A good overlap is seen amongst the 24 positives missed out by any one of the best three classifiers (‘misses’), 5 positive sequences (marked with red asterisk) are the ‘worst misses’ which are not predicted by either of the three.

ROC plot of SVMs based on different protein sequence features depicting relative trade-offs between true positive and false positives. The corresponding Area Under Curve (AUC) is given in brackets in the legends.

The web server predicts fungal adhesins based on the two best classifiers, namely based on PSSM profile (PSSM-a) and the hybrid classifier: ACHM. The two classifiers may be chosen together for a comparative prediction. The server accepts FASTA formatted sequences and allows user defined thresholds of prediction, ranging from −1.5 to 1.5.

(1) A pool of fungal adhesins and non-adhesins sequences was generated from sequence and bibliographic databases (GenBank, UniProt, PubMed). (2) Using CD-HIT, the redundancy of the sequences from both the sets was scaled to 50% threshold, yielding 75 adhesins (positive set) and 341 non-adhesins (negative set). (3) Seven different features of different dimensions (mentioned inside brackets) were extracted using PERL scripts for both the sets. For PSSM-b, lg means lag, i.e. distance along the sequence, for details c.f. [35] (4) LOO CV was done on each of the features and several SVM models with different C and γ generated. The models giving good accuracies and almost equal sensitivity and specificity were selected. (5) Several different combinations of 2, 3, 4 and 5 features were made and LOO CV run on these. Here also the best ones were selected. (6) If the performance of the seven best models trained on different individual features was comparable to or better than the best hybrid models, it was selected for further evaluation. Here the models PSSM-a and PSSM-b were selected. (7) If the hybrid model provided an edge over its constituent individual features or the other hybrid models (in terms of accuracy), it was selected (ACHM) for further evaluation. ACM was another best model amongst the hybrids but offered lower accuracy than ACHM, so was not considered further. (8) & (9) The best SVM models were tested on benchmark data sets. (10) The PSSM-a and ACHM models were implemented on the web server.