Making Protein Localization Features More Robust Meel Velliste Carnegie Mellon University Introduction • 2-D features work well on our own image set • Can they be put to practical use? • The features need to work on other people’s images Images from online journals Fig. 8. Expression of a GMAP-210 mutant form lacking the microtubule-binding site. After transfection, cells were fixed and double labeled for GMAP-210 (a) and the medial Golgi marker CTR433 (b). Red and green image pair is shown in c. Arrow in b indicates a nontransfected cell. Alternatively, cells were stained for GMAP-210 (d, e) and alpha-tubulin (f, g) and image pairs are shown in h and i, respectively. In d, f, and h, a transfected cell is shown. In e, g, and i, a nontransfected cell is presented for comparison. Bars, 10 µm. Online Image Classification Results True Class DNA ER Giantin GPP130 LAMP2 Mitoch. Nucleolin Actin TfR Tubulin Output of Classifier DN ER Gia GP LA Mit Nuc Act TfR Tub 0 0 70 10 0 0 0 10 10 0 0 0 0 0 67 0 0 0 33 0 0 11 28 0 11 0 44 0 0 6 0 0 0 Overall accuracy = 18% 0 0 0 0 0 71 29 Without Texture Features True Class DNA ER Giantin GPP130 LAMP2 Mitoch. Nucleolin Actin TfR Tubulin Output of Classifier DN ER Gia GP LA Mit Nuc Act TfR Tub 0 20 70 0 0 10 0 0 0 0 0 0 0 33 0 0 0 33 33 0 0 6 33 39 0 0 22 0 0 0 0 0 0 Overall accuracy = 34% 0 0 0 0 0 14 86 Other Sensitive Features Removed True Class DNA ER Giantin GPP130 LAMP2 Mitoch. Nucleolin Actin TfR Tubulin Output of Classifier DN ER Gia GP LA Mit Nuc Act TfR Tub 0 0 0 0 20 0 40 40 0 0 0 0 0 0 33 0 0 0 67 0 0 0 61 22 0 0 17 0 0 0 0 0 0 0 0 14 0 14 14 57 Overall accuracy = 45% (c.f. 68% for our images) Conclusions • System can be used to find Golgi or Tubulin images • Some robustness can be achieved by eliminating features sensitive to imaging technique • Need additional robust features to improve performance • Need a more systematic approach to selecting robust features