The thermophilic and mesophilic proteins were obtained from (Lin and Chen, 2011). The psychrophilic proteins were extracted from the UniProt web server, the search keywords including the “psychrophilic, cold-adaptive, and low-temperature.” Firstly, all queried protein sequences must be reviewed and manually annotated; secondly, entries which be a part of other proteins were excluded; finally, to avoid redundancy and homology bias, the CD-HIT program (Huang et al., 2010) was used with a cutoff of 40% sequence identity. The dataset included 2,400 protein sequences, among which the thermophilic, mesophilic, and psychrophilic proteins were 915, 793 and 692, respectively. The training and test sets were split in a 4:1 ratio, so there were 731, 574, and 554 thermophilic, mesophilic, and psychrophilic proteins in the training set, and 184, 219, and 138 in the test set. The sequences of the datasets could be downloaded from the supporting material.

Protein feature descriptors are generated from protein sequences. The feature descriptor extraction and model construction were implemented using iLearnPlus, a machine-learning platform that served as protein sequence analysis and prediction. It has been reported that the AAC and DPC descriptors can discriminate the thermophilic from mesophilic proteins effectively (Zhang and Fang, 2006b; Gromiha and Suresh, 2008; Lin and Chen, 2011; Ai et al., 2018; Guo et al., 2020; Wang et al., 2020), therefore, the two descriptors were calculated for each protein sequence.

AAC refers to the occurrence of each amino acid in the protein sequence (Wang et al., 2011; Khan et al., 2015; Guo et al., 2020; Sun et al., 2020; Charoenkwan et al., 2021), as that there are 20 kinds of naturally-occurring amino acids, that is ACDEFGHIKLMNPQRSTVWY. Therefore, each residue frequency in a sequence can be calculated by the following formula:

DPC calculates dipeptide composition and generates a 400-dimensional feature vector.(Charoenkwan et al., 2021) and it was defined as:

The CHI2 algorithm (Chen et al., 2009) was used for DPC feature selection and dimensional reduction.

Several machine learning algorithms were tested to distinguish between psychrophilic and non-psychrophilic proteins. Considering the reliable performance of SVM, RF, KNN, and Bayes in classifying thermophilic and mesophilic proteins, these algorithms were used in our study (Cortes and Vapnik, 1995; Breiman, 2001). The RF is an ensemble of decision trees. The algorithm performs better than decision trees by building and merging multiple decision trees to obtain more accurate results. For a new sample, the RF assigns the class label based on the prediction by each tree. The n_trees was set to 300. The SVM is a simple but powerful supervised machine learning algorithm used in classification and/or regression. It seeks a hyperplane to classify samples. When the sample is linearly inseparable in the low-dimensional space, the kernel function is used to map the sample to the high-dimensional space to achieve linear separability. The radial basis function (RBF) was selected in the kernel function of SVM and the optimized γ and C were 8.0 and 15.0, respectively. KNN is also one of the most basic algorithms in supervised machine learning. It assumes that similar things are near to each other, and the Euclidean distance between samples was calculated to solve the classification and regression of data. The top K value was set to 3. Naïve Bayes method is a set of supervised machine learning algorithms based on Bayes’ Theorem. It obeys the assumption that every pair of the feature are independent and every feature is equal to the value of the class variable. It states the following relationship, and is mathematically expressed as the following equation:

where A and B are events and P(B) ≠ 0. All the parameters in four machine learning algorithms were optimized by grid search.

The data set was randomly divided into training set and test set in a ratio of 4:1. The 5-fold cross-validation was also used in this study, out of which the datasets were randomly divided into 5 subsets, one of which was used to test the model, and the remaining 4 subsets were used as the training set to train the model and optimize the parameters. This process was repeated 5 times until each subset was used as the test set only once to validate the model. Four indicators were adopted to evaluate the model performance, that is sensitivity (Sn), specificity (Sp), accuracy (ACC), and Matthews correlation coefficient (MCC). The calculation formulas of these indicators were as follows:

Sensitivity = TP TP + FN

Specificity = TN TN + FP

Accuracy = TP + TN TP + TN + FP + FN

MCC = ( TP × TN ) − ( FP × FN ) ( TP + FP ) ( TP + FN ) ( TN + FP ) ( TN + FN )

Where TP, TN, FP, and FN represent the number of correctly predicted positive samples, correctly predicted negative samples, incorrectly predicted positive samples, and incorrectly predicted negative samples, respectively. For a multi-classification task, the ACC was calculated as follows:

Where TP(i), TN(i), FP(i), and FN(i) represent the number of the samples that are correctly predicted as i-th class, the number of samples that are classified correctly as not to be i-th class, the number of samples not in i-th class that is classified wrongly as belonging to i-th class, the number of samples in i-th class that are predicted incorrectly as not in i-th class, respectively. Additionally, ROC (Receiver Operating Characteristic) curves were also utilized to visualize the predictive performance of the classifiers.

