ADAMS: Align Distance Matrix with SIFT algorithm enables GPU-Accelerated protein structre comparison

Requirements

opencv == 4.7.0.72

numpy >= 1.17.2

cuda > 11.x

cupy-cuda111 == 12.2.0 or same as cuda version

biopython == 1.81

scipy == 1.11.2

tqdm == 4.66.1

cuda == 11.x or same as cupy version

pickle

Installation

pip install adams

Please contact: guozy23@mails.tsinghua.edu.cn for more information

Tutorial and description

Introduction

We've developed a method to address the issue of numerous proteins exhibiting high structural similarity despite having no sequence similarities. This problem has become increasingly critical as Alphafold2 continues to predict new structures, resulting in a massive database (23TiB ver 4) that lacks an effective data mining tool.

Foldseek offers a solution by embedding local structure into the sequence and transforming this issue into a sequence alignment problem. It's significantly faster than DALI, TM-Align, and CE-Align and outperforms them on structure comparison benchmarks.

However, according to the Foldseek paper, we observed that Foldseek occasionally underperforms compared to DALI, indicating that some 'overall information' may not be captured within local structure embedding.

Our Align Distance Matrix with SIFT algorithm (ADAMS) is similar to DALI but uses an enhanced version of the renowned computer vision algorithm - Scale Invariant Feature Transform (SIFT). It extracts key features from protein distance matrices at different scales and compares their similarities. Most calculations can benefit from GPU acceleration. This zero-shot model enables more precise structure comparisons at speeds comparable to Foldseek-TM tools. Users can create their own pdb databases on PCs for all-vs-all comparisons with increased speed and reduced memory usage (approximately 500MB - 3GB GPU memory for a 20000 all vs all comparison).

The algorithm is illustrated in Fig.1: The original SIFT algorithm is applied on distance matrixes to extract detectable features across various scales. These features are represented as 128-dimension vectors which are then stacked into an n X 128 matrix for comparison between two structures using cosine similarity calculated between two feature matrices by A X B.T operation. Given these features have nearly identical lengths (512 ± 1.5), feature distances are determined by angles rather than length differences between them; thus when normalized beforehand, similarity calculation becomes straightforward on GPUs.

The performance metrics are as follows - it took between 3-4 seconds to search for the protein structure 'OSM-3' (699aa) within a C.elegans protein structure database (19361 structures) using an Nvidia RTX2080Ti (11GiB) GPU. When loading the entire database onto the dataset, total GPU memory usage was around 4000MB. However, when loaded separately, it only consumed about 500MB of memory. Importantly, these different methods did not impact search speed.

pre-print paper is here: https://www.biorxiv.org/content/10.1101/2023.11.14.566990v1.article-metrics

Tutorial

Installation

pip install adams

1. Download a pdb set and make it a cuda_database or a compatible one'

import adams
from adams.db_maker import *
db = DatabaseMaker(device=0, process=40) # use GPU-0,40*1.5 process.
db.make('./pdb','./pdb_db') # put your pdb dataset in one folder and make your database in another one

2. Match your protein structure to different databases

import adams
from adams.matcher import ADAMS_match
matcher = ADAMS_match('./protein.pdb',gpu_usage=[0,1],threshold=0.95)#use gpu0 and gpu1
result = matcher.match('./pdb_db','tmp',prefilter_threshold = 0.01) # search similar protein structure from a database, return a pandas dataframe. A temp folder is needed, will be created if not exist.

Firstly check the compare_all.py script:

compare_all.py

if permission denied

chmod +x path/to/compare_all.py