SAN DIEGO — In the first months of their startup’s existence, Tom Miller and Fred Manby sent an invitation to the research chemist and drug industry blogger Derek Lowe to join their scientific advisory board.
Lowe, who publishes on Science’s website with takes on the life sciences world, has gained a reputation for his grounded views on AI, guided by nearly 35 years as a drug chemist. Miller and Manby were building the type of AI-focused biotech at which Lowe has often raised an eyebrow. But the co-founders wanted his perspective, acknowledging their own lack of expertise in drug discovery as two theoretical chemists who just left academia. Lowe was to play the role of “the reality check guy” as the startup, now called Iambic Therapeutics, grew into a real lab, using real AI software and making real drugs.
“It hasn’t always been the most cheerful advice for any computational company to hear,” Lowe said in an interview with Endpoints News. “I’m not going to be someone who just comes in and tells them, ‘Yeah, looks great. You guys are wonderful.’ You can get a lot of people doing that for you.”
Iambic is now about four years old, with an AI-designed cancer drug in the clinic and another close behind. The startup has its own fancy AI models that it believes can go toe-to-toe with anyone else. But it has relentlessly focused on applying those models to make actual drugs against carefully chosen targets.
On Tuesday, Iambic added $50 million to an existing Series B, led by new investors Mubadala Capital and Exor Ventures, the company exclusively told Endpoints. The new money brings its total raised to over $200 million.
Endpoints recently visited Iambic’s new San Diego headquarters to see a demonstration of its AI software and to talk with its leaders about how they make medicines. Iambic doesn’t have the head-turning megaround money of other AI bios like Xaira Therapeutics, Generate:Biomedicines, or insitro. Instead, Iambic has built itself into a force to watch by making its own work less theoretical, connecting its AI predictions with a wet laboratory to test and improve them, all with the goal of making better therapeutics.
“AI, in a vacuum, will not lead to drugs,” Miller said in an interview. “This has to be tightly coupled with the making of those molecules, with the testing of those molecules, with closed-loop refinement of the machine-learning models in a tightly integrated platform.”
After four years of watching and advising Miller and Manby, Lowe thinks they have as good a shot as anyone.
“What people want out of AI is new drugs, real fast,” Lowe said. “That is the moment we’re in right now. We’re just to the point where we’re starting to put that to the test with modern technology. I have just no idea what is going to happen. I just hope we can look back and say this is where it began.”
‘Hot and live’
Core to Iambic’s approach is the belief that it will take more than a single AI breakthrough to move the needle in drug R&D. Its internal software system, called Insight, stitches together Iambic’s suite of algorithms — many developed in an ongoing research collaboration with Nvidia — into a coherent, AI-driven way to design and test molecules. Rather than a medicinal chemist booting up a tool for a single structure or property prediction, Insight dictates a new way of working in the lab.
Building a pipeline wasn’t Iambic’s initial focus. When it recruited Lowe to its board, Miller said he hoped to develop AI software to make chemical predictions and sell it to everyone from drug companies to industrial giants like Toyota, Procter & Gamble, and Dow.
But by 2021, Miller said the software worked well enough that they decided to keep it for themselves. Three years later, the startup moved into an unassuming office park about a 15-minute drive from the sandy beaches and scenic cliffs of Torrey Pines. About 40 of its 70 employees work from here, and Miller said they’re already outgrowing the space.
In a corner conference room, a group of Iambic scientists huddled around a laptop to show how they make their drugs. The demo started with KRAS G12C, a notoriously hard-to-drug protein that drives many cancers.
Iambic’s Insight platform starts by using the company’s NeuralPLexer model, similar to Google DeepMind’s AlphaFold in predicting the 3D shape of a protein and a small molecule. With a starting drug candidate, the visualization showed the drug binding — at least a little — to the target. That start gives what’s known in industry lingo as “room to grow into a pocket,” almost like a rough cut of a key that can be refined to better fit a lock.
From there, the chemist opens the next page in the digital notebook, using Iambic’s generative search algorithm called Magnet to make 1,024 possible molecules. Each fills a spreadsheet row, with columns dividing the molecule by its backbone and a building block added through a chemical reaction. The end result is a potential drug candidate.
