Nature Machine Intelligence: Land teams work together to develop antibody research models based on artificial intelligence technology

In the game between humans and viral pathogens, the discovery of powerful neutralizing antibodies (NAB) in treatment is one of the
important "weapons".
In natural antibodies or artificially designed antibodies, the research process of mechanism of action and neutralization ability often requires a large number of experiments to detect and explore, and it is also a key speed limiting link
in the "human and virus" race.
How to quickly and accurately predict the neutralization ability of unknown antibodies and their targets is a key scientific issue
that needs further breakthroughs in the field of traditional antibody drug development.

On November 7, 2022, the land team of the School of Basic Medicine of Fudan University, together with SenseTime, Zhang Shaoting, Zhang Jie and others, published a title: Predicting unseen antibodies' neutralizability via adaptive graph Research paper
on neural networks.

This study proposed for the first time a deep Ab-Ag interaction algorithm (DeepAAI).

DeepAAI is different from the classical sequence alignment method, but through the method of deep learning "dynamically and adaptively" learning the relationship between unknown antibodies and known antibodies (Adaptive Relation Graph), thereby avoiding the problem of cold start of unknown antibodies by AI algorithms and effectively predicting the neutralization ability
of unknown antibodies.
In addition, DeepAAI also has good explanatory properties and can provide clues for the binding sites of antigens and antibodies.
Analyze the similarity between different variants and subvariants of the same virus and recommend possible neutralizing antibodies
for new subspecies of a virus.

Another feature of DeepAAI is that it is based on sequence data
.
While real tertiary structure data can improve the prediction accuracy of AI algorithms, the tertiary structure of a large number of antibodies in the real-world is unknown
.
DeepAAI abandons the tandem AI algorithm mode of first predicting the structure according to the sequence, and then predicting the antigen-antibody interaction according to the predicted structure, and directly extracts enough effective features based on the sequence to predict the interaction
.
This avoids the risk that
"errors in step 1 are accumulated and exponentially magnified in step 2".
At the same time, the large amount of sequence data that exists in the real world can also enhance the practicality
of AI algorithms.

To fully assess the predictive power of DeepAAI, the study conducted in-depth research using HIV, SARS-CoV-2
, infuenza, and dengue as model viruses.
DeepAAI has demonstrated a degree of precision in
predicting the neutralization ability of antibodies to these viruses.

Considering the emergence of a large number of SARS-CoV-2 variants, especially the Omicron subtype, the DeepAAI model reported by the institute may provide ideas
for antibody drug optimization and the development of broad-spectrum antiviral antibodies.

Original source:

Zhang， J.
， Du， Y.
， Zhou， P.
et al.
Predicting unseen antibodies’ neutralizability via adaptive graph neural networks.
Nat Mach Intell (2022).
https：//doi.
org/10.
1038/s42256-022-00553-w.