0:00
Mithrl just raised a $20 million Series A to help drug companies figure out which of their AI-generated hypotheses are actually worth testing in the lab.
0:16
Welcome to Startup Bios, I'm Brian McCullough.
0:18
Like and subscribe to hear about the newest startups coming out of stealth and raising new rounds almost every day.
0:24
About 15 years ago, Vivek Adarsh lost his grandfather to acute kidney disease.
0:29
And at the time, there was no FDA-approved drug aimed specifically at that condition.
0:35
Adarsh says that gap is a big part of why his company exists.
0:40
He told GEN, quote, "the question for me has been, what can technology do to accelerate getting medicines to patients a lot faster?"
0:49
The company is called Mithrl, it's based in San Francisco, and he co-founded it in 2023 with Shara Balakrishnan, who runs the technical side of things.
0:59
This week they announced a $20 million Series A led by Obvious Ventures, with Headline, AGI House, and a handful of pharma executives also coming in.
1:08
That's on top of a $4 million seed round that Bonfire Ventures led back in November 2024.
1:15
Adarsh doesn't come out of biology, either.
1:18
He worked on NVIDIA's CUDA compiler and did distributed systems work at HP Labs, so this is a systems guy who walked into drug discovery.
1:28
What Mithrl does, basically, is help drug companies decide which of their ideas deserve to be tested.
1:32
Adarsh's argument goes like this, quote, "Every player in this category is racing to generate more raw hypotheses, faster.
1:40
But the industry is cracking under the weight of hypotheses it can't triage or validate." end quote.
1:47
We've spoken a lot lately about AI for biology, and most of the attention has gone to the generating side of things.
1:54
You've got models folding proteins, models coming up with new molecules, and so on.
1:57
Coming up with candidates has gotten really cheap.
2:02
What hasn't gotten cheap is actually testing them.
2:04
A wet lab experiment costs real money in real time, and a clinical trial costs a lot more than that.
2:08
So if an AI hands a research team a thousand plausible ideas, you haven't really solved their problem.
2:17
You've arguably given them a new one, which is figuring out which 10 are worth spending money on.
2:21
That's where Mithrl wants to live.
2:25
Adarsh again, quote, "We built Mithrl to answer a different question: which of these will actually hold up in downstream experiments, and ultimately in patients?
2:32
The platform doesn't guess.
2:35
It reasons from trusted studies and each client's in-house evidence the way a client's best scientists do, and it shows its work every time."
2:43
end quote.
2:43
The company calls its product a Scientific Decision Engine, and it helps helps to be clear about what it isn't.
2:52
It's not a biology foundation model in the AlphaFold sense.
2:55
It doesn't predict protein structures and it doesn't generate new molecular sequences.
2:58
And the company says it will never hand you a molecule to go synthesize, that stays with the drug design tools.
3:06
What it does instead is help with the calls all along the R&D process, like which target to pursue, which experiment to run next, and whether something that looked good in lab has a shot at working in actual people.
3:19
At the center of it is something they call a biomedical world model.
3:24
In Mithrl's case that means a big structured knowledge system distilled out of decades of peer-reviewed research, along with public data sets and data from partners.
3:32
It maps out how genes and pathways and diseases relate to each other, but only the relationships that have actually been established in published, verified work.
3:42
The idea is it can connect findings across disciplines and turn up links a single specialist might miss.
3:47
The company's announcement post explains why that's harder than it sounds, quote, "A rise in a gene's expression is not the same as an increase in the activity of the protein it encodes, and a system reasoning about a drug's mechanism has to know the difference."
4:03
In other words, you can't just point a chatbot at a pile of papers and trust whatever comes back, because a chatbot can make two findings sound connected when the details say they really aren't.
4:13
Mithrl claims its system keeps track of the context behind every finding, like what was measured and whether a relationship was actually observed or just inferred.
4:22
And every hypothesis it puts out comes with a confidence score and a record of exactly what the system read to get there, so a scientist can go check the reasoning.
4:33
Sitting on top of that world model is what they call an agentic harness.
4:37
That's the layer that decides which AI model handles a task and what context gets fed to it.
4:41
The two pieces together are Mithrl-1, the second generation of their platform, and early access for it was set to open on September 21st, so it's basically going live right about now.
4:53
Mithril doesn't sell this as an app you log into, which is very much a systems engineer's way of thinking about it.
4:59
The platform gets deployed directly inside each customer's own environment, under the customer's own rules.
5:04
It works with whatever frontier models a customer already trusts, so there's no lock-in.
5:08
Customers plug in their own proprietary data, and from there their scientists can build their own biomedical agents in-house, without waiting on Mithrl's roadmap.
5:18
This is the forward-deployed model we keep running into with enterprise AI companies, which means actual people show up.
5:25
to your prem.
5:25
Adarsh told Jen that most customers need help getting set up, quote, "We have forward deployed scientists that go in, work with them, help them understand what knowledgebase does, and how we can make everything standardized." end quote.
5:41
Mithril claims, that its world model turns up 16 times more primary evidence per answer than frontier models manage on their own.
5:50
On expert-graded benchmarks for scientific correctness, they say it scored 0.96, compared to 0.60 without their knowledge layer.
5:58
And in one recent benchmark, they say the platform used 45 percent fewer tokens than standard workflows running the exact same frontier models out of the box.
