SoftAlign is an advanced alignment method designed to efficiently compare 3D protein structures. By leveraging structural information directly, SoftAlign provides an end-to-end alignment process, allowing for both highly accurate alignments and efficient computations. The method uses the 3D coordinates of protein pairs, transforming them into feature vectors through a retrained encoder of ProteinMPNN. This similarity matrix is then aligned using two strategies: a differentiable Smith-Waterman method and a novel softmax-based pseudo-alignment approach.
Our results demonstrate that SoftAlign is able to recapitulate TM-align results while being faster and more accurate than alternative tools like Foldseek. While not the fastest alignment method available, SoftAlign excels in precision and is well-suited for integration with other pre-filtering methods. Notably, the softmax-based alignment shows superior sensitivity for structure similarity detection compared to traditional methods.
SoftAlign also introduces a novel pseudo-alignment method based on softmax. This approach can be integrated into other models and architectures, even those not inherently focused on structural information. For a more detailed description of the method, please refer to the full paper here.
To facilitate ease of use and reproducibility, we provide three Google Colab notebooks:
-
Training Notebook: Reproduces the training process with the same train-test split as described in our paper.
SoftAlign_training.ipynb -
Inference Notebook: Allows users to experiment with the learned aligners on their own data.
COLAB_SoftAlign.ipynb -
All-vs-All Search Notebook: Performs an all-vs-all search within the SCOPE 40 dataset. Note: This notebook is still in development.
SCOPE_SEARCH_SoftAlign.ipynb
