When Daniel Haders, Ph.D., was working in venture capital, he was on the hunt for an AI-driven drug discovery company that truly deserved his investment. What he found instead would change his professional trajectory.
Haders and his team had no trouble identifying companies full of data and computer scientists trying to solve chemistry or biological problems without understanding the biological mechanisms at play. Other times, they found chemists and biologists who looked at AI as a commodity, rather than something that required its own expertise. And even when they came across companies with a more expansive strategy, those teams never seemed to have successfully developed a drug.
“We were struck by disciplines that didn't seem to be aggregating together, so ultimately, we realized we needed to build that company,” Haders told Fierce on the sidelines of Fierce Biotech Week in Boston Thursday. “We pulled a team together to build what we saw as the next generation of an AI drug discovery company.”
Haders has spent the last seven years doing just that. After founding Model Medicines in 2019, the CEO handpicked employees who understood data science, computer science, chemistry, biology and drug development. While that approach may seem straightforward, Haders knew that to break through the hype of AI drug discovery and build tangible results, he needed to think differently.
“We thought that everybody should be zigging while they were zagging, from a technology point of view,” he said.
Beyond its workforce, the company, which was a Fierce Medtech 15 honoree last year, also tried to rethink how it trained its AI models. Haders explained the philosophy using the interview room as a metaphor for what is possible.
“If each point in this room is a possibility, but all of your training data comes from a blob that's literally right here, your models will learn this one area of chemical space, but it won't be able to discover a drug that's actually in that corner over there,” he said.
In the world of venture capital and drug discovery, breakthrough and first-in-class treatments are king. So Haders built his company to go against the grain. He wanted smaller, tight-knit staffing groups rather than expansive ones. In a similar way, the company used small yet diverse data training sets to predict outlier chemistry.
The AI models were trained on more versatile central processing units (CPUs) rather than the faster graphics processing units (GPUs) often preferred as a way to perform parallel computations required for machine learning.
The result, according to Haders, is a company that has virtually screened 325 billion molecules and discovered drug candidates producing promising data. “We create creative models,” he said.
Haders highlighted MDL-4102 as an example of what the company’s creative AI can do. The preclinical candidate targets androgen receptor and bromodomain 4 (BRD4), a protein that regulates gene expression and was at the center of Johnson & Johnson’s $3 billion acquisition of Halda Therapeutics last year for its prostate cancer drug.
But BRD4 inhibitors have also hit several roadblocks over the years due to dose-limiting toxicity. Researchers in The British Journal of Cancer wrote in 2020 that “there is no doubt that BRD inhibitors have an untapped potential as anticancer drugs. However, there is no obvious ‘low-hanging fruit’ for their development.”
Haders says MDL-4102 is an example of how Model’s AI can climb the tree to pluck fruit from the top. He noted its potential to address oncology, cardiovascular and even more degenerative indications. “It has a combination of potency, selectivity and novelty that supersedes everything in the literature by multiple forms.”
A product of Model’s 325 billion-molecule virtual screen, MDL-4102 is currently in IND-enabling studies, with hopes to request permission next year to enter human studies.
Next year is also when the La Jolla, California-based company hopes to take MDL-001—a broad-spectrum antiviral targeting a conserved viral polymerase mechanism—into the clinic. Model toured various infectious disease and hepatology conferences last year, touting preclinical data that the company said “demonstrated preclinical activity across respiratory and hepatic viruses and high-risk co-infections.”
The preclinical results also included findings that suggested that once-daily doses of MDL-001 carry potentially as much antiviral efficacy as remdesivir, known as the COVID-19 treatment Veklury, and were able to reduce viral accumulation in the lungs better than nirmatrelvir, one of the two active ingredients of Paxlovid.
Haders emphasized the importance of a training model that isn’t bloated as a key to what he hopes will be a fleet of approved drugs developed by AI.
“Outliers are what you're going for. If you have too much data that's too dense, you will only rediscover what has already been discovered. You may discover a lighter yellow or a darker yellow, but you won't discover a new color,” he told Fierce.
“We need to discover new colors,” the CEO added.