200 AI-Designed Drugs in Trials, None Approved Yet — Jun 8, 2026

0:00200 AI-designed drugs are now in human trials. Not one is approved. Welcome to MedAI Times Podcast, your daily update on medical AI. Don't forget to like and subscribe. So I want you to imagine a pharmaceutical waiting room.

0:13And it is just packed with 200 highly-touted, incredibly expensive VIPs. All waiting for the doctor. Right, exactly. Every single one has this impeccable resume. They're backed by billions of venture capital. But here's the catch.

0:28The door to the doctor's office has literally never opened for a single one of them. Not once. Yeah. So today, we're looking at those 200 VIPs. These are the AI-designed drugs that are currently stuck in clinical trials.

0:41And we really want to separate the theoretical promises from what's actually hitting hospital floors and regulatory desks right now. Because there's a lot of noise out there. A ton of noise. We are diving into a stack of industry briefings from today, June 8, 2026.

0:55And we've got you, the listener, squarely in mind as we filter this out. So let's just jump right into the quickfire headlines to set the agenda. Let's do it. All right. 200 AI-discovered drugs are in trials. Zero approved.

1:08The FDA cleared Clarius's bedside ejection fraction AI. GE Healthcare's auto-contouring tool is cleared for radiation planning. That's a big one. Yeah. Phillips and Wellspan signed a seven-year AI imaging alliance.

1:22The Joint Commission launched the first hospital AI certification. The WHO is weighing AI for health policy decisions. Anthropic is scaling its Mythos cyber program to 15 countries.

1:35Moving fast on that. Really fast. Perplexity lets its agent write its own search code. And OpenAI says, chat is dead. And they're pivoting to agents. I mean, that stack just spans the entire spectrum, right? You're going from navigating molecular chemistry in some server farm all the way up to writing global health policy in Geneva.

1:54It's wild. And Friday, we covered that algorithm that designed a coronavirus vaccine from scratch. Right. Which, I mean, that felt like staring straight into the future. But today, we're looking at the sober counterweight to that. Because designing a complex molecule on a computer screen and proving that it won't actually harm a human being in the real world, those are two vastly different things.

2:14Oh, completely. They require just entirely different types of proof. I think we should start by looking at this Biomed Nexus report on the AI drug pipeline. Yeah, let's get into the numbers. Because the industry has effectively, you know, they've solved the initial discovery phase.

2:29It's basically a volumetric now. We're looking at over 200 AI-originated drug candidates in clinical development. Yeah. And if you break that down, it's roughly 94 in phase 1 safety trials, about 56 in phase 2 efficacy trials, and 15 that have reached that massive final hurdle of phase 3.

2:49Right. But, I mean, the critical number hanging over all of that pipeline volume is zero. Exactly. Zero actual FDA approvals. Which is wild. Because when I read these vendor claims about generative chemistry, it's like, OK, the algorithms are clearly exceptional at navigating chemical space.

3:04Sure. They map millions of molecular variations to find one that perfectly binds to a specific disease protein. And that cuts down the trial and error time we used to spend in a physical lab by, like, years. Oh, absolutely. The speed is incredible.

3:17But optimizing for binding affinity in a digital simulation, that doesn't account for the chaotic biology of an actual human patient. I mean, it's like AI is incredible at drawing up the blueprints for a car, but we still haven't proven that the car can survive a crash test on the highway.

3:34That is a perfect analogy. You've isolated the exact mechanism where the delay happens. AI is highly proficient at solving the math problem of protein binding. It generates a molecule that fits a receptor, like a key in a lock.

3:47Right. But human biology is this deeply complex interacting system. So a molecule might fit the target lock perfectly, but on its way through the bloodstream, it might also bind to, I don't know, a completely unrelated enzyme in the kidney.

4:00And just cause cascading failures. Exactly. And predicting that kind of systemic biological chaos is something our current neural networks really struggle with, mostly because our training data on human biology is so incomplete compared to our data on chemical structures.

4:13OK, so this industry aggregation from Biomednexus, it's a broad snapshot, but it really lacks the granular, peer-reviewed clinical data for any specific drug's efficacy. Right. It's just industry numbers.

4:26So we're calling the verdict on this top story a giant waving caution flag. I mean, the pipeline volume is undeniable. The speed of discovery is real. But the ultimate referendum on this technology isn't some press release celebrating how fast a molecule was generated.

