DLFF-ACP: prediction of ACPs based on deep learning and multi-view features fusion

Datasets

In this study, we selected datasets from ACPred-Fuse (Rao et al., 2020b) to develop and evaluate our model. The positive samples were collected from the studies of Chen et al. (2016) and Tyagi et al. (2013), as well as the ACPs database, CancerPPD (Tyagi et al., 2015). The negative samples were partly collected from Tyagi et al.’s (2013), and other negative sequences with no anticancer activity were selected from Swiss-Prot (Bairoch & Apweiler, 2000). To avoid classification bias, we used the CD-HIT program (Li & Godzik, 2006) to remove peptide sequences with a similarity greater than 0.8 in both the positive and negative datasets. Finally, the data set consisted of 332 positive samples and 2,878 negative samples. We randomly selected 250 positive samples and 250 negative samples for the training set. The remaining 82 positive samples and 2,628 negative samples were used as the test set. Because of the imbalance of positive and negative samples in the above test set, we also added a balanced test set (82 ACPs and 82 non-ACPs) from ACPred-FL (Wei et al., 2018) for performance comparison with other models. The datasets used in this study were most representative in the field of ACP identification, which was convenient for the comparison and analysis with other methods. The details of these datasets are shown in Table 1.

Encoding

In order to input the peptide sequences into the deep learning model, we needed to transform the format of the peptide sequences into numerical vector. We assigned a different number 1 to 20 to each of the 20 amino acids. Since the length of the peptide sequences input into the model should be fixed, we amplified the length of each peptide sequence to 210 by padding zero to fit our dataset’s longest ACP (207 amino acids) and non-ACP (96 amino acids). By tuning weights, the model quickly learned to ignore these padded zeros. The encoding process can be seen in Fig. 1.

Handcrafted features

Different features can represent different information from the amino acid sequence. Here, we used three features that are commonly-used in ACP prediction: amino acid composition (AAC) (Bhasin & Raghava, 2004), dipeptide composition (DPC) (Saravanan & Gautham, 2015), and the composition of k-spaced amino acid group pairs (CKSAAGP) (Chen et al., 2018). In addition to these three feature representation methods, we also tested binary, physical and chemical properties (grouped amino acid composition and grouped di-peptide composition) and autocorrelation methods (Moran, Geary). However, they performed poorly in the model with remove the CNN channel, so we did not consider them in the final fused feature model. The feature representation of the peptide sequences was obtained through iFeature (Chen et al., 2018). The details of each feature are discussed in the following sections.

AAC

AAC can be used to represent the frequency of each amino acid in the sequence. It is calculated using the following equation: ${\displaystyle f\left(a\right)=\frac{N\left(a\right)}{L},a\in \left\{A,C,D,\dots ,Y\right\}}$

where f(a) represents the frequency of the occurrence of amino acid type a, N(a) represents the total number of amino acids a appearing in the peptide sequences, and L represents the length of the peptide sequences.

DPC

DPC is defined as the number of possible dipeptide combinations in a given peptide sequence. It can be calculated as: ${\displaystyle D\left(r,s\right)=\frac{N_{rs}}{L-1},r,s\in \left\{A,C,D,\dots ,Y\right\}}$

where N_rs is the number of dipeptides represented by amino acid types r and s.

CKSAAGP

According to their physicochemical properties, the 20 amino acids can be divided into five classes (Lee et al., 2011). Table 2 lists the different physicochemical properties of these five classes and the amino acids contained in each class.

The CKSAAGP is used to calculate the frequency of amino acid group pairs separated by any k residues. Using k = 0 as an example, there are 25 0-spaced group pairs (i.e., G1G1, G1G2, G1G3, …, G5G5). The features can be defined as: ${\displaystyle {\left(\frac{N_{G1G1}}{N},\frac{N_{G1G2}}{N},\frac{N_{G1G3}}{N},\dots ,\frac{N_{G5G5}}{N}\right)}_{25}.}$

Each value represents the number of times the residual group pair appears in the peptide sequence. For a peptide sequence with length L, when k = 0,1,2,3,4,5, the corresponding values of N are L-1, L-2, L-3, L-4, L-5, and L-6, respectively.

