Experts comment on NBT Wang Jun/Chen Yihua group using AI to enable high-throughput mining of microbiome functional macromolecules

Comments | Wang Xiaowo (Tsinghua University), Liu Chenli (Shenzhen Advanced Research Institute) editors | People, animals and plants have a large number of complex symbiotic microbiomes, and their metabolites are closely related to host health and are in many It plays an important role in this important disease, and has gradually become a hot spot of industrial and clinical attention
.

The development of live bacteria or active small molecules produced by metabolism is the main focus of clinical and drug research and development, but at the same time, the highly complex microbiome also encodes and expresses a large number of small reading frames to produce small proteins or polypeptides.
In recent years, it has been considered that It is a huge microbiome "dark matter" library, containing a variety of microbial macromolecules with important biological functions
.

A variety of cyclic peptides and linear peptides found in various microorganisms in the early stage can directly play antibacterial, anticancer, and host immunity regulation after catalytic modification or without any modification in the later stage.
Among them, polypeptides with antibacterial functions (antibacterial peptides) Some of them work by perforating or lysing on bacterial cell walls or lipid membranes, which are not easy to cause bacteria to rapidly develop drug resistance or horizontal transfer
.

In the context of the rapid development of sequencing technology, there is already a large amount of human microbiome data that can be used for further mining and transformation, especially the functional molecules of polypeptides, which can be further optimized and applied as lead molecules of various drugs in the future.

.

However, for a long time, the mining and verification of microbial macromolecules have relied on experimental methods; for functional peptides with short sequences, high diversity and low similarity, the existing bioinformatics methods also have great limitations
.

Applying the latest and more efficient artificial intelligence deep learning research methods to realize the systematic mining of microbial functional macromolecules has important basic research and application value
.

On March 3, 2022, Wang Jun's research group and Chen Yihua's research group of the Institute of Microbiology, Chinese Academy of Sciences jointly published an article Identification of antimicrobial peptides from the human gut microbiome using deep learning in Nature Biotechnology, which integrates natural language analysis in a variety of artificial intelligence fields ( Natural Language Processing) neural network method, taking antimicrobial peptides as the first research object, after learning thousands of existing antimicrobial peptide sequences, constructed an antimicrobial peptide prediction method with an accuracy rate of more than 90%, far exceeding the previous Published model for antimicrobial peptide determination based on amino acid composition and properties
.

Figure 1: Schematic diagram of the flow of this study
.

The researchers then used a large number of published microbiome sequencing data to predict small proteins and mine antimicrobial peptides in more than 10,000 sequencing data, and combined metatranscriptomic data and metaproteomic data to further screen the actual expressed peptides.

.

Combined with the correlation with species abundance in the microbiome, more than 240 peptides were finally identified as candidate molecules with good potential
.

Among the 216 peptides successfully synthesized by chemical synthesis, 180 were finally confirmed to have antibacterial activity by experiments, reaching an excavation accuracy rate of over 83%
.

The activity (minimum inhibitory concentration MIC) of some of the peptides is close to the most active eukaryotic antimicrobial peptides reported in nature, and in further experiments, it has shown that it is resistant to a variety of clinical drug-resistant bacteria (Escherichia coli, Baumann et al.
Escherichia coli, Klebsiella pneumoniae), and no significant resistance was induced in the treatment of common Escherichia coli for up to 30 days
.

Further analysis showed that the similarity of antimicrobial peptides mined by deep learning to known antimicrobial peptides was very low (mostly <60%), the structure and mechanism were not single (not limited to wall breaking or membrane breaking), and they were finally tested in animal experiments.
Among them, 3 peptides showed good therapeutic effect against Klebsiella pneumoniae infection model in vivo, and the load of pathogenic bacteria was reduced by more than 10 times, and there was no visible toxicity or harm to the animals themselves
.