4:42No. Nobody cares about that if it doesn't work. Right. The only referendum that matters is whether the drug cures the patient without causing severe side effects. And we're getting close to a put-up or shut-up moment for the industry. The defining test is going to be these upcoming 2026 phase 3 readouts.

4:59Which ones are you watching specifically? We're closely watching the data for drugs like zosylcetinib. And we are waiting on further data for rentocere tib, which targets idiopathic pulmonary fibrosis, or IPF.

5:12And IPF is a brutal disease. It's a severe progressive lung scarring disease. And it's notoriously difficult to treat because we don't fully understand the underlying mechanism of the scarring. So if an AI-designed molecule can actually halt or reverse that kind of complex fibrotic pathway in a phase 3 trial, well, that validates the entire methodology.

5:32Wow. And if those succeed, when do analysts actually project the first FDA approvals? If the trial succeed, we're probably looking at 2027 or 2028 for the first actual approvals. OK, so while molecular biology takes years to prove itself, the software side is literally hitting hospital floors right now.

5:50Yeah. We're seeing these highly narrow incremental tools sliding directly into clinical workflows. They're designed to automate really tedious diagnostic boundaries, not to act as like a generalized artificial doctor.

6:04Yeah, let's look at the clarious ejection fraction AI. The FDA just cleared this on June 2nd. This is the ultrasound one, right? Exactly. It's a handheld ultrasound app. A clinician uses the probe at the bedside, and in about 90 seconds, the software returns an automated measurement of the patient's left ventricular ejection fraction.

6:21Right, and for clinicians in the ER or critical care, ejection fraction is huge. It's the percentage of blood pumped out of the left ventricle with each contraction. It's basically the dividing line when a patient crashes. Exactly. If someone comes in with undifferentiated dyspnea, meaning they can't breathe and you have no idea why you need to know immediately if their heart is feeling to pump or if they have like a pulmonary embolism in the lungs.

6:45You need that answer fast. But OK, here's my issue with the DAYC report on this clearance. There are zero published sensitivity or specificity numbers, none. The vendor just claims the AI was trained on, quote, thousands of annotated images.

7:01Well, the software maps the pixel gradients of the ultrasound image. It identifies the changing boundaries of the ventricle wall during a heartbeat and calculates the volume change mathematically. OK, but they're pitching it as a replacement for the visual estimate.

7:14Right, where an emergency doc without advanced echocardiogram training just eyeballs the bouncing heart on the screen and guesses if the pumping action looks normal. But an experienced physician's eyeball estimate is surprisingly accurate, especially in a chaotic trauma bay.

7:30I mean, if I'm an ER doctor, I am not accepting a black box pixel calculator over my own clinical judgment. Not without data. Right. Not unless the vendor gives me hard public performance data showing how often their algorithm miscalculates because of, say, poor image quality.

7:47The thing is that missing public data stems from how the regulatory landscape actually handles software. An FDA 510K clearance just means the agency agrees the device is substantially equivalent to something already on the market.

8:00Oh, wow. Yeah, or that it meets a baseline safety threshold. It doesn't force the company to publish a peer-reviewed academic paper detailing their false positive rates for the public. So the credibility signal is just the clearance itself.

8:13Pretty much. Vendors rely on the regulatory clearance. But clinicians integrating these tools into life-or-death triage pathways are totally justified in demanding the underlying performance metrics directly from the vendor. Yeah, you'd want to see the receipts before deploying that.

8:28And that brings up this fascinating tension about how much we trust software to define physical boundaries. Oh, definitely. We see this even more clearly in the second major clearance. This is GE Healthcare's MIM Contour Proto-JAI Plus 2.0.

8:42Say that five times fast. Right. It's a tool that auto-segments anatomy for radiotherapy planning. Right. And they just added specific models for the MR brain and the CT male pelvis. The clinical reality of radiation oncology requires a level of precision that is just difficult to overstate.

8:59Before they can fire a radiation beam into a patient's body, a dosimetrist or a physicist has to manually outline the tumor. And all the healthy organs nearby. Right, every single organ at risk. And they do this slice by slice on a CT or MR scan, which often means drawing hundreds of individual polygons.

9:17It can consume hours of manual labor per patient. Not to mention interclinician variability. Exactly. The exact boundary of a prostate gland might look slightly different depending on which expert is actually drawing the lines.

9:29So this deep learning model analyzes the image, predicts the anatomical structures, and just draws the lines automatically. It turns the clinician's job from manually drawing to just reviewing and editing the AI suggestions.

