Artificial intelligence was used to aid the discovery of high-affinity antibodies, which are typically identified through directed evolution—a process that requires rounds of mutagenesis and selection.
Researchers used the pipeline (called RESP) to identify a new antibody that binds a major cancer target 17-fold tighter than an existing antibody drug. The authors say the approach could accelerate the discovery of novel antibody drugs against cancer and other diseases such as COVID-19 and rheumatoid arthritis.
This is published in Nature Communications in the paper, “The RESP AI model accelerates the identification of tight-binding antibodies.”
The process to identify a high-affinity antibody is long and expensive. In addition, many of the resulting antibodies fail to be effective in clinical trials. Deep learning techniques could accelerate the process, but the existing AI-based methods call the reliability of the predictions into question.
In this new study, University of California, San Diego (UCSD) scientists designed a machine learning algorithm to accelerate and streamline these efforts. The approach starts with researchers generating an initial library of about half a million possible antibody sequences and screening them for their affinity to a specific protein target. But instead of repeating this process over and over again, they feed the dataset into a Bayesian neural network which can analyze the information and use it to predict the binding affinity of other sequences.
More specifically, the authors explained that they developed, “a learned representation trained on over three million human B-cell receptor sequences to encode antibody sequences” and a “variational Bayesian neural network to perform ordinal regression on a set of the directed evolution sequences binned by off-rate and quantify their likelihood to be tight binders against an antigen.”
“With our machine learning tools, these subsequent rounds of sequence mutation and selection can be carried out quickly and efficiently on a computer rather than in the lab,” said Wei Wang, PhD, professor of cellular and molecular medicine at UCSD School of Medicine.
One particular advantage of their AI model is its ability to report the certainty of each prediction. “Unlike a lot of AI methods, our model can actually tell us how confident it is in each of its predictions, which helps us rank the antibodies and decide which ones to prioritize in drug development,” said Wang.
Another advantage is that it can assess sequences not present in the directed evolution library, which greatly expands the search space to uncover the best sequences for experimental evaluation.
To validate the pipeline, the researchers set out to design an antibody against programmed death ligand 1 (PD-L1)—a protein highly expressed in cancer and the target of several commercially available anti-cancer drugs. Using this approach, they identified a novel antibody that bound to PD-L1 17 times better than atezolizumab (brand name Tecentriq), the wild-type antibody approved for clinical use by the FDA.
The researchers are now using this approach to identify promising antibodies against other antigens, such as SARS-CoV-2. They are also developing additional AI models that analyze amino acid sequences for other antibody properties important for clinical trial success, such as stability, solubility, and selectivity.
“By combining these AI tools, scientists may be able to perform an increasing share of their antibody discovery efforts on a computer instead of at the bench, potentially leading to a faster and less failure-prone discovery process,” said Wang. “There are so many applications to this pipeline, and these findings are really just the beginning.”