Detection and identification of macromolecular complexes in cryo-electron tomograms using support vector machines

Detection and identification of macromolecular complexes in cryo-electron tomograms is challenging due to the extremely low signal-to-noise ratio (SNR). While the state-of-the-art method is template matching with a single template, we propose a 3-step supervised learning approach: (i) pre-detection of candidates, (ii) feature calculation, and (iii) final decision using a support vector machine (SVM). We use two types of features for SVM: (i) correlation coefficients from multiple templates, and (ii) rotation invariant features derived from spherical harmonics. Experiments conducted on both simulated and experimental tomograms show that our approach outperforms the state-of-the-art method.