Since when did stock analysts start designing proteins?

An investment research firm, not a biotech company, spent roughly $200,000 this month designing proteins to bind scorpion venom. That’s a pretty unusual way to spend an investment research budget, considering that firms like Citrini Research normally make their money studying companies, not experimenting with toxins. But Citrini wanted to find out what AI could actually do in biology, so it teamed up with Anthropic and gave it a try.

On October 20, they published their results. Five people from Citrini, three with absolutely no science background and two with limited experience, worked with Anthropic’s researchers and an internal version of Claude. Together, they generated more than 318,000 protein designs in about 64 hours. They then selected 211 candidates and sent them to two external laboratories, Adaptyv Bio and Twist Bioscience, for testing. Of those, 17 met the criteria for binding to scorpion toxins.

Now, before anyone gets too excited, these aren’t antivenoms. Binding to a toxin doesn’t necessarily mean a protein can neutralize it, let alone work as a treatment in humans. The team also had considerable help from Anthropic, including access to a model that isn’t publicly available. So we’re not quite at the point where someone can open a chatbot and develop a medicine over the weekend.

But what’s interesting to me is how this whole thing was organized. Citrini didn’t need to hire an entire team of biologists or build a lab. They had a research objective, used AI to help design the experiments, and paid third-party labs to do the physical testing. About $50,000 went toward computing and another $150,000 toward the lab work. So basically, the contributed money.

And I wonder whether this could eventually become a rather normal way of doing research, particularly as autonomous laboratories become more capable and cheaper to use. We’ve covered them in past weeks and the general rush of automating bench work.

I feel at this point we are overlooking how much of an impact this could have. Consider how investors evaluate biotech companies today. Suppose a startup claims to have developed a promising cancer drug. It presents data showing how well the candidate drug binds its target, how selective it is, and perhaps some impressive results from animal studies. Investors can hire scientific advisers to review the evidence, but they’re still largely evaluating experiments that the biotech company itself has performed. They’re essentially evaluating material based on trust, that the company is truthful in their findings.

Now suppose an investor is considering putting $50 million into that company. Instead of relying entirely on the company’s data, maybe they could spend another $50,000 commissioning a few independent experiments. An AI could examine the published results and identify something worth investigating, maybe a potential off-target interaction or an experimental condition that hasn’t been tested. Assuming the compound is available, the investor could then pay a lab to test it.

That might sound a little far-fetched, but part of the infrastructure already exists. In April, Adaptyv Bio launched an API that allows AI agents to order certain lab experiments and retrieve the results. Some protein screening services start at $149 per protein. Obviously, there’s a huge difference between measuring protein binding and validating a cancer medicine, but individual experiments are already becoming something software can order.

What is perhaps the more valuable role for AI wouldn’t even be predicting whether a drug will succeed. I doubt we can reliably predict Phase 3 outcomes from a handful of lab measurements anytime soon. But AI might become quite good at identifying which experiment would tell an investor the most about a particular risk. If you’re about to invest millions and years waiting for a potential return, spending a few thousand dollars to investigate a scientific problem would absolutely be worthwhile.

Now take this a little further. If experiments become inexpensive enough, large investment firms could start routinely testing drug candidates from companies they're following. Over time, they might accumulate their own databases of binding measurements, toxicity findings, or other experimental results, some of which the drug developers themselves have never collected.

This is where I think things get a little strange.

In academic research, there’s generally an incentive to publish what you discover. That’s how scientists share knowledge and build their careers. An investment firm operates rather differently. If it pays for an experiment that reveals something important about a drug candidate, that information could give it an advantage over other investors. Publishing the finding will eliminate that advantage.

We’ve worried for years about negative results that never get published, especially when pharmaceutical companies have little incentive to share unsuccessful experiments. But in this scenario, the company developing the medicine might not even know that someone else has tested its drug and found a potential problem.

Imagine scientists spending years developing a drug candidate while an investment firm’s private database contains experimental evidence that could have changed the direction of their research.

Of course, the same technology could work in the opposite direction. I do think there’s a potential upside we’re not exploring. For example, patient advocacy groups might eventually commission experiments on rare diseases that attract little commercial interest. Smaller research organizations could gain access to sophisticated laboratory capabilities without having to build the infrastructure themselves. There’s a pretty compelling case for making this kind of research more accessible.

In this particular case, Citrini did publish their work, so none of this is a criticism of what they did. We’re also still a long way from inexpensive, automated testing that can meaningfully predict whether medicines will work in patients.

But I think the possibility is worth considering. We’re accustomed to thinking that cheaper experiments should accelerate scientific progress, because more people can investigate more ideas - that probably will happen.

What I’m less certain about is whether all that additional knowledge will actually make its way into the scientific literature.

We might eventually have a situation where the people who know the most about a particular drug aren’t necessarily the scientists developing it, but the investors secretly paying to test it.

Here’s something from our drug development database that puts this into perspective.

We have 66 AI drug companies that have raised a combined $6 billion in disclosed funding, even though every drug programme we track for them is still preclinical.

The biggest are XtalPi at $860 million, Earendil Labs at $790 million, NewLimit at $750 million, and insitro at $640 million. So we're talking about billions of dollars invested before any of these programmes have even reached human testing. And much of the evidence investors have to work with comes from the companies themselves.

That's fairly normal in early-stage biotech, of course, but imagine if investors spent just 0.1% of that $6 billion checking some of those claims independently. That's $6 million for laboratory experiments, which could go quite a long way, especially as automated testing becomes cheaper.

I can see why investors would find that valuable. What I’m less certain about is whether they’d have much reason to share the results, particularly if those findings gave them an advantage over everyone else.

OpenAI is moving further into drug development with its acquisition of Orchestra Bio, a company that builds AI agents to help pharmaceutical companies manage their research programs. Think of all the different studies, manufacturing steps, and timelines involved in developing a drug, where a delay in one area can affect everything else. Orchestra's software helps keep track of these moving parts, and OpenAI plans to integrate it into Rosalind, its platform for life sciences. [Link]

The FDA has finalized new guidance for drug companies offering apps alongside their medicines. Some of these apps are fairly simple, like medication reminders, while others connect to medical devices or help patients manage their treatments. The guidance clarifies when information from these apps needs to be included in official drug labeling and what may require FDA review. It's becoming increasingly relevant as more medicines come with some kind of digital component. [Link]

China is considering new export restrictions on 40 technologies, including advanced AI models and biotechnology platforms. The proposed list covers AI agents, cell and gene therapy technologies, and RNA-based drug platforms. Now, this doesn't necessarily mean companies outside China would lose access to these technologies, since certain licensing arrangements could still be permitted. But transferring ownership or exporting the underlying technology could become more complicated. The proposal is open for public comment until November 15, while separate export-registration requirements have already taken effect. [Link]