Search

Abstract Background Protein biomarkers play important roles in cancer diagnosis. Many efforts have been made on measuring abnormal expression intensity in biological samples to identity cancer types and stages. However, the change of subcellular location of proteins, which is also critical for understanding and detecting diseases, has been rarely studied. Results In this work, we developed a machine learning model to classify protein subcellular locations based on immunohistochemistry images of human colon tissues, and validated the ability of the model to detect subcellular location changes of biomarker proteins related to colon cancer. The model uses representative image patches as inputs, and integrates feature engineering and deep learning methods. It achieves 92.69% accuracy in classification of new proteins. Two validation datasets of colon cancer biomarkers derived from published literatures and the human protein atlas database respectively are employed. It turns out that 81.82 and 65.66% of the biomarker proteins can be identified to change locations. Conclusions Our results demonstrate that using image patches and combining predefined and deep features can improve the performance of protein subcellular localization, and our model can effectively detect biomarkers based on protein subcellular translocations. This study is anticipated to be useful in annotating unknown subcellular localization for proteins and discovering new potential location biomarkers.

Given the limitation of technologies, the subcellular localizations of proteins are difficult to identify. Predicting the subcellular localization and the intercellular distribution patterns of proteins in accordance with their specific biological roles, including validated functions, relationships with other proteins, and even their specific sequence characteristics, is necessary. The computational prediction of protein subcellular localizations can be performed on the basis of the sequence and the functional characteristics. In this study, the protein–protein interaction network, functional annotation of proteins and a group of direct proteins with known subcellular localization were used to construct models. To build efficient models, several powerful machine learning algorithms, including two feature selection methods, four classification algorithms, were employed. Some key proteins and functional terms were discovered, which may provide important contributions for determining protein subcellular locations. Furthermore, some quantitative rules were established to identify the potential subcellular localizations of proteins. As the first prediction model that uses direct protein annotation information (i.e., functional features) and STRING-based protein–protein interaction network (i.e., network features), our computational model can help promote the development of predictive technologies on subcellular localizations and provide a new approach for exploring the protein subcellular localization patterns and their potential biological importance.

High throughput screening determines the effects of many conditions on a given biological target. Currently, to estimate the effects of those conditions on other targets requires either strong modeling assumptions (e.g. similarities among targets) or separate screens. Ideally, data-driven experimentation could be used to learn accurate models for many conditions and targets without doing all possible experiments. We have previously described an active machine learning algorithm that can iteratively choose small sets of experiments to learn models of multiple effects. We now show that, with no prior knowledge and with liquid handling robotics and automated microscopy under its control, this learner accurately learned the effects of 48 chemical compounds on the subcellular localization of 48 proteins while performing only 29% of all possible experiments. The results represent the first practical demonstration of the utility of active learning-driven biological experimentation in which the set of possible phenotypes is unknown in advance.