9:42Which is a huge time saver. It is. It's like an automated, incredibly high states coloring book. But keeping the radiation inside those specific lines is literally a matter of life and death for the healthy tissue. What's really interesting here is the regulatory maneuver GE used.

9:58They utilized a relatively new FDA framework called a Predetermined Change Control Plan, or PCCP. How does that work? Well, historically, if a manufacturer wanted to update an algorithm to recognize a new organ, they had to submit a completely fresh start to finish 510k application.

10:15Which takes forever. Right. But a PCCP shifts the regulatory focus from the final product to the engineering process. The FDA essentially approves the manufacturer's testing and validation methodology in advance. Oh, I see.

10:28So as long as GE follows that strict, pre-approved protocol for verifying the software's accuracy, they can roll out new anatomical regions, like these brain models, much faster. OK, a streamlined process is great for iteration. But the fundamental flaw of deep learning models and imaging is out of distribution data.

10:46The AI is trained on thousands of textbook examples of human anatomy. But what happens when it encounters, say, a post-surgical pelvis? Like, the prostate has already been removed, the bladder shifted into the empty space, and there's extensive scar tissue everywhere.

11:01The pixel gradients no longer match the statistical weights the network learned during training. Exactly. And then the algorithm can just fail silently. It'll draw a boundary that makes absolutely no anatomical sense.

11:13Which is why the human reviewer remains the sole fail-safe against those out-of-distribution failures. We have all these highly specific AI tools flooding into hospital workflows right now, accelerating tasks from triage to radiation planning.

11:28It's moving so fast. It is. This adoption curve is vastly outpacing our traditional oversight mechanisms. And that leads directly into this current scramble to establish institutional governance. Yeah, because someone has to oversee this. On June 1st, the Joint Commission, which is the major accreditation body for U.S. hospitals, they launched the Responsible Use of AI in Healthcare Certification, or RUAIH.

11:51It's a big step. It certifies health organizations across five domains. Governance, data management, risk and bias reduction, performance monitoring, and transparency and training. It really represents the first accreditation-grade AI program built specifically for the operational side of health systems.

12:10I mean, maybe. Yeah. But looking at the documentation, the structural weakness of this certification is glaring to me. Awesome. It's entirely voluntary. And it explicitly dates that the commission does not validate the individual commercial AI tools a hospital purchases.

12:27Well, they aren't auditing the code. Exactly. They aren't auditing the code of the GE contouring software or the Clarius ultrasound. They're just certifying that the hospital has a committee that talks about AI. Okay, that's a little harsh. Is it? If it doesn't validate the tools, isn't this just a shiny marketing sticker for a hospital to slap on its billboard?

12:44While the actual algorithms running on their servers remain completely unexamined by the accrediting body. Okay, you're right that it doesn't have the technical teeth to audit proprietary neural networks, but the value lies in establishing an external yardstick where literally none existed before.

13:01I guess that's fair. Prior to this, hospital chief technology officers were building ad hoc governance structures in a total vacuum. By forcing a hospital to formalize a data management protocol just to get the certification, it prevents them from treating AI vendors as infallible black boxes.

13:21It mandates a paper trail. It mandates institutional responsibility. But you're right that the ultimate pressure test doesn't happen on paper. It happens in the messy reality of community health systems. Like the Philips and WellSpan Health Alliance.

13:35They announced that on June 4th. That's a perfect example. It's a seven-year deal making Philips the preferred vendor across all 12 WellSpan hospitals. And they have a specific stated target of recovering more than 500,000 workforce hours a year.

13:48And that structure really matters. Testing an algorithm at a massive academic research center with pristine, perfectly formatted data is one thing. Right, but reality is different. Deploying it across a regional network of 12 community hospitals with varying legacy systems and completely different patient demographics, that is where the technology actually proves its worth.

14:08So WellSpan is basically functioning as a co-developer here. They are. They're gonna measure the AI against tangible operational metrics. Actual patient throughput, verifiable cost reduction, genuine workflow integration.

14:21Because projecting half a million saved hours is really easy math for a marketing brief. Oh yeah, but proving those hours are actually saved over a seven-year contract, that is the real validation. But if we scale that risk assessment up from a regional hospital network to the entire globe, it reveals a much darker potential.

14:38A WHO paper. Yeah. The WHO released a policy paper weighing the use of AI to synthesize evidence for health policy decisions. They map out the potential for rapid literature review, sure, but they heavily emphasize the risks of opacity, bias, and hallucination.