In the background, Iambic’s system allows its scientists to restrict these 1,024 molecule ideas to what is actually synthesizable in the lab. The AI doesn’t dream up compounds that can’t be made. Even more practically, the team can filter molecules to what can be immediately made a few feet away in their San Diego lab, instead of needing to buy chemicals from vendors, which can add days to each cycle of development and testing.
Scrolling through the rows, an Iambic chemist adds a column, generating predictions of the stability of each molecule. This action, again, taps into yet another homegrown AI model — this time a neural network called PropANE.
In two, maybe three seconds, 1,024 predictions are added. Miller calls this the “hot and live” nature of Iambic’s software. Multiple scientists can simultaneously view and analyze the same online page, and property predictions come in seconds, instead of minutes or hours. To make that happen, Iambic pays multiple millions per year to Lambda, a cloud provider of Nvidia’s powerful computing chips.
The speedy predictions help Iambic’s chemists select which of the 1,000-odd options are worthy of real-world testing. By clicking a “BUILD PLATE” button, the work shifts to Iambic’s wet lab, where lab scientists and robotic arms assemble the compounds, test them for purity and run a battery of experiments to see how the AI predictions fared.
That design-make-test cycle typically takes nine days. While every drugmaker uses at least some computational techniques, Iambic’s workflow places models and predictions at the heart of the process rather than as an add-on. The experimental data serve the purpose of fueling and retraining its algorithms, making even better guesses the next time around.
HER2 to be Iambic’s first test of AI
The KRAS demo ends with finding more potent compounds, showing how the system is supposed to work. It’d be a “dereliction of duty” to show a demo that isn’t sleek and impressive, Lowe said. But Iambic’s pipeline has results — the biotech’s HER2 drug candidate is now in a Phase 1/1b study, going from research idea to human testing in two years, far faster than the six-year industry average.
The drug, called IAM1363, shows the promise and limitations of where AI is today. Miller said there are two attributes that could help differentiate it from other HER2-targeting drugs on the market like the longtime blockbuster Herceptin or Pfizer’s Tukysa.
First, Iambic’s drug is designed to enter the brain. Brain metastases are a leading cause of morbidity for HER2-driven cancers, Miller said, and preclinical testing showed Iambic’s compound was 10 times more brain-penetrant than Tukysa.
Secondly, IAM1363 is about 1,000-fold more selective in hitting HER2 instead of the EGFR protein. Miller said boosting selectivity with tyrosine kinase inhibitors, or TKIs, could boost its effectiveness.
“We entered into it with a strong biological hypothesis that one of the reasons people were leaving efficacy on the table with TKIs is they simply weren’t pushing that selectivity as far as it can go,” Miller said. “The tools allowed us to really go after that biological hypothesis.”
While Iambic’s technology was crucial in getting those two qualities, that hypothesis is still human-generated. AI isn’t able to tell Iambic what makes a good target for a drug, and it isn’t able to say which specific properties would translate to clinical differentiation.
Lowe said the selectivity and brain penetrance aren’t impossible to attain with typical drug discovery techniques, but acknowledged Iambic progressed faster with less searching in achieving that combination.
“The question is: Will that selectivity and brain penetrance buy you anything in cancer therapy?” Lowe said. “That’s what’s going to be decided in the clinic. We don’t know. It seems like it should, but there’s an awful lot of ideas that go into the clinic that really seem like they should help and don’t.”
A major challenge for the technology remains that the areas where AI is most helpful in drug discovery are also the least impactful at addressing the core problems driving the industry’s high failure rate and low productivity, Lowe said.
Even with Iambic, Lowe said he hasn’t seen anything suggesting they can turn the technology to choose the right drug targets — the critical first step of the R&D process — or how it can speed up clinical trials, which eat up the vast majority of time and money instead of laboratory work.
“You’ve got to actually have people dosing people and collecting the data on that,” Lowe said. “There’s nothing you can do to simulate your way out of that.”