Towards Stable Representations for Protein Interface Prediction

Authors: Ziqi Gao, Zijing Liu, Yu Li, Jia Li

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on various benchmarks demonstrate that ATProt consistently improves the performance for protein interface prediction. Moreover, our method demonstrates broad applicability, performing the best even when provided with testing structures from structure prediction models like ESMFold and Alpha Fold2.
Researcher Affiliation Collaboration Ziqi Gao1,2, Zijing Liu3*, Yu Li3, Jia Li1,2* 1Hong Kong University of Science and Technology 2Hong Kong University of Science and Technology (Guangzhou) 3 International Digital Economy Academy (IDEA)
Pseudocode No The paper presents mathematical formulations and describes the framework, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks with structured steps.
Open Source Code Yes The code is provided in https://github.com/ATProt/ATProt.
Open Datasets Yes We evaluate our method on the complexes from Docking Benchmark 5.5 (DB5.5) [45], a gold standard dataset with high-quality, and Database of Interacting Protein Structures (DIPS) [43], which collects 41,876 complexes mined from PDB [4].
Dataset Splits Yes The two datasets are randomly divided into training, validation, and testing sets with the following sizes: 203/25/25 (DB5.5) and 39,937/974/965 (DIPS).
Hardware Specification Yes The training process takes around 0.5 hours with 1 Nvidia 4090 GPUs with 24GB RAM.
Software Dependencies No The paper lists hyperparameters in Table 4 but does not specify software dependencies like programming language versions or library versions (e.g., Python 3.x, PyTorch 1.x) that are crucial for reproducibility.
Experiment Setup Yes The hyper-parameters used in this paper are listed in the following table [Table 4]. For example, Batch size 4, Learning rate 3e-4, Optimizer Adam, Dropout rate 0.2, etc.