The predictive performance of the machine learning model based on AAC, DPC, and the combination of the two descriptors were listed in Table 1. Among the models using with AAC descriptor, the SVM model achieved the highest prediction accuracy with 80.6%. The prediction accuracy of RF was lower than 0.4% of the SVM. And the two other models, Bayes and KNN, the accuracies were less than 80%, especially the Bayes model had the lowest prediction accuracy with 73.8%. All the trained models were public in github¹website.

DPC descriptor generates 400-dimensional vectors, and the CHI2 algorithm was used for feature dimension reduction. The results indicated that the prediction accuracy of the models based on DPC descriptor decreased compared with AAC descriptor, which declined by about 5–7%. Similar to the AAC descriptor, the model with DPC descriptor using the SVM algorithm also achieved the best accuracies.

In addition, the two descriptors were integrated to construct the classification model. Compared with the DPC descriptor, the accuracies of the AAC + DPC descriptors had been improved to varying degrees. While compared to AAC descriptors, the accuracies of SVM and KNN models were almost unchanged, RF and Bayes models even dropped by 1.2 and 2.4%, respectively. The models constructed by AAC have achieved best accuracy via four machine learning algorithms in this study. Of course, DPC is also an important feature to distinguish psychrophilic proteins from non-psychrophilic proteins, which has also achieved relatively good prediction accuracy. However, the addition of DPC to the descriptor may cause redundancy of features, which makes the accuracy decrease slightly. In a report of using AAC and DPC to distinguish thermophilic and mesophilic proteins, AAC and DPC achieved 0.9256 and 0.9157 prediction accuracy, respectively. The accuracy of AAC and DPC combination to distinguish thermophilic and mesophilic proteins also decreased, though DPC contained more parameters (Lin and Chen, 2011).

The ROC curves of four models using AAC and DPC descriptors were plotted (Figure 1), it also showed that the AAC descriptors outperformed the DPC descriptors. In a comparison of the frequencies of amino acids between thermophilic and non-thermophilic proteins, it is proposed that the AAC captures the thermostability of the protein (Sun et al., 2020). Same as thermostability, it is also demonstrated that the psychrophilicity is highly related to the AAC descriptor in this study.

In addition to the higher predictive performance of the AAC descriptor, it is easy to find in Figure 1 that the SVM model achieved the best predictive accuracy among the four models (AUC 0.8609). It has been shown in many studies that the SVM model based on AAC descriptor had achieved good predictive performance in discriminating thermophilic from mesophilic enzymes. For example, the SVM model constructed by Michael Gromiha et al. using AAC descriptors could distinguish thermophiles from mesophiles with an accuracy of 89% (Gromiha and Suresh, 2008). And employing AAC descriptors with only 16 dimensions to discriminate thermophilic and non-thermophilic proteins with 93% accuracy (Guo et al., 2020).

To verify the feasibility of ternary classification, the scatter diagram of the three types of enzymes was calculated employing the K-means clustering method, where psychrophilic proteins were labeled as 1, mesophilic proteins were labeled as 0, and thermophilic proteins were labeled as 2. As seen from Figure 2, three types of proteins had different distribution patterns on principal component 1 and principal component 2, which indicated that multi-class classification is feasible. Therefore, the ternary classification model was established, and the predictive accuracies of the models for psychrophilic (P), mesophilic (M), and thermophilic (T) proteins were listed in Table 2; Supplementary Table S1.

The results showed that the accuracies of ternary classification were slightly lower than that of binary classification. The model predictive accuracy of AAC + DPC descriptors by SVM method was 76.1%, which was 4.0% lower than that of binary classification model with the same descriptors and method. In general, the SVM model performed well in discriminating three types of enzymes. As an ensemble classifier, the RF also achieved relatively good prediction accuracy with 73.8% solely using AAC descriptor. Among the four models, the predictive accuracy of the KNN model was relatively lower than other models, the prediction accuracy based on DPC descriptor was 67.1%. Taken together, the AAC descriptor achieved the highest prediction accuracy, which indicated the capacity of the amino acid composition in distinguishing psychrophilic proteins.

The frequencies of 20 amino acids in psychrophilic, mesophilic, and thermophilic proteins were computed (Figure 3). Ala, Gly, Ser, and Thr amino acids in psychrophilic enzymes were higher than those in non-psychrophilic proteins, whereas the other amino acids Glu, Lys, and Arg were lower than, the non-psychrophilic proteins, and aliphatic amino acids Ile,Val, and Leu were slightly lower than non-psychrophilic proteins.

Many studies have demonstrated that psychrophilic proteins maintain their high catalytic activity at low temperatures mainly due to their more flexible structures (Siddiqui and Cavicchioli, 2006; Santiago et al., 2016; Åqvist et al., 2017; Arcus and Mulholland, 2020). Several factors contribute to conformational flexibility, such as reduced inter-domain and inter-subunit interactions, fewer inter-protein disulfide bonds, and reduced hydrogen bonds and electrostatic interactions. Glycine and alanine are very small amino acids, and the side chains are a hydrogen atom and a methyl group, respectively. And the comparative analysis focused on the dataset from psychrophilic and mesophilic proteins also showed that Ala and Gly residues are over-represented. Increased levels of Gly residue have been suggested to be related to psychrophilicity.