6:06
The logic on that one makes sense, I think.
6:09
If your agents over curated, validated biology instead of searching everywhere for an answer, they probably waste less effort and wasted effort cost compute.
6:17
None of these numbers have been independently checked, at least not yet.
6:20
As for customers, Mithrl says it's deployed right now at top-10 pharma companies, clinical-stage biotechs, and genomics platform partners.
6:29
The one name they've put out there so far is LFS Biosciences, which studies live tumor tissue outside the body.
6:39
Back in April the two announced a collaboration pairing that tumor profiling with Mithrl's AI analysis, to look for new signals about which patients respond to immunotherapy.
6:48
It also says that more customer announcements could come in the fourth quarter.
6:51
The number Adarsh seems to care about most, though, isn't any of those benchmarks, it's patents.
6:58
According to Mithrl, discoveries made on its platform have already contributed to more than half a dozen patent filings, and those filings belong to the customers.
7:05
The company says those are only the ones it knows about, and that it never set out to build a patent engine, it just sort of happened.
7:13
Adarsh told GEN that his north star is, quote, "to be able to give scientific value and uniqueness to our partners."
7:20
And he was pretty direct about why that matters, quote, "What truly draws the revenue, and the topline for pharmaceutical companies, is uniqueness."
7:28
the uniqueness he pointed to was patents.
7:32
the market has gotten a lot easier to sell into, too according to Adarsh quote, "I think there is a systemic level of education that has gone on over the past 12 months, mostly led by Frontier Labs, which is good for us." and He's seeing new pharma executives come in who already know what they want out of AI. the new money is going to two places, and Adarsh ranked them, for Jen quote, "In the order of priority, we have more demand to do deployments of our platform within big organizations, which obviously requires headcount.
7:59
Second is expansion across therapeutic areas."
8:03
The company is about 30 people right now.
8:05
The therapeutic areas they're going after first are the ones their customers care about most, like oncology, immune-related disease, diabetes and metabolic disease, and cardiovascular.
8:16
The stated mission at Mithril is to help R &D teams go from idea to IND 50 % faster.
8:21
And IND is the FDA filing that lets you start testing a drug in humans.
8:26
And Adarsh brings it right back to his grandfather.
8:31
If this all works, he says, developing a new medicine "won't take 15 years" but more like "five years."
8:35
Quote, "That is the personal motivation that led us to create Mithrl and the path we're on."
8:39
Rohan Ganesh, a partner at Obvious Ventures, laid out the bet like this, quote, "The AI-for-biology stack is getting crowded at the app layer, but the durable advantage lives in the middle, of the proprietary data, the world models, and custom infrastructure each biopharma team needs.
8:57
Everyone's requirements are different.
9:01
There's exactly where Mithrl is building."
9:04
So the claim is that Mithril isn't competing with the molecule designers.
9:09
Companies like Jira Therapeutics or Isomorphic Labs, which are designing new molecules and antibodies meant to hit a target, are on a completely different layer of the stack.
9:16
The closer comparisons right now are infrastructure plays like Argon AI, which raised a $5.5 million in a seed to tie together pharma's scattered commercial and clinical data.
9:29
And then there's Inherent, which isn't biology-specific at all.
9:33
Well, it claims Its research agent, Faraday, can beat Claude Opus 4.8 and GPT-5.5 at replicating published research, which is basically a bet that general scientific reasoning wins without a curated knowledge base underneath it.
9:45
The risk I'd keep an eye on is the frontier labs.
9:50
though.
9:50
Adarsh gives them credit for educating his market, and that's helping him right now.
9:53
But same labs are going after science very directly and very hard recently.
9:57
And if a general model gets good enough at reading the literature and keeping its facts straight, a curated world model starts to look less like a moat and more like a feature.
10:07
I don't think that happens quickly.
10:10
What Mithril really has, I think, is the provenance, the stuff sitting inside the customer's firewall, and the relationships with the scientists, and a new model release doesn't hand you any of that.
10:20
I'd also want to know how the business model scales.
10:22
Forward-deployed scientists inside every customer is great for adoption, and in pharma it's probably the only way in the door.
10:29
But it also means that, at least for now, this is part software and part services, and they're saying outright that the first use of this money is headcount to handle deployments.
10:37
With about 30 people and top-10 pharma customers, that can get stretched thin fast.
10:41
Palantir made the forward-deployed thing work eventually, but it took a long time.
10:45
I do think framing the value around patents is smart.
10:51
Pharma companies don't really pay for productivity so much as they pay for exclusivity, so if you can plausibly say your platform contributed to a customer's IP, that's a much easier budget conversation than tokens saved.
11:03
I just keep in mind that "contributed to" a filing is a loose standard, a filing isn't a granted patent yet, and a patent isn't a drug.
11:13
Still, for a company this young, being close enough to the science that customers credit you with anything is a good sign, and it tends to get you the next contract.
11:20
That's it for this episode.
11:23
Please like and subscribe.
11:23
If you are a startup coming out of stealth and you'd like to get your founding team on the show, I will interview you at about 45 minutes.
11:31
Get in touch at hello at bios .tech.
11:36
And if you'd like to get this in your ears every morning, check out the podcast, Startup Bios, wherever you get your podcasts.
11:44
[untranscribed]