On the one hand, this study combines microbiome big data and the latest deep learning models to provide a good example of artificial intelligence (AI)-enabled macromolecular mining and transformation of the microbiome; on the other hand, it also shows that the use of computational methods directly It is highly feasible to mine the active molecules (such as proteins or RNAs) directly encoded by the genome as therapeutic molecules from genomic data or other genomic data
.

At the same time, the research also expands the scope of AI methods applied to biomedical research, and adds the applicable scenarios of natural language models on the basis of previous medical image processing and small molecule drug screening
.

Considering the accumulation of more microbial big data in the future and the theoretical high diversity of polypeptide molecules, mining more polypeptide molecules with the treatment of infection, metabolism and immune diseases has great development potential
.

Doctoral students Ma Yue, Xia Binbin, Zhang Yuwei and assistant researcher Guo Zhengyan of the Institute of Microbiology, Chinese Academy of Sciences are the co-first authors, and Wang Jun and Chen Yihua are co-corresponding authors
.

Wang Jun's research group has long been focusing on bioinformatics methods and applied research of the microbiome.
Since its establishment four years ago, it has published 20 articles in journals including Nature Biotechnology, Nature Genetics, Advanced Science, Microbiome, etc.
remaining articles
.

At this stage, there are many cutting-edge projects in the field of machine learning, artificial intelligence and microbiome.
Postdoctoral fellows who are interested in this direction are welcome to join the research group.
If you are interested, please contact junwang@im.
ac.
cn Expert Comments Wang Xiaowo (Tsinghua University) The analysis of protein structure and function has long been the focus of computational biology research
.

The recent success of AlphaFold2 has set off an upsurge in the intersection of artificial intelligence and biology
.

In addition to the well-known structural proteins, there are also a large number of small proteins or polypeptides in organisms that function as hormones and signaling molecules in animals and plants, as well as functional molecules in microorganisms
.

For these functional polypeptides with short sequences, high diversity and low similarity, there is still a lack of understanding of their functions at the structural level, and there are very few available biological information analysis methods
.

 In this study, the research group of Wang Jun and Chen Yihua proposed a deep learning method based on the sequence of the natural language processing (NLP) model for the functional analysis of peptide sequences, and effectively mined the functional information hidden in the sequence
.

This study cleverly used the large amount of existing data of the microbiome, and used the natural database containing a large number of peptides to identify antimicrobial peptides, as well as follow-up step-by-step screening, mechanism verification and animal experiments
.

The study shows that the human health-related proteins contained in the genomic and metagenomic data are not only enzymes that catalyze primary and secondary products, but also a large number of smaller polypeptide protein components that can play a more direct role
.

This work well shows that, guided by prior knowledge and designing targeted artificial intelligence methods to mine massive genomic data, it is expected to achieve a leap in the discovery of macromolecular drugs; In the context of the continuous expansion of diseases, the use of computational methods to develop more functional peptides can provide more powerful support for life and health and disease treatment
.

  Expert Comments Liu Chenli (Shenzhen Advanced Research Institute, Researcher) The cooperation between Wang Jun's research group and Chen Yihua's research group has beautifully solved an obvious but unresolved important problem
.

This is an exciting and pioneering work, which has both important scientific significance and strong application value
.

As two frontier fields in the 21st century, the cross-border integration of artificial intelligence and synthetic biology is expected to become the engine of a new round of technological revolution
.

The rational design of the underlying components of synthetic biology is still very difficult at this stage, mainly due to the extremely complex nonlinear mapping relationship between sequence and function, and the black-box model based on artificial intelligence technology can effectively handle this complex relationship
.

The work published today not only provides a new method for the effective use of the current massive omics resources, but also provides an important technical means for mining and designing new functional synthetic biological components
.

Original link: https://doi.
org/10.
1038/s41587-022-01226-0 Reprint notice [Non-original article] The copyright of this article belongs to the author of the article, and you are welcome to forward and share it.
, violators will be prosecuted
.