Protein-Nucleic Acid Complex Modeling with Frame Averaging Transformer

Authors: Tinglin Huang, Zhenqiao Song, Rex Ying, Wengong Jin

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

Reproducibility Variable Result LLM Response
Research Type Experimental To validate the advantage of our model architecture, we evaluate FAFormer on two tasks: (1) protein-nucleic acid contact prediction and (2) unsupervised aptamer virtual screening. In the first task, our model consistently surpasses state-of-the-art equivariant models with over 10% relative improvement across three protein complex datasets. For the second task, we collected five real-world protein-aptamer interaction datasets with experimental binding labels. Our results show that FAFormer, trained on the contact map prediction task, is an effective binding indicator for aptamer screening.
Researcher Affiliation Academia 1Yale University, 2Carnegie Mellon University, 3Northeastern University, Khoury College of Computer Sciences 4Broad Institute of MIT and Harvard
Pseudocode No The paper describes the model architecture and its components but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes https://github.com/Graph-and-Geometric-Learning/Frame-Averaging-Transformer
Open Datasets Yes We collect the complexes from PDB [8], NDB [7], RNASolo [2] and DIPS [77] databases.
Dataset Splits Yes Each dataset is equally split into validation and test sets.
Hardware Specification Yes The experiments are conducted on a single Linux server with The AMD EPYC 7513-32 Core Processor, 1024G RAM, and 4 Tesla A40-48GB.
Software Dependencies Yes Our method is implemented on Py Torch 1.13.1 and Python 3.9.6.
Experiment Setup Yes For all the baseline models and FAFormer, we fix the batch size as 8, the number of layers as 3, the dimension of node representation as 64, and the optimizer as Adam [42]. Binary cross-entropy loss is used for contact map identification tasks with a positive weight of 4. The gradient norm is clipped to 1.0 in each training step to ensure learning stability. We report the model s performance on the test set using the best-performing model selected based on its performance on the validation set. All the results are reported based on three different random seeds. The learning rate is tuned within {1e-3, 5e-4, 1e-4} and is set to 1e-3 by default, as it generally yields the best performance. For each model, we search the hyperparameters in the following ranges: dropout rate in [0, 0.5], the number of nearest neighbors for the GNN-based methods in {10, 20, 30}, and the number of attention heads in {1, 2, 4, 8}.