Model architecture

In order to improve the recognition ACPs, we designed a new model called DLFF-ACP based on dual-channel DNN. We used Keras framework (Chollet, 2018) to build the model and TensorFlow (Abadi et al., 2016) deep learning library back-end. The DLFF-ACP model structure is shown in Fig. 2. It consists of three parts: the handcrafted feature channel (M1) for processing handcrafted feature information CNN channel (M2) for extracting features using CNN, and the classification module (M3) for classify the ACP and non-ACP based on the fused features. The sequence was converted into a numerical vector with a length of 210, processed by the embedding layer, and lastly, fed into CNN to extract features. Handcrafted features were fed into a network containing two full connection layers and then connected with the features extracted by CNN. Finally, this fused feature was used as the input of the M3 and output the probability value in the range [0, 1].

Performance evaluation

To evaluate the performance, we adopted four metrics that are widely-used in machine learning for two-class prediction problems: sensitivity (SE), specificity (SP), accuracy (ACC) and Matthew’s correlation coefficient (MCC) (Chu et al., 2019; Le et al., 2020; Yue, Chu & Xia, 2020). The four metrics are defined as follows:

${\displaystyle SE=\frac{TP}{TP+FN}}$ ${\displaystyle SP=\frac{TN}{TN+FP}}$ ${\displaystyle ACC=\frac{TP+TN}{TP+TN+FP+FN}}$ ${\displaystyle MCC=\frac{TP\times TN-FP\times FN}{\sqrt{\left(TP+FN\right)\times \left(TP+FP\right)\times \left(TN+FN\right)\times \left(TN+FP\right)}}}$

where TP, TN, FP, and FN represent the number of true positives (i.e., ACPs classified correctly as ACPs), true negatives (i.e., non-ACPs classified correctly as non-ACPs), false positives (i.e., ACPs classified incorrectly as non-ACPs), and false negatives (i.e., non-ACPs classified incorrectly as ACPs), respectively. In order to better measure the classifier’s overall performance, we also used area under curve (AUC) (Shi et al., 2019) as the threshold-independent evaluation metric. AUC is defined as the area bounded by the receiver operating characteristic (ROC) curve and coordinate axis. The closer the AUC is to 1.0, the higher the authenticity of the detection method (Bin et al., 2020; McDermott et al., 2019).

Compositional analysis

In order to better understand the differences between ACPs and non-ACPs, we conducted different types of analyses on the total training and test datasets. The AAC analysis results Fig. 3A show that ACPs had a higher tendency to contain A, F, K, L, and W compared to non-ACPs, while D, E, N, Q, and S were more abundant in non-ACPs. This suggested that ACPs contain more positively-charged amino acids than non-ACPs. The selective toxicity of ACPs is also thought to be due to the close electrostatic interaction between positively-charged ACPs and negatively-charged cancer cells. The DPC analysis results Fig. 3B show that dipeptides AK, AL, FA, KA, KK, KL, KW, LA, LK, and LL were significantly more dominant in ACPs than in non-ACPs, and dipeptides DL, DV, EA, EE, EG, EL, EV, LD, LE, NL, TL, and VE were more abundant in non-ACPs than in ACPs. Additionally, the analysis results showed that 43% of dipeptides differed between ACPs and non-ACPS (p < 0.01). The CKSAAGP analysis results are shown in Fig. 3C, where X represents the interval between amino acids. As seen, the ratios of G1G1, G1G3, G1XG3 G1XXXG1, G1XXXXG1, G1XXXXG3, G1XXXXG4, G1XXXXXG3, G2XXXXG3, G2XXXXXG3, G3G1, G3G3, G3XG1, G3XXG2, G3XXG3, G3XXXG3, G3XXXXG1, G3XXXXXG1, G1G3, G1G1, and G1XXG3 were higher in the ACPs than in non-ACPs. However, G1XXG5, G4G1, G5XXXXXG1, G5XG1, G4XG1, G5XXXG1, G5XXXXG1, G5G1, and G5XXG1 ratios in ACPs were higher than those in non-ACPs. It is worth noting that this is highly similar to our AAC and DPC analysis results. Across 150 compositions of k-spaced amino acid pairs, 121 showed differences between ACPs and non-ACPS (p < 0.01). In conclusion, we chose these features as the input features of our proposed method due to the significant differences in the analyses of these components.

