Chung-Chih Lin

Boosting multiclass learning with repeating codes and weak detectors for protein subcellular localization

MOTIVATION Determining locations of protein expression is essential to understand protein function. Advances in green fluorescence protein (GFP) fusion proteins and automated fluorescence microscopy allow for rapid acquisition of large collections of protein localization images. Recognition of these cell images requires an automated image analysis system. Approaches taken by previous work concentrated on designing a set of optimal features and then applying standard machine-learning algorithms. In fact, trends of recent advances in machine learning and computer vision can be applied to improve the performance. One trend is the advances in multiclass learning with error-correcting output codes (ECOC). Another trend is the use of a large number of weak detectors with boosting for detecting objects in images of real-world scenes. RESULTS We take advantage of these advances to propose a new learning algorithm, AdaBoost.ERC, coupled with weak and strong detectors, to improve the performance of automatic recognition of protein subcellular locations in cell images. We prepared two image data sets of CHO and Vero cells and downloaded a HeLa cell image data set in the public domain to evaluate our new method. We show that AdaBoost.ERC outperforms other AdaBoost extensions. We demonstrate the benefit of weak detectors by showing significant performance improvements over classifiers using only strong detectors. We also empirically test our methods capability of generalizing to heterogeneous image collections. Compared with previous work, our method performs reasonably well for the HeLa cell images. AVAILABILITY CHO and Vero cell images, their corresponding feature sets (SSLF and WSLF), our new learning algorithm, AdaBoost.ERC, and Supplementary Material are available at http://aiia.iis.sinica.edu.tw/

Tracking of secretory vesicles of PC12 cells by total internal reflection fluorescence microscopy.

Total internal reflection fluorescence microscopy is used to detect cellular events near the plasma membrane. Behaviours of secretory vesicles near the cell surface of living PC12 cells, a neuroendocrine cell line, are studied. The secretory vesicles are labelled by over‐expression of enhanced green fluorescent protein‐tagged Rab3A, one of the small G proteins involved in the fusion of secretory vesicles to plasma membrane in PC12 cells. Images acquired by a fast cooled charge‐coupled device camera using conventional fluorescence microscopy and total internal reflection fluorescence microscopy are compared and analysed. Within the small evanescent range (< 200 nm), the movements of the secretory vesicles of PC12 cells before and after stimulation by high K+ are examined. The movements of one vesicle relative to another already docked on the membrane are detected. Total internal reflection fluorescence microscopy provides a novel optical method to trace and analyse the exocytotic events and vesicle specifically near a cell membrane without interference of signals from other parts of the cell.

Involvement of Rab3A in Vesicle Priming During Exocytosis: Interaction with Munc13‐1 and Munc18‐1

Rab3A is a small G‐protein of the Rab family that is involved in the late steps of exocytosis. Here, we studied the role of Rab3A and its relationship with Munc13‐1 and Munc18‐1 during vesicle priming. Phorbol 12‐myristate 13‐acetate (PMA) is known to enhance the percentage of fusion‐competent vesicles and this is mediated by protein kinase C (PKC)‐independent Munc13‐1 activation and PKC‐dependent dissociation of Munc18‐1 from syntaxin 1a. Our results show that the effects of PMA varied in cells overexpressing Rab3A or mutants of Rab3A and in cells with Rab3A knockdown. When Munc13‐1 was overexpressed in Rab3A knockdown cells, secretion was completely inhibited. In cells overexpressing a Rab‐interacting molecule (RIM)‐binding deficient Munc13‐1 mutant, 128‐Munc13‐1, the effects of Rab3A on PMA‐induced secretion was abolished. The effect of PMA, which disappeared in cells overexpressing GTP‐Rab3A (Q81L), could be reversed by co‐expressing Munc18‐1 but not its mutant R39C, which is unable to bind to syntaxin 1a. In cells overexpressing Munc18‐1, manipulation of Rab3A activity had no effect on secretion. Finally, Munc18‐1 enhanced the dissociation of Rab3A, and such enhancement correlated with exocytosis. In summary, our results support the hypothesis that the Rab3A cycle is coupled with the activation of Munc13‐1 via RIM, which accounts for the regulation of secretion by Rab3A. Munc18‐1 acts downstream of Munc13‐1/RIM/Rab3A and interacts with syntaxin 1a allowing vesicle priming. Furthermore, Munc18‐1 promotes Rab3A dissociation from vesicles, which then results in fusion.

