Faster Science, Better Drugs

We can frame this in high-level philosophical goals, such as accelerating scientific progress. While that may not be tangible for everyone, the most important thing is that science operates in the real world. Unlike AI research, which moves quickly through GPU iterations, real-world science requires physical manipulation. We must move atoms and transfer liquids from tube to tube to create life-changing medicines. These processes occur in real-time, involving the growth of cells, tissues, and animals. The promise of applying machine learning to biology is the ability to massively accelerate and parallelize these efforts.

We built ARC as an organizational experiment to see what happens when you bring together neuroscience, immunology, machine learning, chemical biology, and genomics under one physical roof. By increasing the collision frequency across these five distinct domains, we hoped to uncover a vast space of problems that would otherwise be inaccessible. While these individual fields are represented across various universities or geographical regions, people are often distributed. We aimed to bring everyone together.

There's controversy surrounding what it means to be an accurate biological simulator or a virtual cell. We can't measure everything, such as metabolites, with high throughput and spatial resolution. Capability will evolve through different phases. Initially, individual cells will be modeled, then pairs of cells, followed by cells within a tissue, and eventually in a broader, physiologically intact animal environment. These scaling and complexity layers will aggregate and improve over time.

If the goal is to reach an "AlphaFold moment" for virtual cells, where a model consistently provides useful, folded structures 90% of the time, how far are we from that? For instance, if we ask a virtual cell model to shift a cell from state A to state B, and 90% of the time the suggested perturbations successfully achieve this shift experimentally, that would be a significant milestone. How close are we to achieving that level of reliability and predictive power in virtual cell modeling?

I believe we can add many biological evaluations to these models over time. These are tangible, textbook examples, unlike what early generations of models do today. Current models focus on quantitative metrics like mean absolute error over differentially expressed genes. While these are standard machine learning benchmarks, we aim to increase sophistication. We want to create something explainable to an experienced professor who has never used a computer terminal.

The challenge in our industry is that bottlenecks persist. The biggest bottleneck, which is necessary, involves proving that our creations are suitable for their intended purpose. We must ensure they are as de-risked as possible before human application. This bottleneck is crucial and should remain. I'm not suggesting we eliminate it. However, we need to find ways to reduce the cost and time associated with navigating the bottleneck of human clinical trials.

GLP-1s have demonstrated the immense value created by targeting large patient populations. This has significantly boosted the industry's ambition, both for investors and drug developers. We should continue to pursue this trend. The positive trajectory of GLP-1s is undeniable. The value generated through their increased use, and the subsequent transfer of value to companies like Lilly and Novo, is well-deserved.

That's the bottleneck. We joke about it, but you have to find a molecule that can be tested in mice, then rats, then monkeys, and finally humans. This process takes a long time and is incredibly difficult to compress. Therefore, when you undertake this journey, it must be worth the effort. Failing at the end of such a process is obviously devastating. The industry desperately needs ways to ensure that this path leads to a successful journey as often as possible.