Researchers stimulate antibodies against Covid-19 with AlphaFold from DeepMind

Researchers stimulate antibodies against Covid-19 with AlphaFold from DeepMind

The structural biology world was stunned at the end of 2020 by the arrival of AlphaFold 2, a second version of the deep learning neural network developed by Google’s artificial intelligence unit, DeepMind. AlphaFold solved a decades-old problem with how proteins folda key fact governing their function.

Recent research shows that the approaches pioneered by AlphaFold extend to the whole community of biologists. In an article published this month in the journal PNAS, “Deep learning-guided optimization of a human antibody against SARS-CoV-2 variants with broad neutralization“, scientists describe the modification of a known antibody against Covid-19, so as to strengthen its effectiveness against several variants of the disease.

“We are enabling to expand the scope of the antibody and improve its potency by 10 to 600-fold against SARS-CoV-2 variants, including the Delta variant”, starting from an antibody that had no no efficacy against the latter, write the scientists. They even find promising signs that the approach may work against the Omicron variant.

L’approche deep learning

The research, led by Sisi Shan and colleagues at Tsinghua University in Beijing, the University of Illinois at Urbana-Champaign, and the Massachusetts Institute of Technology (MIT), uses deep learning for two very important reasons. .

The first is to expand what is called the search space, ie the set of potential solutions for modifying antibodies. Existing approaches, like random mutagenesis, while valuable, “are time-consuming and labor-intensive.” The use of deep learning is therefore a way to automate and therefore accelerate efforts.

Second, approaches like random mutagenesis can remove the good parts of an antibody at the same time as conferring benefits. The result may be suboptimal. Using a deep learning approach, the authors hope to extend efficiency while maintaining what has been achieved. They are not just looking to improve, but also to optimize.

Fix the objective function

We speak of a network of graphs when a set of things can be evaluated according to its relationship, such as individuals on a social network. AlphaFold 2 uses protein information to build a graph of the proximity of different amino acids. These graphs are then manipulated by the attention mechanism to calculate the degree of relationship of each amino acid with another amino acid.

Sisi Shan and his colleagues take the same approach and apply it to the amino acids of the virus, the antigen, and the amino acids of the antibody. They are comparing so-called wild types with mutated forms of both, to determine how the antibody’s binding to the antigen changes when the amino acid pairs of the two change between the wild type and the mutated version.

To train a deep neural network for this task, they set a goal, what in machine learning is called the objective function, i.e. the target that the neural network is trying to reproduce. In this case, the objective function is the change in free energy, the change in energy of proteins from wild type to mutant, symbolized by the Greek letters delta-delta and G, or ΔΔG. Given a target free energy, the neural network is expanded to a point where it can reliably predict which amino acid pair changes will most closely approximate the target free energy change.

Sisi Shan and colleagues describe their approach as follows: “To estimate the effect of mutation(s), we first predict the structure of mutated protein complexes by folding side chains around mutation sites and encode the type complexes. wild (WT) and mutated using the array to obtain the WT and mutant embeddings. Then, additional layers of the neural network compare the two embeddings to predict the effect of the mutation measured by ΔΔG”.

Coding from scratch

Although Sisi Shan and his team refer to AlphaFold 2, and employ the AlphaFold 2 approach, they did not use DeepMind’s code. The new work, called ΔΔG Predictor, was written entirely from scratch, co-author Bonnie Berger of MIT tells ZDNet.

Since the ΔΔG Predictor and AlphaFold 2 are both open source, you can review them for yourself and compare their code. ΔΔG Predictor code can be viewed on the GitHub pageet AlphaFold 2 on his.

After training the neural network to predict significant antibody and antigen mutations, the authors work backwards from evidence of antibody efficacy, in the case of the Alpha, Beta, and Gamma versions of the Covid-19. They use this data to predict which mutated antibodies will prolong their effectiveness.

As the authors state, “Our method generated a library of in silico mutations of antibody CDRs, ranked by trained neutral geometric networks such that they should not only enhance antibody binding to CDR Delta, but also maintain the link to the CDR of other variables of concern (VOC)”. The “CDR”, or complementarity determining region, is the part of an antibody that binds to an antigen. The “RBD”, the receptor binding domain, is the main target of the virus.

Test with a virus synthesized in the lab

Researchers derive doubly, triply and even quadruple mutated antibodies. By testing them in the laboratory against a synthesized virus, they find that the reduction in the concentration of antigen is greater and greater as the mutations accumulate. They conclude that something “binds” better between the mutated antibody and the virus. “Antibodies HX001-020, HX001-024, HX001-033 and HX001-034 with three or four mutations were also stronger than HX001-013 with only two mutations,” they write. “The increased binding affinity may contribute to the increased neutralizing activity of these antibodies against SARS-CoV-2 WT and its variants. »

Among the provocative results, it is possible that it is enough for a mutated antibody to avoid a problematic mutation of the virus to be more effective. In a structural analysis, they found that part of the original antibody collides with a particular part of the antigen, and the two repel each other. “Because the R103 antibody and the R346 antigen both have very long side chains and carry positive charges, proximity between the two can introduce a strong repulsion that can significantly reduce the binding affinity between the antibody. and antigen,” they observe.

The scientists substitute the normal antibody particle and “we no longer observe a direct interaction with R346 on the Delta RBD. This factor could explain the markedly improved neutralization against the Delta variant”.

This research is all the more interesting since the antibody on which the authors are working is an antibody which was introduced last year by Sisi Shan and her colleagues. Called “P36-5D2”, the antibody was obtained by being isolated from the blood serum of a recovering patient infected with Covid-19. It was determined by studies in animal models by Sisi Shan and her team to be a “broad, potent and protective antibody”.

This new work therefore marks what could be a milestone in artificial intelligence: the extension of conventional wet lab methods in the treatment of infectious diseases by refining a traditional biological product using new computational methods.

Source :

Leave a Reply

Your email address will not be published. Required fields are marked *