein lysine modifications database (PLMD) (Xu et al., 2017) and Uniprot database (UniProt Consortium, 2018). The PLMD was devoted to collect lysine modification, currently hosting 284,780 modification events in 53,501 proteins for 20 types of lysine modification such as ubiquitination, methylation, and sumoylation. The Uniprot was a comprehensive database of protein sequence and function annotation. We firstly downloaded 192 proteins containing 413 propionyllysine sites from the PLMD http://plmd.biocuckoo.org/download.php. We then retrieved 18 propionylation proteins from the Uniprot database. After merging two dataset of proteins and removing repeated proteins, we obtained 207 unique proteins. Functions of protein including protein modification would rely more or less on homology. To reduce or remove influences of homology on the proposed method, we applied the sequence clustering software CD-HIT (Li and Godzik, 2006) to perform sequence clustering. The sequence identity was set to 0.7. Finally, we obtained 189 proteins as experimental data, of which sequence similarities between any two was less than 0.7. We selected randomly 4/5 of 189 proteins (151) as positive training samples which containing 304 sites, the remaining (38) as positive testing ones containing 104 sites. Lysine sites largely outnumbered lysine propionylation sites, so positive and negative samples were unbalanced, i.e., negative samples extremely outnumbered positive ones. The unbalance between positive and negative samples would cause the trained model to prefer to negative samples. Therefore, we randomly selected sites of lysine which does not undergo PTM from these proteins as negative samples at a ratio of positive to negative 1:1. The training set consisted of 304 positive and 304 negative lysine sites, while the testing set of 104 positive and 104 negative lysine sites. All the positive and the negative sites are listed in the Supplementary Material.

We also downloaded 3,429 malonylated proteins containing 9,584 malonylation sites. Similarly, we randomly chose the same number of lysine sites as nonmalonylation sites, These lysine sites did not undergo malonylation events as negative samples. Therefore, the malonylation set contained 9,584 malonylation sites and 9,584 nonmalonylation lysine sites.

As shown in Figure 1, the proposed method consisted of three main steps: feature encoding, training classifier, and predicting propionylation, or eight modules: segmenting sequences, constructing a deep RNN model, training the deep RNN model, extracting features, constructing SVM model, optimizing the window size and the super-parameters in the SVM model, training the SVM model, and predicting propionylation with trained SVM model. We used the malonylation dataset to train the RNN model and then fine tuned the trained model by the training set of propionylation data. Propionylation sequences were inputted into the fine-tuned and trained deep RNN model and the outputs in its last second layer were viewed as features of the propionylation sequences. The subsequent workflow was the same as the common machine learning method.

In the case of regression or classification question, there are generally four types of validations: hold-out validation, k-fold crossvalidation, leave-one-out, and independent test. In the hold-out validation, the training set was splitted into two parts: one for training and another for validation. In the k-fold cross validation, the training set was divided into k parts. Each part was tested by the model trained over other k − 1 parts. Leave-one-out was an extreme cross validation, where the number of samples is equal to k. We used 10-fold cross validation and independent test to examine the proposed method.

To quantitatively compare performance of methods, the following metrics: sensitivity (SN), specificity (SP), accuracy (ACC), and Matthews correlation coefficient (MCC), were used, which were computed by

In the equations above, TP is the number of the true positive samples, TN the number of the true negative samples, FN the number of false-negative samples, and FP the number of false-positive samples. SN, SP, and ACC ranges from 0 to 1, 0 meaning completely wrong and 1 completely correct. For example, SN = 0 implied that all the positive samples were predicted as negative ones. MCC ranges from −1 to 1, 1 meaning perfect prediction, 0 random prediction, and −1 the prediction completely opposite to the true.

The receiver operating characteristic (ROC) curve was used to depict performance, which plotted true positive rate against false positive rate under various thresholds. The area under the ROC curve (AUC) was used to quantitively assess the performance. The AUC ranged from 0 to 1, 0.5 meaning random guess and 1 perfect performance.

We presented a transfer learning-based method and an online webserver¹ for computationally predicting propionylation. The method took advantage of crosstalk between propionylation and malonylation. The advantage of the method was to avoid artificially designing features. Statistical enrichment analysis implied that propoinylation was associated with metabolism.

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding authors.

AL and YD: conceptualization, funding acquisition, and writing – original draft. YT: data curation, formal analysis, and software. AL, MC, and YD: methodology, validation, writing – review, and editing. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.