Parameters of CNN

In our model, we used CNN to extract the spatial features of a sequence. In order to find the best hidden layer setting, we selected different number of filter layers. Since our training set was small 500 samples, we considered using at most two layers when selecting CNN architectures. Using deeper layers meant introducing too many parameters, which could potentially cause overfitting of the model. We chose three different filter sizes: 32,64, and 128. Table 3 shows the performance of 10-fold cross validation on the training set with different number of filter layers. The model with 32 filters achieved the best performance and the highest AUC value of 0.89. Therefore, we chose to use 32 filters when building our model.

Channel comparisons

In our model, the fusing feature information used for classification came from M1 and M2. To have a deeper understanding of the specific performance of these two channels, we compared M1 and M2. In M1, we chose three handcrafted features for comparison. Since there was only a single channel, we removed the model’s concatenate process. The output of a single channel was directly input into the M3 and the final prediction result was obtained. All the results were obtained through 10-fold cross-validation on the training set. The final comparison results are shown in Table 4 and Fig. 4. M2 achieved the highest SP (0.80), ACC (0.78), and MCC (0.57) values. Additionally, M2 still achieved the highest AUC. Therefore, the features extracted by the CNN channel achieved the best performance across the four different features.

Comparison across different fusion features

To select the most effective feature fusion method with handcrafted features, we used CNN extracted features connected with three kinds of handcrafted features: AAC+CNN, CKSAAGP+CNN, and DPC+CNN. In order to verify whether the fused features were conducive to model performance, we compared the fused feature model with the model constructed using only the CNN channel (hereinafter referred to as the CNN model) with the best performance in individual feature models. Table 5 and Fig. 5 show the model performance across different scenarios in the training set. When compared with the CNN model only, the fused feature model achieved better performance and all the three fusion feature models were superior to the CNN model in performance. The CNN+AAC group had the highest SP (0.84) and the CNN+CKSAAGP group had the highest SE, ACC, and MCC, at 0.82, 0.82 and 0.65 respectively. When compared with the CNN model, the independent threshold evaluation method’s AUC of the fusion feature model also improved. To better understand the distribution of feature information, we used t-distributed stochastic neighbor embedding (t-SNE) (Maaten & Hinton, 2008) to visualize the feature information, and the results are shown in Fig. 6. Figure 6A shows the feature of CKSAAGP, Fig. 6B shows the sequence after encoding, and Fig. 6C shows the fusion of the output information of M1 and M2. The fused information can effectively differentiate ACPs from non-ACPs, fully demonstrating the effectiveness of using fusion features.

Comparing the proposed model with existing methods

To verify the performance of our proposed model, we compared it with six traditional machine learning methods: AntiCP (Tyagi et al., 2013), Hajisharifi’s method (Hajisharifi et al., 2014), iACP (Chen et al., 2016), ACPredFL (Wei et al., 2018), PEPred-Suite (Wei et al., 2019), ACPred_Fuse (Rao et al., 2020b), and two deep learning methods, DeepACP (Yu et al., 2020) and ACP-MHCNN (Ahmed et al., 2020). We used the results of traditional machine learning method comparisons from the literature (Rao et al., 2020b). It should be noted that all models were used on the same training set, and the comparison results are shown in Table 6. Our model performed better in terms of SE, with an improvement of 0.05−0.28 compared with the other models. Our model’s MCC was similar to that of ACPred-Fuse and higher than those of the other models. However, our model’s AUC was higher than the other models. Our model was overall superior to the other models at distinguishing ACPs from non-ACPs. To further verify the performance of our model, we compared our method with other methods on the ACPred-FL’s dataset. To make a fair comparison, we used this training set to train all models, and tested the test dataset. The performance comparison results of all the models are illustrated in Table 7. Compared to AntiCP, Hajisharifi’s method, iACP, and DeepACP, our model showed an improvement of 0.04−0.13 and 0.08−0.25 in ACC and MCC, respectively. Although our model’s performance in this data set was a litter lower than that of ACPred-FL and ACP-MHCNN, it was generally better than most of the other methods.