Researchers from New York plan to screen a library of 1010 molecules to train an artificial intelligence (AI) model. The model, developed by software startup DeepSeq.AI, aims to help manufacturers assess the likelihood of protein-based drugs, especially harder-to-produce ones, being successfully expressed by Chinese Hamster Ovary (CHO) cells at a commercial scale.
“For 15 years or so, the holy grail on the development side [of protein production] has been trying to figure out which candidate molecules [are manufacturable],” explains Susan Sharfstein, PhD, professor of nanoscale science and engineering at the University at Albany.
“Researchers often discover a molecule with amazing binding properties [to the drug target], but when it comes into development, it aggregates or simply doesn’t express well at scale.”
To address this challenge, Sharfstein teamed up with DeepSeq.AI with a plan to express a library of single-chain antibodies (variable domain of a heavy chain-only antibody, or VHH) on the surface of CHO cells, a cell type commonly used for antibody manufacture.
The collaborators hope they can eventually compile a library that AI can use to understand the impact of each amino acid and, in the longer term, predict the likelihood that an antibody can be manufactured within CHO.
“We’re looking at the impact that each amino acid has on the expressibility of VHH molecules and, if this works well, we can look at whole antibodies and then biospecifics, which are molecules that nature never intended,” she says.
According to Sharfstein, after DeepSeq.AI first developed its tool, they used it to help them develop the initial library. Sharfstein and her team now plan to clone the library into an appropriate expression vector. They will then transfect the library into CHO cells, label them with a fluorescent antibody, and use that to sort good from fair to poor expressors, she says.
After isolating the DNA from the good expressors for the whole library, they then plan to send that back to DeepSeq.AI for next-generation sequencing and use those results recursively to train the AI further, generate a new library, and do more experimental analysis.
According to Sharfstein, her approach working with DeepSeq.AI is quite unusual for using a small number of molecules expressed on CHO.
“We have a low number of molecules for each cell and what’s also unique is that when people do cell surface display, they often sort and amplify, whereas we’re doing a next-generation sequencing step, which allows for a greater [coverage] of the [screening] space.”
Once they’ve worked on VHH molecules, Sharfstein hopes to move on to screening a library of single-chain FV molecules—fragments of antibodies that are the smallest unit capable of binding an antibody to a target molecule.
“When you’re making bispecific molecules, if your single-chain FV doesn’t express well, the bispecific doesn’t express well, and FV is simpler than [screening] entire antibodies,” she says.