Fusion of systems for automated cell phenotype image classification

Automated cell phenotype image classification is related to the problem of determining locations of protein expression within living cells. Localization of proteins in cells is directly related to their functions and it is crucial for several applications ranging from early diagnosis of a disease to monitoring of therapeutic effectiveness of drugs. Recent advances in imaging instruments and biological reagents have allowed fluorescence microscopy to be extensively used as a tool to understand biology at the cellular level by means of the visualization of biological activity within cells. However, human classification of fluorescence cell micrographs is still subjective and very time consuming, thus an automated approach for the systematic determination of protein subcellular locations from fluorescence microscopy images is required. Existing approaches concentrated on designing a set of optimal features and then applying standard machine-learning algorithms. This paper takes into consideration the best methods proposed in the literature and focuses on the study of ensemble machine learning techniques for cell phenotype image classification. Two techniques are tested for the classification: a random subspace of Levenberg-Marquardt neural networks and a variant of the AdaBoost. Each of these two methods are tested with different feature sets, moreover the fusion between the two ensembles is studied. The best ensemble tested in this work obtains an outstanding 97.5% accuracy in the 2D-Hela dataset, which to the best of our knowledge is the best performance obtained on this dataset (the most used benchmark for comparing automated cell phenotype image classification approaches).

A spectral graph theoretic approach to quantification and calibration of collective morphological differences in cell images

Motivation: High-throughput image-based assay technologies can rapidly produce a large number of cell images for drug screening, but data analysis is still a major bottleneck that limits their utility. Quantifying a wide variety of morphological differences observed in cell images under different drug influences is still a challenging task because the result can be highly sensitive to sampling and noise. Results: We propose a graph-based approach to cell image analysis. We define graph transition energy to quantify morphological differences between image sets. A spectral graph theoretic regularization is applied to transform the feature space based on training examples of extremely different images to calibrate the quantification. Calibration is essential for a practical quantification method because we need to measure the confidence of the quantification. We applied our method to quantify the degree of partial fragmentation of mitochondria in collections of fluorescent cell images. We show that with transformation, the quantification can be more accurate and sensitive than that without transformation. We also show that our method outperforms competing methods, including neighbourhood component analysis and the multi-variate drug profiling method by Loo et al. We illustrate its utility with a study of Annonaceous acetogenins, a family of compounds with drug potential. Our result reveals that squamocin induces more fragmented mitochondria than muricin A. Availability: Mitochondrial cell images, their corresponding feature sets (SSLF and WSLF) and the source code of our proposed method are available at http://aiia.iis.sinica.edu.tw/. Contact: chunnan@iis.sinica.edu.tw Supplementary information: Supplementary data are available at Bioinformatics online.

Morphological filter improve the efficiency of automated tracking of secretory vesicles with various dynamic properties

Membrane trafficking is a very important physiological process involved in protein transport, endocytosis, and exocytosis. The functions of vesicles are strongly correlated with various spatial dynamic properties of vesicles, including their types of movements and morphology. Several methods are used to quantify such dynamic properties, but most of them are specific to particular populations of vesicles. We previously developed the so‐called PTrack system for quantifying the dynamics of secretory vesicles near the cell surface, which are small and move slowly. To improve the system performance in quantifying large and fast‐moving vesicles, we firstly combined morphological filter with two‐threshold image processing techniques to locate granules of various sizes. Next, Kalman filtering was used to improve the performance in tracking fast‐moving and large granules. Performance evaluation by using simulation image sequences shown that the new system, called PTrack II, yields better tracking accuracy. The tracking system was validated using time‐lapse images of insulin granules in βTC3 cells, which revealed that PTrack II could track better than PTrack, averaged accuracy up to 56%. The overall tracking results indicate that PTrack II is better at tracking vesicles with various dynamic properties, which will facilitate the acquisition of more‐complete information on vesicle dynamics. Microsc. Res. Tech., 2009.

