osFP: a web server for predicting the oligomeric states of fluorescent proteins

Predicting FP oligomeric states

Figure 1 illustrates the flowchart of the workflow used to predict and analyze the oligomerization of FPs. In the study, six classes of protein features were benchmark to provide a better picture on which protein features can be considered sufficient to provide insights on the oligomerization of FPs. To avoid the possibility of obtaining prediction results that may arise from chance correlation from a single calculation, the multivariate analysis was performed for 100 independent iterations where each run involves random data splitting to an internal and external sets consisting of 80 and 20%, respectively.

Judging from the performances, although different descriptor sets may capture different aspects of amino acids however all descriptor sets afforded similar level of performance. This may indicate that all descriptor sets capture the oligomerization space well and can be used as a features to train predictive QSPR models as assessed via Ac, Sn, Sp and MCC. However, the best performing method was AAC/DPC/TPC descriptors whereas AC descriptors ranked last, indicating that the amino acid composition descriptors are capable of capturing information on the oligomerization of FP. The predictive performance of the six classes of protein descriptors were further discussed in the following paragraphs.

The internal set was used to construct a predictive model by means of the J48 algorithm as to discriminate FPs to either monomers or oligomers. The predictive model was fine tuned using tenfold CV as to prevent overtraining on the internal set and then tested on an external set in order to assess its ability to accurately predict unknown samples. Table 2 provides the mean performance comparison amongst the various types of protein descriptors as assessed by the training set, the tenfold CV set and the external set.

It was observed that models built with AAC/DPC/TPC outperformed the others with Ac, Sn, Sp and MCC of 83.01 ± 2.04%, 83.26 ± 2.19%, 82.95 ± 2.27% and 0.66 ± 0.04, respectively, for the tenfold CV set. Furthermore, AAC/DPC/TPC also afforded the best performance on the external set with Ac, Sn, Sp and MCC of 83.26 ± 3.58%, 83.77 ± 5.14%, 83.37 ± 4.45% and 0.67 ± 0.07, respectively. On the other hand, the autocorrelation descriptors afforded the lowest performance with Ac, Sn, Sp and MCC of 78.48 ± 4.76%, 78.65 ± 5.66%, 78.89 ± 5.62% and 0.57±0.10, respectively. The predictive models built using CTD, Ctriad, QSO and PseAAC provided moderate performance with MCC of 0.60 ± 0.10, 0.62 ± 0.10, 0.63 ± 0.08 and 0.63 ± 0.09, respectively.

When looking into the relationship between protein sequence features and oligomerization, the phylogenetic relationships between sequences in the data sets should be taken into account. By not considering homologous relatedness amongst the FP samples, a problem in which FPs are the products of site-directed mutagenesis from a few wild-type sequences may arise. On the other hand, one site mutation may convert oligomeric FP to the monomeric state. For instance, the Ala206Lys mutation could convert a weakly oligomeric GFP to the monomeric form. Therefore, homologous reduction with the threshold of 100, 99 and 95% were considered.

Table 2 shows the results of the predictive models for the data set that removes all identical homologous sequence via the use of an identity threshold of 100% (non-redundant data set), it can be seen that the top performing tenfold CV set was built using amino acid composition, which afforded the highest performance with Ac, Sn, Sp and MCC of 83.07 ± 2.04%, 83.26 ± 2.19%, 82.95 ± 2.27% and 0.66 ± 0.04, respectively, whereas the performance of tenfold CV of autocorrelation descriptors (i.e. normalized Moreau-Broto autocorrelation, Moran autocorrelation and Geary autocorrelation) was 78.49 ± 4.76%, 78.36 ± 2.36%, 78.67 ± 2.30% and 0.57 ± 0.04 for Ac, Sn, Sp and MCC, respectively. However, the J48 model built with CTD, Conjoint, QSO and PseAAC were comparable with Ac, Sn, Sp and MCC in the ranges of 80.15–81.13, 79.66–81.19, 79.82–81.14 and 0.60–0.62, respectively.

The performance of models with homologous sequence reduction set at 99% is shown in Table 3. Again, J48 model built using AAC/DPC/TPC descriptors was the top performing model as assessed via tenfold CV with Ac, Sn, Sp and MCC with 79.40 ± 2.75%, 80.78 ± 1.86%, 77.98 ± 3.20% and 0.59 ± 0.06, respectively, when compared to other J48 models built with different descriptor sets. As for the external set, the J48 model with the lowest performance was made with AC descriptors having Ac, Sn, Sp and MCC of 72.73 ± 6.11%, 74.89 ± 6.72%, 71.43 ± 7.57% and 0.46 ± 0.12, respectively. Nevertheless, it can be observed that J48 models built with different descriptor set performance well as assessed via tenfold CV set and external set.

For the performance of the sequence homologous reduction at 95%, it can be seen that the top performing model of the tenfold CV set resulted in Ac, Sn, Sp and MCC of 72.13 ± 4.18%, 79.83 ± 3.66%, 61.03 ± 5.34% and 0.42 ± 0.09, respectively, was from the model built using amino acid composition. On the other hand, the other models built using different descriptors were comparable as shown in Table 4. As for the external set, again, model built with amino acid composition outperform others with Ac, Sn, Sp and MCC of 72.89 ± 7.08%, 79.85 ± 6.92%, 64.16 ± 11.20% and 0.43 ± 0.15, respectively.