Chongfan Technology
News
10
2024
-
08
Nature sub-journal | Li Yu of the Chinese University of Hong Kong and other teams have jointly proposed an ultra-fast and highly sensitive homolog detection method!
Author:
In large databases, when using regular methods (such as protein sequence alignment) for protein homologs identification, the distal homologs are often missed.
On August 9, 2024, Li Yu of the Chinese University of Hong Kong, Mark Gerstein of Yale University and Sun Siqi of Fudan University jointly communicatedNature Biotechnology(IF=33)Published online entitled“Fast, sensitive detection of protein homologs using deep dense retrieval”of research papers.Based on the protein language model and dense retrieval technology, an ultra-fast and high-sensitivity homolog detection method-Dense Homolog Retrieval (DHR) is proposed.
Its dual encoder structure produces different embeddings to the same protein sequence, and it is easy to locate homologs by comparing these representations. Its alignment-free properties increase speed, and the protein language model contains rich evolutionary and structural information in the DHR embedding. Compared with the previous method, the sensitivity of DHR is increased by> 10%, and for samples that are difficult to identify using alignment-based methods, the sensitivity at the superfamily level is increased by> 56%. It is 22 times faster than traditional methods such as PSI-BLAST and DIAMOND, and 28,700 times faster than HMMER.The new remote homologs discovered exclusively by DHR help to uncover the connections between the characterized proteins and improve the understanding of protein evolution, structure and function.
Protein homology detection is a vital and fundamental component of computational biology, and is essential for almost all studies related to biological sequences, such as protein structure prediction, biomolecular function analysis, transcriptional regulation research, new enzyme discovery and phylogenetic reconstruction. In biological sequence databases, homologs represent evolutionarily related protein sequences with similar structure and function. Therefore, homolog detection has been widely used as a first step in evolutionary analysis and is beneficial for drug discovery, disease diagnosis, biomarker prediction, and protein structure prediction. Prior to the publication of AlphaFold2 (AF2), protein homology identification and threading were considered to be effective methods for predicting the three-dimensional (3D) structure of proteins.Despite the impressive power of deep learning in AF2, the importance of protein homologs and multiple sequence alignments (MSAs) in the AF2 framework has been recognized from the CASP15 (Critical Assessment of Structure Prediction) competition; hybrid MSAs models based on AF2 still rank high compared to single-sequence based approaches.
Identifying homologs in a sensitive and rapid manner remains crucial. In the past few decades, homology detection has been extensively studied, and several methods of three main types have been proposed. The most popular classes are the traditional alignment-based methods such as PSI-BLAST, MMseqs2 and HMMER. The classical BLAST (basic local alignment search tool) algorithm is fast, but it does not take into account the overall dependence and omits remote homologs with low sequence similarity.HMMER can implicitly learn complex location-specific rules that indicate evolution, but is highly dependent on the quality of the profile, which can only be obtained from a profile that is not always available.MSAsobtained in. Other areas for improvement include the inability to detect structural homology and the lack of protein domain information.
DHR Framework Overview
Recently, approaches that pre-train large datasets of protein sequences and then fine-tune specific downstream tasks related to sequencing, structure, or function have shown promise. Bepler and Berger proposed the use of a bidirectional long short-term memory (LSTM) model and a multitasking framework to encode structural information for knowledge transfer between structurally related proteins. A similar language model target is also applied in AF2, which further improves the accuracy of structure prediction. In the field of sequencing homology, Bileschi utilize the embedding of protein language models to focus on family-specific classification tasks, thereby facilitating the annotation of new proteins. Heinzinger and Hamamsy use protein language models to compute embedding similarity and fitting structure similarity measures, such as template modeling scores (TM-score).Therefore, they inherently exhibit the restriction that they can only perform optimally when comparing sequences of similar length. Although these methods are innovative in their respective fields, they also have certain limitations in application.
Here, the authors propose DHR, an efficient and accurate homolog detection framework. We use a contrastive learning strategy to train dual encoders to create fixed-dimensional embeddings. This method of embedding proteins allows us to accurately and quickly identify related sequences by considering the long-term dependence information of various protein families. By using the dot product, the correlation can be retrieved with high accuracy, which exceeds the existing method in terms of speed. Compared with the profile-based method, the sensitivity and AUC-1FP rate are improved by more than 10%, and the acceleration is increased by 28,700 times. Specifically, the MSA, built purely from DHR, enables the AF2/ColabFold to match the performance of its default settings. In addition, integrating DHR with other MSA pipelines can enhance performance. Compared with the standard AF2/ ColabFold pipeline, the performance of CASP14 has increased by 0.4 Å.DHR provides solutions to important computational challenges in computational biology and bioinformatics. It has the potential to unlock a wide range of applications in protein analysis, such as homology detection and sequence alignment.
Chongqing Chongfan Technology Co., Ltd. Optical platform film thickness meter OCT optical coherence tomography LabVIEW control
LATEST NEWS
Thank you for visiting the official website of Chongfan Technology. If you have cooperation intentions or suggestions, please contact us through the following methods, and we will reply as soon as possible, thank you!
Address: Room 403, Building 6, Phase III of R&D, No. 36 Xiyong Avenue, High tech Zone, Chongqing, China.
Telephone: +86-13658337211
E-mail: Sales@cfkeji.net
Website: www.cfkeji.net
Mobile Version