Broadly neutralizing antibodies (bnAbs) are one of the most powerful weapons against HIV. And like everything that is effective in the fight against HIV, they are hard to come by.

In his plenary talk at the 2022 Conference on Retroviruses and Opportunistic Infections (CROI), Mark Feinberg, the president and CEO of the International AIDS Vaccine Initiative (IAVI) told the audience that only about 10% of HIV-infected persons end up producing bnAbs to the virus, and most of those bnAbs only bind weakly. At a CROI session on Frontiers in Laboratory Science, Regina Barzilay, who is a Distinguished Professor for AI and Health at the Massachusetts Institute of Technology, described using artificial intelligence (AI) methods for identifying bnAbs from their amino acid sequence.

In the development of small-molecule drugs, the use of AI has progressed to the point where it is widely used in the pharmaceutical industry.

There, "it doesn't look like a science, it looks like a technology," Barzilay told her audience. In 2020, Barzilay and her colleagues used the approach to identify halicin (SU-3327), a molecule that had originally been identified as a potential antidiabetic drug, as a broad-spectrum antibiotic.

In her talk, Barzilay described applying the same type of algorithms to "design the perfect antibody."

What's perfection anyway?

Perfection, like beauty, is in the eye of the beholder – or in this case, in the definition of the experimenter.

Barzilay and her colleagues defined perfection as the ability of an antibody to neutralize a viral strain.

"So the first question we need to address is, given viral strain Y and antibody X... will they neutralize each other or not?"

In their work, the team used the CATNAP (Compile, Analyze and Tally NAb Panels) database, which is a relatively large database of different antibodies and their ability to neutralize HIV strains. CATNAP contains the sequences of the variable chains of 300 antibodies, about 1,000 viral strains, and roughly 32,000 virus/antibody interactions.

In the method Barzilay described, the sequences corresponding to the antibody and the virus were used to generate mathematical vectors that encode different characteristics. Those vectors were then fed into a model that learned to predict whether a specific antibody could neutralize a given viral strain.

The model learned to predict binding strengths for new pairs, although the predictions were stronger if a new viral sequence was paired with a previously encountered antibody or vice versa. But the model was able to make predictions even if both the specific antibody and the specific viral sequence were new to it.

The team tested its model on several real-world scenarios, and showed that it was able to identify people living with HIV that made bnAbs, based on the sequences of their B-cell repertoire. The model also predicted treatment-resistant strains that arose after analytical treatment interruption with the bnAb VRC-01.

Those experiments showed that the method has potential to be used in clinical applications, where it could improve the therapeutic use of existing bnAbs. At the conference, researchers from the National Institute of Allergy and Infectious Diseases reported using assays to predict the neutralizing capacity of VRC-01 LS and VRC-07-523LS in a clinical trial.

While Barzilay and her team have focused mostly on broad neutralization, other characteristics of antibodies can be used to define perfection as well, if there are existing measurements the model can be trained on. An additional characteristic to breadth of neutralization that the models consider is making sure that "the generated sequence actually looks like an antibody, that it is not some random garbage," she said.

In some of their experiments, the team has also considered binding strength as well as neutralization breadth. And more mundane but equally important aspects such as ease of manufacturing can also be considered.

In her talk, Barzilay also noted that especially in combination with complementary efforts to predict viral evolution, it is theoretically possible to identify antibodies that would be broadly neutralizing against viral strains that are most likely to appear.

"Currently, what we are doing is optimizing on the past, and breadth is defined [by] what we've seen, not what is likely to happen," she said.

And even "when people are looking into the future, they are not necessarily conditioning... for instance on particular medications. So you are just generically assuming what may come."