DeepProtein: Deep Learning Library and Benchmark for Protein Sequence Learning

• Key Innovation: DeepProtein introduces a comprehensive deep learning library tailored for protein science, offering advanced tools for tasks like protein function prediction, localization, and interaction prediction. The library integrates cutting-edge neural network architectures, including CNNs, RNNs, transformers, and graph neural networks (GNNs).

• Why It’s Exciting: DeepProtein sets a new standard by benchmarking 8 neural network architectures on 7 protein learning tasks, showcasing superior performance in areas like antibody developability and CRISPR repair outcome prediction.

• User-Friendly Focus: A major strength is its accessible, one-command-line interface, designed to help domain experts apply complex models with ease, building on the drug discovery library DeepPurpose.

• Comprehensive Benchmarking: The library offers a curated set of protein datasets and evaluates methods on diverse tasks, such as protein-protein interaction prediction and epitope prediction, which are essential for advancements in biotechnology and medicine.

• Deep Learning on Proteins: Beyond traditional sequence learning methods, DeepProtein incorporates structural-based methods like GNNs and graph transformers, enhancing protein structure representation and function prediction.

• Scalability: The library has been tested on powerful hardware (NVIDIA GPUs) and is optimized for large-scale protein tasks, although it also provides guidance for managing computational complexity on smaller hardware setups.

• Broader Impact: DeepProtein contributes to reproducibility in research with extensive documentation and tutorials, aiming to democratize deep learning applications in proteomics and therapeutic science.