Human-level Protein Localization with Convolutional Neural Networks

Authors: Elisabeth Rumetshofer, Markus Hofmarcher, Clemens Röhrl, Sepp Hochreiter, Günter Klambauer

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present the largest comparison of CNN architectures including Gap Net-PL for protein localization in HTI images of human cells. Gap Net-PL outperforms all other competing methods and reaches close to perfect localization in all 13 tasks with an average AUC of 98% and F1 score of 78%. On a separate test set the performance of Gap Net-PL was compared with three human experts and 25 scholars.
Researcher Affiliation Academia 1 LIT AI Lab, Johannes Kepler University Linz 2 Institute for Machine Learning, Johannes Kepler University Linz {rumetshofer,hofmarcher,hochreit,klambauer}@ml.jku.at 3Department of Medical Chemistry, Center for Pathobiochemistry and Genetics, Medical University of Vienna clemens.roehrl@meduniwien.ac.at
Pseudocode No The paper presents architectural diagrams (e.g., Figure 3) but does not include any pseudocode or algorithm blocks describing the methodology.
Open Source Code Yes Code and dataset are available at: https://github.com/ml-jku/gapnet-pl
Open Datasets Yes All experiments were conducted on a dataset released for the Cyto Challenge 2017 by the Congress of the International Society for Advancement of Cytometry (ISAC). The main challenge dataset contains 20,000 samples taken from the Cell Atlas (Thul et al., 2017) which is part of the Human Protein Atlas.
Dataset Splits Yes The pre-processed samples were then randomly split into a training (80%), a validation (10%) and a test (10%) set.
Hardware Specification Yes The batch size for the models depends on their memory consumption and is chosen as large as possible to fit on a typical GPU with 11GB memory (NVIDIA GTX 1080 Ti).
Software Dependencies No The paper mentions activation functions (SELU) and optimizers (SGD) but does not provide specific software library names with version numbers (e.g., PyTorch 1.x, TensorFlow 2.x, scikit-learn 0.x.x).
Experiment Setup Yes For our experiments we use an initial learning rate of 0.01, dropout in the fully connected layers of 30% and a batch size of 40.