My graduate advisor used to say that I could do all the work I wanted on a computer to develop a hypothesis, but no matter what, I would have to work with living things to prove it. By developing its own in silico assets, the generative AI company Ordaōs is ready to do the functional work independently.
After using contract research organizations (CROs) for several years to assist in bringing these in silico mini-proteins to life, Ordaōs officially opened its lab space at JLABS @ NYC in June 2023 to kick off its partnership with Johnson & Johnson Innovation. As a JLABS resident company, the lab space will optimize Ordaōs’ R&D efforts and grow its capabilities to fuel its novel approach to drug design. Ordaōs works on overlooked rare diseases, primarily focusing on cancer, with programs in triple negative breast cancer and pancreatic cancer.
Ziwei Liang, PhD, chief science officer at Ordaōs said that the residency at the JLABs facility in the SoHo district of Manhattan will add three main streams of value. First, it will close the continuous learning loop from in silico to in vitro by providing high throughput and quality data generation; this can feed back to the in silico engine at Ordaōs with proprietary data for training and improvement. Second, it will fast track asset development by bringing in binding and functional assays designed in-house to fit specific targets. And third, it will provide the opportunity for technological innovation from the biological side.
“We are open for business,” said David Longo, CEO of Ordaōs. “We’re an asset foundry that has both an internal and an external pipeline that we are happy to partner with both biotech and pharma. And we recently opened our lab and are excited to continue to grow and serve the world’s ailing populations.”
The in silico funnel
Longo isn’t the first CEO I’ve met recently who notes that there is a sentiment in the drug development industry that nobody is actually using AI and that they’re one of the few actually using AI in the industry (Alex Zhavoronkov of Insilico Medicine is the other).
“Ordaōs is a company that brings the order of digital bits to the chaos of physical atoms, using generative AI technology in protein design,” said Longo.
But it’s not the garish and borderline-gimmicky elevator pitch that convinces me that Ordaōs, which is a portamento of “order” and “chaos”, is actually using AI. Rather, it is how Longo explains the AI-driven drug discovery funnel at his company. Ordaōs is designing mini-proteins, which Longo said is a greenfield modality that provides many advantages over antibodies. “They’re twenty times smaller, more stable, easier to produce and manufacture—and we’re creating them from scratch with AI,” said Longo.
All of their internal and external programs pass through stages of QC (for purity and aggregation) and characterization for binding to the target epitope, proteins, and cells before ever going into functional testing.
“This approach is antithetical to the way that most in pharma do it, because it’s going from a component perspective, where it’s typically done starting with functional assays,” said Longo. “You’re starting with the end goal, and the only thing you can really see is success, which is a very coarse signal because you don’t really know where it’s binding, and the failures are also very coarse and opaque.”
Ordaōs makes predictions on every assay at the component level so that as it moves to functional assays, the company hypothesizes what would cause that functional result.
“We can confirm every step of the way if it’s the binding affinity that’s off, or the epitope that’s off, or maybe it’s something in the purity or aggregation,” said Longo. “The false positives are far more costly than the false negatives in AI-driven drug design, but as we are collecting so much negative data, we have a smaller likelihood of false positives.”
Hotdog, not-hot dog
To explain how generative AI works, Longo recalls a skit in the television show Silicon Valley: developer Jìan-Yáng (played by comedian Jimmy O. Yang), unveils his new app. Using his phone, he takes a picture of a hot dog, and the app spits back a green check mark next to the word “hotdog.” Then, he takes a picture of a pizza, and the app spits back a red X next to the words “not hotdog.” This kind of neural network results from training an algorithm to detect images of hotdogs—after being fed many images of hotdogs.
For generative AI, the neural networks have to come up with what they are “seeing” on their own. After being fed pictures of hotdogs and pepperoni pizza, the generative AI may come up with two classes of items: one a red cylinder, the other a circular or slice shape with spots of red amongst a white or yellow background.
Another metaphor that Longo leans on is two “children” that walk into a zoo: one is a “predictive” child, and the other a “generative” child. After walking past the tigers and elephants, they come across a zebra. The “predictive” child will classify the zebra as most likely a tiger because both have stripes, whereas the “generative” child will compare the zebra to the tiger and elephant and say that, though this has four legs like the others, it has stripes and hooves, doesn’t have a mane or a trunk, and it doesn’t roar or trumpet but instead brays—though it’s kind of like a tiger, this must be something else.
From bits to benchtop
Over the past three years, Ordaōs has been designing, testing, and back-training the development of mini-proteins in silico, and testing them functionally, using CROs.
In December 2022, Ordaōs Liang became CSO to drive the drug discovery strategy, manage multiple research programs, and lead a team of machine learning and AI scientists to develop an internal in silico pipeline for de novo protein generation. The mini-proteins generated by the AI engine at Ordaōs are not constrained to have a particular mechanism of action. For that reason, the lab will allow Ordaōs to design and execute the right assay to drive the asset forward.
“It’s very important for us to control the quality and assay development in-house, so we can really fast track this process and maintain the high-quality control,” said Liang.
With JLABS, Ordaōs will have access to facilities like tissue culture rooms and key equipment, such as an HPLC or flow cytometer, that may be difficult for an early-stage company to obtain but is necessary for high throughput and high-quality data generation and characterization. In addition to the quality, care, and safety of the established structure at Johnson & Johnson, the collaboration provided a mentor.
While JLABS provides flexibility and control over complicated assays, Ordaōs has not fully detached from its use of CROs. “We are not eliminating the option of using CROs for very productionized standard experiments, like standard protein production that is cell-based,” said Longo. “They do it very fast and have very high purity and quality.”
That doesn’t mean that the lab at Ordaōs will not look into protein production. Longo mentions that he’d like to explore cell-free protein production because it’s a unique technology that can give them a real advantage in bringing down the cost and increasing the production quality for mini-proteins.
Running of the bulls
The appetites of biotech investors for companies expanding in AI have been picking up in 2023. In mid-July, Recursion Pharmaceuticals (RXRX), an AI-focused drug developer, announced a $50-million investment by Nvidia (NVDA), a company increasingly focused on AI. This triggered a doubling of the stock price of RXRX, jumping from $6.78 on July 11 to close last month at $14.12.
While Insilico Medicine hasn’t gone public, they announced the investigational new drug (IND) filing of another preclinical candidate developed leveraging its proprietary AI drug discovery platform, Pharma.AI. The IND application is for ISM001-055, an anti-fibrotic small molecule inhibitor developed for treating Idiopathic Pulmonary Fibrosis (IPF). In the past year, Insilico nominated a handful of preclinical candidates and generated positive topline Phase I data in human clinical trials with ISM018-055 for IPF received Orphan Drug Designation from the FDA, and is nearing Phase II clinical trials.
On top of that, Moderna and IBM are teaming up to use generative AI and quantum computing to advance mRNA technology, the development at the core of Moderna’s COVID vaccine. The agreement allows Moderna to access IBM’s quantum computing systems and generative AI model. The agreement comes as Moderna navigates its post-pandemic boom driven by its successful mRNA vaccines.
By working on mini-proteins, Ordaōs may be able to cut a slice of the AI generative pie that no one else will touch. However, if Insilico Medicine and Recursion Pharmaceuticals are real generative AI companies, they may have gotten hits for mini proteins. And according to ChatGPT3, they’re not alone; there are a few other AI companies in biotech: Atomwise, Deep Genomics, BenevolentAI, Berg Health, Zymergen, and Ginkgo Bioworks. But if they are truly generative AI and are using their own programs and data sets, they are unlikely to have generated identical assets.