Table 3 Accuracy of fold class assignment on SCOPe.

Method	Accuracy
Method	Overall	\(\alpha\)	\(\beta\)	\(\alpha \beta\)	Small proteins
Expert handmade (without optimization)	0.852	0.683	0.771	0.961	0.357
Expert handmade (optimized)	0.880	0.759	0.889	0.928	0.500
Multinomial logistic regression	0.863	0.916	0.861	0.851	0.818
SVM (linear)	0.445	0.991	0.927	0.069	0.548
SVM (RBF kernel)	0.896	0.947	0.869	0.896	0.861
Bagged SVM (RBF kernel)	0.915	0.943	0.882	0.937	0.621

Fold classes were assigned to AlphaFold2 models based on secondary structure content and sequence length. Here we show the benchmark results from optimizing these classifiers on the original manually curated SCOPe fold classes. For the expert handmade classifiers, secondary structure content and protein length conditions were defined for each fold class. The first classifier without optimization used the following conditions: \({{{{{{\rm{length}}}}}}} < 50{aa}\to {{{{{{\rm{small}}}}}}}\); else \({{{{{{\rm{helix}}}}}}}\ge 60 \% \to \alpha\); else \({{{{{{\rm{sheet}}}}}}}\ge 35 \%\) and \({{{{{{\rm{helix}}}}}}} < 20 \% \to \beta\); else \(\to \alpha \beta\). The second one optimized the actual threshold values by parameter sweep of an increment of 5% for secondary structure content and increments of 5aa for the sequence length. The optimized mapping was: \({{{{{{\rm{length}}}}}}} < 55{aa}\to {{{{{{\rm{small}}}}}}}\); else \({{{{{{\rm{helix}}}}}}}\ge 55 \% \to \alpha\); else \({{{{{{\rm{sheet}}}}}}}\ge 25 \%\) and \({{{{{{\rm{helix}}}}}}} < 20 \% \to \beta\); else \(\to \alpha \beta\). For the other classifiers, lengths and secondary structure proportions were used directly as features. For each classifier, accuracy is shown both overall and per-class.

Quick links

Search