Sponsored Insight
We know the scientific method to be sequential and empirical: Ask, observe, hypothesize. Test, analyze, report. And then do it all over again, and again, and again.
For better or worse, that sequence has defined drug discovery — bringing remarkable new therapies along with heaps of failed ones. But as generative AI advances, it might be time for a refresh. Instead of asking questions first and testing molecules, proteins, and other targets one by one, maybe we could run simulated tests for everything at once and then derive questions from there — iteratively.
AI-designed antibodies from Genentech are a good example. Using a machine learning model, researchers at Prescient Design, a Genentech accelerator, have applied an optimization method known as “walk-jump sampling” that can generate antibodies with a 92% expression rate, rivaling those of B-cell clones. The paper won an outstanding award at the International Conference on Learning Representations, a global AI-focused industry event.
The finding is revolutionary, but so is how it came to be. Genentech’s discovery was born from a new strategy called the “lab in a loop” that uses data from the laboratory and clinical studies to train models that can then be used to design better experiments and better medicines. It’s an iterative cycle that transfers knowledge from one project to the next. Richard Bonneau, Ph.D., vice president of machine learning for drug discovery at Genentech and coauthor of the paper, explained:
“We’ve gone through — at this point, 12 cycles — of moving the model, using all the data to update it, then taking that updated model and selecting designs that both give us good shots on the goal but also optimize our chances of improving,” said Bonneau, who spoke at a fireside chat during the 2024 Breakthrough West Summit in San Francisco. “After 12 rounds of that, the models have gotten substantially better. And now for any particular target, we can perform a single iteration to kickstart the projects, this can be applied to any new project at Genentech from the start.”
Uniting different points of research timelines
Genentech’s interest in iteration has led to new possibilities in drug discovery, with antibody engineering among the most exciting — or at least, provable — so far. Also on their radar: small- and large-molecule design plus other types of therapeutics. Not only could iterative machine modeling bring these new assets to preclinical and clinical junctures faster, but it could also help prioritize the jobs those targets take on.
“Molecules can do multiple more things at once,” Bonneau said at the fireside chat. “So how do we actually think about designing a good target when we can have three or four valences, when we have affinities that are recruiting degradation processes, or are pH or tumor environment sensitive? When we know that we have the capability of making molecules that bring these functions together, well, now we have to really up our game.”
Another benefit of having readily available target options is that they could help hedge against immunogenicity and other risks later on, he added.
That ability to find backups when Plan A targets peter out helps give research a nonlinear, time-jumping feel — where late-stage insights inform early days and vice versa. This makes drug discovery more agile, and ideally brings new medicines to market faster.
“At any point in the drug discovery process, you can have the considerations of the other parts,” Bonneau said in a STAT Brand Studio multimedia lounge interview during the Breakthrough West event. “So if you’re late and you run into a dead end, you have something that can go back to that ensemble of other pretty good designs to pick back up without a committee meeting. Instead of that meeting, you get the answer in a minute.”
A meeting of multiple superpowers
Genentech’s work is supported by strategic partnerships with industry leaders, something Bonneau considers essential in what he calls a “zeitgeist” moment of R&D.
“Things are happening,” he said in the multimedia lounge interview. “Machine learning is driving science … No matter how big a company is, it’s great to partner to see where your blind spots are, to see where you could accelerate things.”
Among those team-ups is a multi-year collaboration with NVIDIA to run Genentech’s algorithms and models on NVIDIA’s computing infrastructure. As one of the world’s first biotechnology companies, Genentech has a legacy in life sciences and NVIDIA pioneered the AI accelerated computing platform. In turn, the collaboration made sense not only because of the two organizations’ complementary strengths and shared vision for AI-driven drug discovery, but also because they’re inclined to push each other’s limits. As they work together, their iterative process flows both ways — with Genentech tweaking its models and NVIDIA its machines.
“[We’re] able to be intellectually challenged every single day,” said Kimberly Powell, vice president of healthcare at NVIDIA during the interview. “[We’re asking each other questions like] ‘If you had all the memory in the world, what kind of new architectures for biology might you invent? And if you could think beyond what today’s computers can do … what would a million times more computing mean to the scientific discovery process?’”
Better science is more representative science
Iterative AI-powered drug discovery feeds on health care data — something Genentech has no shortage of after decades spent generating and digitizing it. And as the undercurrent of biological engineering, that data has to be not just plentiful, but also diverse.
It’s a responsibility Genentech says it takes seriously. If researchers keep operating under the ask-then-test model, you have to wonder if those ‘asks’ are biased. Perhaps reconfiguring the scientific method could make for not just better science — but more representative science, too.
“It is important that we stand collectively, undaunted, and continue to challenge the status quo; the lives of patients, of course, depend on it,” said Erica Taylor, Ph.D., vice president and chief marketing officer at Genentech, in the summit’s opening remarks. “What is the purpose of developing these medicines — advancing science and treatment for novel and new modalities — when people who can benefit the most can’t access them?”
Read more at genentech.com.