In-Depth Fluorescence Lifetime Imaging Analysis Revealing SNAP25A-Rabphilin 3A Interactions

The high sensitivity and spatial resolution enabled by two-photon excitation fluorescence lifetime imaging microscopy/fluorescence resonance energy transfer (2PE-FLIM/FRET) provide an effective approach that reveals protein-protein interactions in a single cell during stimulated exocytosis. Enhanced green fluorescence protein (EGFP)-labeled synaptosomal associated protein of 25 kDa (SNAP25A) and red fluorescence protein (mRFP)-labeled Rabphillin 3A (RPH3A) were co-expressed in PC12 cells as the FRET donor and acceptor, respectively. The FLIM images of EGFP-SNAP25A suggested that SNAP25A/RPH3A interaction was increased during exocytosis. In addition, the multidimensional (three-dimensional with time) nature of the 2PE-FLIM image datasets can also resolve the protein interactions in the z direction, and we have compared several image analysis methods to extract more accurate and detailed information from the FLIM images. Fluorescence lifetime was fitted by using one and two component analysis. The lifetime FRET efficiency was calculated by the peak lifetime (taupeak) and the left side of the half-peak width (tau1/2), respectively. The results show that FRET efficiency increased at cell surface, which suggests that SNAP25A/RPH3A interactions take place at cell surface during stimulated exocytosis. In summary, we have demonstrated that the 2PE-FLIM/FRET technique is a powerful tool to reveal dynamic SNAP25A/RPH3A interactions in single neuroendocrine cells.

Feature space transformation for semi-supervised learning for protein subcellular localization in fluorescence microscopy images

As rapid acquisition of large collections of fluorescence microscopy cell images can be automated, large-scale subcellular localizations of GFP-tagged fusion proteins can be practically accomplished. Semi-supervised learning has the potential of using a large set of unlabeled images for the recognition of subcellular organelle patterns, but the performance still has room for improvement. This paper presents a feature space transformation method based on the spectral graph theory to improve semi-supervised learning. Experimental result shows that our feature space transformation method can improve the classification accuracy substantially.

Adaptive Image Enhancement for Fluorescence Microscopy

Fluorescent cell micrographs contain inhomogeneous contrast levels due to fluorescence intensity variations and the existence of out-of-focus objects. A novel image enhancement method based on adaptive local region sizes is proposed, which can correctly highlight salient objects from changing background intensity. The local region size for each pixel is determined adaptively, then locally normalized intensity values based on the regions are obtained, automatically taking inhomogeneous contrast and background intensity into account. Object background binarization can be done by applying simple thresholding to the enhanced image. The method is validated by comparison with ground truth segmentations of cell micrographs, which shows that after applying the proposed signal enhancement, binarization using common global thresholding methods produces segmentations much closer to the ground truth. Segmentations produced by the proposed method and simple global thresholding are also compared to some recent adaptive segmentation methods, showing that the proposed method is more efficient and appropriate for cell micrograph analysis.

Dynamics of Mitochondria and Mitochondrial Ca2+ near the Plasma Membrane of PC12 Cells: A Study by Multimode Microscopy

Abstract: The goal of this study is to examine whether there is a difference in the regulation of Ca2+ between mitochondria near the cell surface and mitochondria in the cytosol. Total internal reflection fluorescence and epifluorescence microscopy were used to monitor changes in the mitochondrial Ca2+ ([Ca2+]mt) between the mitochondria near the plasma membrane and those in the cytosol. The results show that [Ca2+]mt near the plasma membrane increased earlier and decayed slower after high K+ stimulation than average mitochondria in the cytosol. In addition, the changes in [Ca2+]mt in the mitochondria near the cell surface after a second stimulation were larger than those induced by the first stimulation. The results provide direct evidence to support the hypothesis that mitochondria in different subcellular localization show differential responses to the influx of extracellular Ca2+.