Although expression titers have increased significantly since the 1990s, biopharmaceutical scientists still have much to learn about engineering cell lines for biologics manufacturing, especially as therapeutic proteins grow in diversity and complexity. At the BPI Theater during the 2024 BIO International Convention, BioProcess Insider editor Millie Nelson moderated a panel discussion about strategies for improving cell-line development (CLD). Conversation focused especially on how advanced computational tools, including some based on artificial intelligence/machine learning (AI/ML), could enhance host-cell genome engineering and protein design. Joining Nelson were Larry Forman (chief executive officer (CEO) of cell-line engineering company CHO Plus),

Nathan Lewis (professor of pediatrics and bioengineering at the University of California in San Diego, UCSD), and Steve McCloskey (CEO of Nanome).

Engineering for Predictable Protein Expression

Nelson asked first about designing cell lines for maximum yield and quality. Lewis noted that CLD scientists historically have focused on engineering cell-culture components and conditions to enhance protein expression. Thus, the industry still needs technologies for engineering cells themselves. Lewis’s laboratory at UCSD investigates methods for targeted genome engineering, especially for cell lines with complex phenotypes.

Another possibility, Forman explained, is to take a broad, whole-genome approach to cell-line engineering. That entails disruption of the Chinese hamster ovary (CHO) cell genome, random amplification of chromosomes, and selection of advantageous phenotypes. The approach might seem like applying “a sledgehammer,” but CHO Plus has had much success with it, and screening processes can provide nuance.

McCloskey added that synthetic biology and AI/ML are raising opportunities to optimize protein sequences for manufacturing. In some cases, it makes sense to start with sequence optimization and then to “work your way back” to host-cell genome editing for additional adjustments to protein expression.

Such tools, he continued, can be applied to increase understanding of posttranslational modifications (PTMs). Although the AlphaFold 3 model (Google DeepMind and Isomorphic Labs) has visualized hundreds of thousands of protein structures, biological factors still make PTMs difficult to anticipate. “We can’t reproduce the cellular environment

in silico right now,” McCloskey reported. But as industry feeds biomanufacturing data into ML models of increasing sophistication, scientists will begin to understand and control PTMs.

Lewis recalled that PTMs can have negative or positive effects on a given protein. Some changes can compromise product quality — e.g., by initiating proteolysis. The industry has addressed some such problems by engineering them out of host-cell genomes. But some PTMs provide proteins with features that are necessary for their biological activity. “We really don’t know much about what controls the addition of such PTMs,” Lewis explained. “There is no amino-acid sequence that you can just encode; there’s something more inherent.” Some of that adjustment comes from host cells, which will express specific machinery to add or remove given modifications.

Researchers are making headway in understanding necessary PTMs. Lewis’s group has investigated protein codes for glycan structures and has developed a “template” to help with encoding glycosylation patterns. Echoing McCloskey, Lewis said that generating more data and enhancing algorithms will enable researchers to modulate PTMs.

Forman noted that although scientists’ understanding of PTMs must improve, the industry knows enough about protein expression to forge ahead with important CLD work. He noted that researchers could investigate the influences of cellular Golgi bodies on protein quality. Scientists already know how to engineer cells to modulate Golgi number and quality, so that could be a fruitful entry point for research into PTMs. Other changes — e.g., to intracellular pH and oxidation–reduction (redox) reactions — can be engineered into cells to influence protein assembly. “Although we don’t understand [PTMs] perfectly yet . . . the cells know what they are doing, and we are guiding the cells,” Forman said.

When asked whether cell lines could be designed for complete predictability, the panelists agreed that some randomness is inherent to working with living cells. Forman compared the ideal of complete predictability to developers’ enthusiasm for platform production approaches. For decades, companies investigated the possibility of using the same expression vector, cell line, culture medium, and manufacturing process for an array of recombinant antibodies. As it turns out, each antibody requires a tailored expression system. “We would like to think that we can engineer cells for predictability, but the best thing that we can do is to have a constellation of cells with different phenotypes.” The same can be said for production of adenoassociated virus (AAV) vectors: Slowly, the industry is realizing the impossibility of creating a single human embryonic kidney (HEK) cell line for all AAV serotypes.

Considering the difficulty of establishing platform approaches, Lewis emphasized the importance of libraries for cell lines, vectors, and related raw materials. A panel of cell lines could be engineered to perform specific PTMs for most recombinant proteins. As needed, AI/ML could be applied to identify ideal expression systems. Such an approach could facilitate early screening activities.

McCloskey highlighted related efforts in the small-molecule pharmaceutical industry, particularly a movement toward open science. Chemical libraries with billions of compounds are already available worldwide. Similar initiatives could work in the biopharmaceutical industry, with companies releasing data related to commonly used cell lines.

Making Space for Innovation

Nelson asked the panelists about what role AI/ML could play in the industry. McCloskey observed that current conversations about AI/ML center on data analytics — e.g., for identifying patterns in genomic data. But some companies are exploring computer vision and other ML methods for continuous bioprocess monitoring. Computer vision could develop deep insights into cellular processes, enabling dynamic process control. McCloskey added that AI-based analytics already are linked inextricably with biopharmaceutical research and computational biology. Companies that delay implementing such technologies could fall behind the rest of the industry.

Lewis responded that AI will take considerable time and resources to implement effectively. AI utility depends on the quality of the data fed into it. And even with excellent algorithms and training data, an AI-based system will be unhelpful if it cannot respond adequately to user queries. Data scientists must engineer prompts carefully. Adopting an AI system also will require significant investment to establish data warehouses and automated systems for data generation, transfer, and storage.

Forman countered McCloskey, saying that generation of a sufficiently comprehensive model — e.g., about the productivity of cells with different phenotypes — requires time and labor. A business case could be made for developing algorithms for complicated analyses. “In the meantime, we are not going to wait” for AI to be developed. It might be more fruitful to hone the predictive value of existing tools.

Conversation turned to the rate of innovation in the biopharmaceutical industry. Forman remarked that only a small part of a large company has sufficient bandwidth and resources to explore new processes and technologies. Innovation occurs incrementally. “But big changes are going to come from [taking] small stabs at a big idea.”

McCloskey emphasized the value of industry consortia in accelerating innovation. Citing the COVID-19 pandemic, he said that research and development (R&D) can occur rapidly amid a willingness to share data. He advocated for a “distributed” innovation ecosystem in which networks of companies share intellectual property (IP) as well as costs, risks, results, and rewards from R&D activities.

Lewis added that emerging technologies could facilitate consortia. Contributors could pool data into a secure system that leverages AI to mask proprietary information while enabling analysis of anonymized data. Setting up consortia with such capabilities could foster innovation through collaboration.

Plotting the Path Ahead

Nelson concluded the discussion by asking how biomanufacturing might look in five years. McCloskey expressed excitement about the number of companies that would be applying AI/ML capabilities to transition products from idea to proof of concept. Lewis said that ML developments are encouraging, but drug makers should expect some resistance from investors, regulatory agencies, and other stakeholders. Historically, the industry has adopted technologies methodically. Regulators will expect drug companies and technology developers to demonstrate that emerging tools will provide high-quality insights, especially as companies adopt AI to accelerate development activities — sometimes hastily so. Forman ended the session with a nod to cell and gene therapies, predicting that many such products will reach the market soon and that their costs will begin to decline. He welcomed advances, AI included, that would increase advanced-therapy availability.

Brian Gazaille, PhD, is managing editor of BioProcess International; [email protected]; 1-212-600-3594.

Watch Online

See the full panel discussion online at https://bioprocessintl.com/category/bpi-theater/bpi-theater-bio-2024.