A higher percentage of serine and threonine is also found in the psychrophilic proteins. The study of Subbiah et.al on the classification rules for psychrophilic and mesophilic proteins showed that when the percentage of Ser and Thr is higher than certain values, the proteins would be classified as psychrophilic proteins (Nath et al., 2012). Meanwhile, a pairwise comparison of proteins from cold-adapted archaea revealed that there was higher content of non-charged polar residues, especially threonine (Berthelot et al., 2019; Bargiela et al., 2020). Ser and Thr are uncharged polar amino acids and prefer to reside on the surface of the psychrophilic proteins (Jahandideh et al., 2007), therefore they tend to have more interactions with water molecules around proteins (Sun et al., 2020). Structural and molecular dynamics (MD) analysis of homologous psychrophilic, mesophilic, and thermophilic counterparts of serine proteases (Tiberti and Papaleo, 2011; du et al., 2017) and serine hydroxy methyltransferases (Zhang Z. B. et al., 2021) reported that psychrophilic proteins formed more hydrogen bonds with solvent water molecules. Further analysis revealed that the content of serine in psychrophilic proteases and hydroxy methyltransferases is greater than in homologous mesophilic and thermophilic proteins. Although these studies only include several types proteins, it seems that serine and threonine involve increasement of surface hydrophilicity via forming more H-bonds with water molecules to enhance the mobility and flexibility of psychrophilic enzymes.

The charged amino acids in proteins are divided into two groups: basic amino acids which are lysine, arginine, and histidine; while acidic amino acids including glutamic acid and aspartic acid. Basic and acidic amino acids have positive and negative charges under physiological conditions and thus form higher number of salt bridges and electrostatic interactions. Therefore, more charged residues were found in the non-psychrophilic than in psychrophilic proteins to maintain the conformational stability of protein structures (Gianese et al., 2002; Tiberti and Papaleo, 2011; Wu et al., 2017). However, Figure 3 indicated that Asp amino acid favors psychrophilic proteins. It seems that Asp is unstable at high temperatures, thus the increased content of Asp contributes to the structural flexibility of the psychrophilic proteins. Another acidic amino acid, Glu, contributes to the formation of helical structures, and structure comparative analysis shows that the content of helical structures is lower in the psychrophilic proteins than in the mesophilic proteins (Metpally and Reddy, 2009), thus the reduced content of Glu maintains the thermolability of psychrophilic proteins. In contrast, the higher charged amino acids in thermophilic proteins are essential to protein stabilization at high temperatures (Zhang and Fang, 2006a,b; Gromiha and Suresh, 2008; Taylor and Vaisman, 2010; Ai et al., 2018). For example, a model only using Lys residue feature to classify thermophilic and non-thermophilic proteins reached 76.41% accuracy, a striking difference between the thermophilic and non-thermophilic proteins (Guo et al., 2020).

The content of three aliphatic acids (valine, leucine, and isoleucine) in psychrophilic is slightly lower than in the non-psychrophilic proteins. The aliphatic amino acids maintain the conformational stability of the protein structure through hydrophobic interactions. Many findings have demonstrated that psychrophilic enzymes possess reduced core hydrophobicity (Lonhienne et al., 2000; Feller and Gerday, 2003; Siddiqui and Cavicchioli, 2006; Åqvist et al., 2017; Arcus and Mulholland, 2020). Such as fewer Ile residue were found on the core of the psychrophilic citrate synthase, trypsins, and AHA (Siddiqui and Cavicchioli, 2006). In other comparative studies, fewer Leu residues were proposed to contribute to the reduced hydrophobic interaction within the protein (Zhou et al., 2008).

In conclusion, psychrophilicity is the consequence of numerous characteristics, and different families of psychrophilic enzymes may adopt one or several strategies to adapt to low temperatures, which causes no structural features that is always presented in all psychrophilic enzymes.

To identify the key amino acids, the influence of different features subset on the accuracy of the model was investigated. According to the residue differences between psychrophilic and non-psychrophilic proteins, the feature of hydrophobic (ILV), charged (KRED), aromatic (WYF), and polar uncharged (STQ) residues were explored. These residual features were removed, respectively, and the remaining residues were used to build the classification model. It is demonstrated that by removing the descriptors, the performance of all established models was decreased, especially the sensitivity values decreased significantly (Table 3). The largest degradation in performance was the models that excluded from the KRED and STQ residues. It is deduced that the charged amino acids and non-polar amino acids play a vital role in discriminating psychrophilic from non-psychrophilic proteins. However, the Acc and MCC values did not decrease significantly, because the number of psychrophilic proteins was smaller than that of non-psychrophilic proteins, thus the subtle change of TP values had little effect on Acc and MCC.