Twelve.
That is how many scientists Superluminal Medicines—which, in the beginning of 2022, was just a concept to develop drugs to membrane-bound proteins with multiple conformations computationally—needed to create a drug pipeline that now includes six programs.
“We were running six programs with 12 people, which is really tight—and they’re all [chemists],” Cony D’Cruz, CEO of Superluminal Medicines, told Inside Precision Medicine. “That can happen when you use computation effectively.”
By the end of last year, D’Cruz said that the board could see that we were making significant progress, particularly on their lead program, which sent Superluminal Medicines into a Series A. This week, Superluminal Medicines announced that the funding round had closed, raising $120 million from a syndicate of investors led by RA Capital Management. Existing investors Insight Partners, NVentures (NVIDIA’s venture capital arm), Gaingels, and new investors Catalio Capital Management, Eli Lilly and Company, and Cooley LLP participated in the financing.
The funds will move Superluminal’s lead program from animal testing to the clinic, expand the pipeline, and advance their computational and data-heavy platform. They’ll also expand the team.
Computer speak
No matter their size, biopharma companies looking to incorporate computational technologies into drug development face the challenge of integrating the most recent algorithms and models into their workflows—it is often a struggle.
D’Cruz stated, “For years, [computational approaches] did not work, and you were always beaten by medicinal chemists, who would say, ‘None of your calculations actually work.’”
What has allowed Superluminal Medicines to move so quickly is that they began with a “green field” that they seeded with select people who were open to incorporating novel approaches into the workflow rather than just approaches they believed in based on their experience.
“When [the computational approach] finally starts working, you do not want to deviate from it because you find what works and stick with it,” D’Cruz said. “Because there are so few people [at Superluminal], we cannot have a particular bias or mantra like, ‘I did things this way for years, and it worked.’ That’s not going to work. We find novel ways of deploying technology.”
The researchers at Superluminal Medicine are chemists, not computer scientists who write code, so they have combined freely available machine learning algorithms, like ChatGPT, to create a drug discovery engine that facilitates a “predict-design-test” architecture, as the company describes.
“It is all about how you deploy [the algorithms], and we have managed to find ways of deploying publicly available algorithms in a unique sense, as well as combining them with proprietary methods to give us these answers,” D’Cruz said.
Remote viewing
Cony and the computationally literate chemists at Superluminal Medicines are trying to solve the problem of generating candidate-ready compounds in silico for complex and unsolved multi-conformational membrane proteins like GPCRs. The crux is that capturing the structure of the protein in its active state necessitates understanding how to elicit that conformation.
“It is almost like the chicken and egg problem—you need the ligand to stabilize the protein to get the structure,” D’Cruz explained.
Superluminal Medicines’ approach starts by using data to model a GPCR target in multiple states, which isn’t dependent on a highly accurate protein structure. What they need is a predicted activating binding site in the membrane-bound protein that’s unique to begin designing the matching small molecule. However, for lead optimization, cryo-EM becomes almost necessary to enable running physics-based methods that properly model the ligand-binding interaction.
Fittingly, Superluminal has access to more cryo-EMs at academic centers worldwide than scientists manning its pipeline. The company sends specially produced protein samples that the company’s scientists scan and analyze remotely using the same technology astronomers use to control telescopes worldwide from their desktops. With this approach, they’ve generated structures of multiple conformations for all six of their programs for GPCR targets.
High value, low risk
Superluminal Medicines is not ready to share its targets and indications. D’Cruz said the reason is that drug discovery is complicated and unpredictable, and because all sorts of things can happen that lead to the dismantling of a program, the company is being very judicious in terms of revealing the targets it’s working on until it reaches an inflection point.
Even the target class of GPCRs is a large family of integral membrane proteins that are drug targets for ~35% of all approved drugs. Yet 70% of the more than 800 GPCRs are undrugged, and only 138 have experimental active-state protein structures.
All D’Cruz was willing to share was that they were looking for safe bets regarding the line to and through the clinic, even in a crowded space, and immense value.
“As we were building the platform and the company, we didn’t want also to take biology risks, so we selected targets that were very well understood from a biology perspective in that they’ve been drugged, with proven drugs on the market,” said D’Cruz. “The mode of action may differ—maybe peptides, injectables, etc.—and we were going after small molecules. So, the route to the clinic—the biomarkers, animal models, translatability—is very well understood.”
D’Cruz continued, “In neuroscience and Alzheimer’s, there’s very little translatability between the animal models and humans. We’ve cured mice from Alzheimer’s but not had the same success in humans, whereas the translatability of the targets we’re working on is really there.”
Protein agnostic, process benefactor
D’Cruz fully expects to be able to expand Superluminal’s platform to other receptors within the membrane because of the large number of uncharacterized membrane-bound targets. When asked about the applicability of this approach to other proteins, D’Cruz responded, “It’s all about the protein prep—if you can prep it appropriately. It would be a challenge with something like intrinsically disordered proteins, but anything else is up for grabs, especially if it’s in the context of the membrane.”
The potential universal applicability of the approach is likely a major reason Superluminal has attracted noteworthy investors. D’Cruz said that with Eli Lilly Company, there’s no obligation attached to any of Superluminal’s programs—it’s purely an investment. “I think that’s good for us,” said D’Cruz. “It gives us optionality, but given their investments, there is an intent to get to know us more.”
In the case of NVIDIA, Superluminal is receiving a ton of compute support in addition to the financial investment. “They have extraordinary capabilities that we can work with them together on,” said D’Cruz. “The thing that they do spectacularly well is make sure that you can maximize the use of their hardware. We’re very compute heavy, and if you think about it, the limitation that we have is compute power.”
To do this, Superluminal uses not only NVIDIA’s GPU infrastructure but also the massive Google Cloud Platform. According to D’Cruz, this move positions the company to do something perhaps unexpected—share its process.
“We’re more than happy to share because any advances in how people do it, then obviously we can adopt,” said D’Cruz. “We’re not provincial and thinking that we have all the good answers. So, we’re more than happy for other people to start along the path with us so that they could keep on improving the algorithms, processing, software, etc., so that we can benefit.”
Automated and totally de novo
To further advance its drug discovery engine, Superluminal is moving from conditional to total de novo drug design.
“We started with the conditional de novo design, which is an idea that you feed the algorithm a couple of data points in terms of compounds that are hit, so it will come up with compounds that look similar or should behave similar,” said D’Cruz. “Then you move to the next step, which is total de novo design based on the information around the target and the target family.”
D’Cruz dreams of a scalable and fully automated process where proteins are continually developed in the background. All that would be required is to enter the target, which the system would then generate multiple conformations for and perform a virtual screen to identify compounds. To D’Cruz, that’s the easy part. The hard part comes after.
“The roadblocks come downstream in terms of actually synthesizing those compounds and testing them, but even that is becoming automatable,” said D’Cruz. “It’s restricted at the moment in that the amount of chemistry space isn’t that broad. However, there’s a starting point to get you some early compounds that could be tested. Of that, significant time is still with cell-based assays, but even now you can do membrane-on-chip assays. The throughput is relatively low, but it’s increasing. So, you can envisage a time when the whole process is really automated. I don’t think it’s going to be in the next few years, but certainly in the next 5 to 10 years.”
Instead of six programs, Superluminal could have 60 or 600. What’s more, they may even be able to go back to just having 12 people.
