Extracting Cellular Location of Human Proteins Using Deep Learning

In 2007, Chebria’s team used Machine Learning to recognize and classify the major subcellular locations using 2D HeLa single-cell images dataset. They extracted multiresolution (MR) features and trained a basic two-layer Neuron Network on the dataset. Their accuracy on the dataset with a set of multiresolution features was 95.3% (Chebira et al., 2007). However, the model they trained on the 2D HeLa single-cell images dataset does not generalize to other cell types with more than one cell in an image.

In 2012, Fan Yang, Ying-Ying Xu, and Hong-Bin Shen published another gradient based classifier to classify 7 major subcellular classes using the Human Protein Atlas (HPA) database. They selected important feature subsets such as wavelet Haralick features and local binary patterns and ensembled the models. After all, the model achieved 84% accuracy on the test set and 98% accuracy on the most confident classifications (Yang, Xu, & Shen, 2012). This research successfully classified and located different human proteins in reproductive tissues. But more researches are required for obtaining meaningful information from the location of the protein in other tissues.

In 2017, experimenters in the University of Tartu also used an 11-layer neural network to classify the subcellular locations of yeast proteins. They classified two channel images (denoting protein of interest and subcellular location) into 12 subcellular zones using basic Convolution Neuron Network (CNN) architecture as the feature extractor. The model achieved 91% per cell classification accuracy, and 99% per protein accuracy (T & L, 2017). Although this approach established a baseline locating the yeast proteins, the structure of yeast cells is relatively simple compared to human cells of different types, making it unsuitable to generalize the distributional patterns of human proteins.

Inspired by the articles, our objective is to go a step further to generalize the solution to more proteins in more tissues with more subcellular categories without the limitation to specific tissues, cells or proteins. The method we present would localize proteins in 27 different human cell types with 28 different subcellular locations, far greater than the historical approaches above. Subcellular protein distribution may reflect the current status of a cell, making the classification of cells with abnormal cell growth easier. For example, exportin-1 (XPO1), or chromosomal region maintenance 1 (CRM1), transport other proteins between the cytoplasmic area and the nucleus to maintain the normal functions of a cell (Parikh, Cang, Sekhri, & Liu, 2014). When there is an imbalance of proteins distribution inside the nucleus, the chance of viral replication, inflammatory development, and malignant tumor transformation would also increase (Cast Pharma, 2015). This means that by correctly localize proteins in human cells, researches can make use of a large number of unanalyzed microscopy images to identify various types of infection within a cell.

In this experiment, we used the Human Protein Atlas (HPA) Image Classification data on Kaggle and an additional HPA v18.1 dataset for testing. HPA dataset has 1000:1 data unbalance which may result in the gradient explosion problem when using unweighted loss functions. We split the data into 10 folds cross-validation by stratifying using the label (The Human Protein Atlas, n.d.).

We augmented the training set randomly based on our augmentation algorithm as showed above in 3.

Our network consists of ideas from Aggregated Residual Block from ResNext (Kaiming, Zhuowen, Piotr, Ross, & Saining, 2016). We also added Squeeze and Excitation Block from SENet (Hu, Shen, Albanie, Sun, & Wu, 2017).

We trained the network on 16 GPU with 48 GB memory and a Nvidia Tesla P100 CPU for more than 24 hours using PyTorch. To get a faster startup, we used pre-trained SE-ResNext on ImageNet (Russakovsky et al., 2009). During the training process, we recorded several different losses. For the first stage of the training (10 epochs), we used Binary Entropy Loss (BCE) to warm up the gradient. (If we use the focal loss from the beginning, the precision would reach 99% while the recall would remain low.) After the first 10 epochs, we switch to the combination of soft F1 Loss and negatively weighted BCE. We used a batch size of 64 with an initial learning rate of 0.1 on Adadelta optimizer. Instead of decreasing learning rate each epoch, it is more reasonable to decrease it on each step consider the size of the dataset. Based on the training F1 loss, We decrease the learning rate on the plateau by a factor of 0.5 with initial learning rate set up to 0.1. We used the rest of the evaluation data for epoch evaluation. We implemented 4 times test time augmentation (TTA) so that the result would be more stable and precise. We stopped training after seeing a horizontal fluctuation of the evaluation metrics for 5 epochs. The final training F1 Score is 0.75 with the precision of 77.28% and the recall of 75.28%.

Because we discovered the F1 Score of major classes plateaued from the threshold of 0.1 to 0.9 and the best thresholds from the last epoch contains a lot of noise, we hand-picked a maximum value, 0.268, using its score from the evaluation set.

After 40 epochs trained on 90% of HPA Kaggle dataset, the F1 score reaches its maximum value. The accuracy of our model is 63.07% with binary accuracy 96.87% (compared to human accuracy of 27.29% with binary accuracy 91.15%), while the F1 Score of our model is 0.3407. With the most confident class, the model can reach the accuracy of 79.68%. Within the subcellular locations, the Nucleoli has the highest accuracy of 79.68% among all classes due to more available data for the class. In general, the machine can reach 35% higher accuracy than a trained human. Besides the accuracy, the result shows that the model’s processing speed (78.125 images per second) is way faster than that of a human.