14:55Which are massive risks at that scale. Right. We know large language models confidently invent facts. If an LLM hallucinates a detail while summarizing a single patient's chart, the impact is severe, but it's isolated to one individual.

15:08But at the WHO level. Exactly. If an LLM hallucinates data or injects systematic bias while synthesizing evidence for a national ministry of health that's drafting global coverage policies, the error scales to millions of people simultaneously.

15:23The amplification of risk is exponential there. And the WHO is stepping in right now because the fundamental architecture of these AI systems is undergoing a substantive change. The shift from chatbots to agents.

15:35Exactly. We are moving away from passive chatbots that just retrieve text toward active agents that can execute complex autonomous tasks. And this agentic shift is basically the undercurrent of the entire tech ecosystem right now.

15:49Look at Anthropic. They just expanded their cloud mythos security initiative to roughly 150 organizations across more than 15 countries. And that's a cyber vulnerability program. Right. It's an AI driven program that actively hunts for cyber vulnerabilities.

16:04And Anthropic is specifically rolling this out to critical infrastructure operators. Which includes hospitals. It directly encompasses hospital networks and fleets of connected medical devices. They are deploying an autonomous AI agent to hunt for flaws in the exact networks that keep intensive care units online.

16:24And the technical leap from chatbot to agent is demonstrated perfectly by what Perplexity is currently doing. They just introduced this search as code feature. I saw that. It's fascinating. Historically, an AI system used a simple tool call to query a fixed search engine, right?

16:40It just pulled whatever results the engine handed back. But Perplexity's model is now writing its own executable Python code on the fly. Wait, to do what? To run, filter, sort, and deduplicate its own database searches.

16:52It bypasses the standard search engine entirely. It reportedly cuts their token costs by up to 85%. And it's outperforming open AI on internal benchmarks. Okay, so it's like we're moving from a digital encyclopedia that you have to read to a digital intern who actually runs the errands for you.

17:09A chatbot is the encyclopedia. You ask a query, it displays text, and you do the work of synthesizing it. What Perplexity has built is closer to a digital librarian who doesn't just hand you a book.

17:22They actually invent a custom Dewey decimal system in real time. Exactly. They write a script to index the entire library, filter out the unreliable authors, and deliver the exact synthesized conclusion you need.

17:35And open AI is aggressively pivoting in the same direction. They're stating publicly that the era of the chatbot is over and they're moving to bundle tools into this task-completing agent ecosystem. Because the mechanism driving all of this is Retrieval Augmented Generation, or RJ, healthcare IT departments are currently spending millions on clinical RJ systems just to search disparate electronic health records.

17:58Right, to find patient data. But the moment medical AI vendors adopt this agentic approach, those systems won't just pull up a static PDF of a patient's history. The clinical AI will write its own background routines to hunt down complex, multivariable patient histories across completely different hospital databases.

18:14Autonomously deciding what's relevant and what to ignore. Exactly. It makes the diagnostic tool vastly more powerful. But it means the AI is executing logical steps in the background that the physician cannot easily audit or reconstruct.

18:26Which loops directly back to the WHO's warning about opacity. We're deploying tools that write their own operational rules inside environments where human lives depend on understanding the methodology. And tying this all together requires looking at the financial scaffolding underneath the entire industry.

18:42The VC funding. Yeah. In the first quarter of 2026, digital health hit roughly $4 billion in venture capital funding. And according to J.T. Morgan, approximately 75% of those deals were specifically AI-focused.

18:56That is a staggering amount of money. The entire market is deeply leveraged on this foundational belief that AI will inevitably deliver superior clinical outcomes. So if we tie that $4 billion quarter back to our top story, you know, the 200 AI-designed drugs currently in trials with zero approvals, the stakes become incredibly clear.

19:16Right. Because if the phase three trials for molecules like Zasocitinib and Rentocitib fail to show efficacy this year, the clinical reality will violently decouple from the investment hype.

19:28Oh, wow. The market is banking on these algorithms understanding biology as well as they understand math. If they don't, we may be hurtling toward a massive systemic market correction by the end of the year. It's basically a multi-billion dollar house of cards waiting on a handful of clinical trial readouts.

19:44So since the drug discovery pipeline is the main through line dictating the future of this sector, we have a specific proposition for you. Do you want us to build a dedicated tracker of the specific 2026 and 2027 phase three readouts, logging the drug, the company, the disease indication and the expected date so we can mark each one as verified signal or expensive noise the moment the data lands?

20:05Thanks for listening. Find us on YouTube and your favorite podcast app. See you tomorrow.