Programming Life: AI, Medicine, and the Next Frontier of Human Health

📖 Read: https://helioxpodcast.substack.com

🎥 YouTube: https://www.youtube.com/channel/UCd5BbCEeC3Z6dp-nNjWRbBw
🎙️Available for Broadcast: https://exchange.prx.org/group_accounts/253118-heliox_where_evidence_meets_empathy 

For centuries, medicine worked by accident. Alexander Fleming came back from vacation to a contaminated petri dish and stumbled into the antibiotic age. In this episode of Heliox: Where Evidence Meets Empathy, we explore what happens when that era ends — and the age of programmable biology begins.

Drawing on a May 2026 Nature feature by Melanie Sr., we examine how generative AI is fundamentally restructuring drug discovery: not just sorting through existing molecules, but designing entirely new ones from scratch, engineered at atomic precision for specific biological targets.

In this episode:

🔬 Why NVIDIA, Google Ventures, and OpenAI are pouring billions into medicine — and what their "biology as information processing" thesis actually means

🔬 AlphaFold2, Levinthal's Paradox, and the Nobel Prize–winning solution to a 50-year structural biology puzzle

🔬 The lab-in-the-loop: how AI platforms tethered to robotic wet labs are compressing five-year discovery timelines to 17 months

🔬 Synthetic macrocycles using "alien" D-amino acids — molecules the human digestive system cannot break down, potentially enabling oral delivery of complex biologics

🔬 Genetic switches that activate only inside tumor cells, forcing cancer to manufacture its own immune signal

🔬 The Sec61 translocon: selectively blocking disease-causing proteins at the cellular tollbooth without shutting down normal cell function

🔬 The real bottlenecks: clinical trial infrastructure, manufacturing constraints, and why no de novo AI drug has yet cleared commercial approval

🔬 The horizon question: when does therapeutic medicine become directed human enhancement?

This is Heliox: Where Evidence Meets Empathy

Independent, moderated, timely, deep, gentle, clinical, global, and community conversations about things that matter.  Breathe Easy, we go deep and lightly surface the big ideas.

Transcript

Speaker 1: 0:25

This is Heliox, where evidence meets empathy. Independent, moderated, timely, deep, gentle, clinical, global, and community conversations about things that matter. Breathe easy, we go deep and lightly surface the big ideas. For centuries, if you really look at it, human medicine was essentially just this massive game of trial and error.

Speaker 2: 0:57

Oh, absolutely. Yeah.

Speaker 1: 0:57

I mean, you look at the historical record and it's largely a chronicle of foraging, right?

Speaker 2: 1:01

Right. Yeah. Like early healers just sort of chewing on willow bark to see if it would cure a headache.

Speaker 1: 1:06

Exactly. And, you know, eventually that led us to isolate salicylic acid, which is basically the precursor to aspirin. But it started with just chewing on bark.

Speaker 2: 1:15

Or you think about Alexander Fleming in 1928.

Speaker 1: 1:18

Oh, the penicillin story.

Speaker 2: 1:20

Yeah, literally returning to a disorganized lab, finding this rogue spore of mold on a culture plate, and just ushering in the era of antibiotics through sheer blind luck.

Speaker 1: 1:32

It's wild, because we carry this cultural perception of medical research as this highly calculated deterministic science, but structurally, it's been a massive, uncoordinated screening process.

Speaker 2: 1:44

We've basically just been rummaging through nature's chemical inventory.

Speaker 1: 1:46

Right. Throwing molecular shapes at human biology and just hoping one of them happens to bind to the right receptor.

Speaker 2: 1:53

Right. And, you know, hoping it does that without causing catastrophic side effects in the process.

Speaker 1: 1:58

Right. But the premise we are exploring today in this deep dive flips that entire model completely on its head.

Speaker 2: 2:05

It really does. It's a fundamental shift.

Speaker 1: 2:07

Because what if instead of discovering drugs through like high throughput screening or just sheer serendipity, we could actually program them?

Speaker 2: 2:15

Yeah, program them the exact same way a software engineer sits down, defines a computational problem, and writes a sequence of code to solve it.

Speaker 1: 2:22

That is the core of what we are looking at today. And to anchor this deep dive for you, we are pulling from a news feature published in Nature by Melanie Sr. in May of 2026.

Speaker 2: 2:33

The piece is titled Programming Biology. Next-gen AI firms raise billions to design better medicines.

Speaker 1: 2:40

And our objective for you today is to cut through all the residual noise of the AI hike cycle. Because, let's be real, there's a lot of hype.

Speaker 2: 2:47

There is. But the speculative phase is over. We are actually moving into the execution phase now.

Speaker 1: 2:53

Exactly. We are going to look at the tangible mechanics of how generative artificial intelligence is currently restructuring drug research. We're going to examine how neural networks are engineering completely synthetic molecules down to the atomic level.

Speaker 2: 3:08

And how we are using biological components to rewrite DNA without breaking it. Plus, we'll get into how engineered switches are being deployed to basically weaponize a disease against itself.

Speaker 1: 3:19

It is so fascinating. But first, to kind of contextualize the scale of what is happening right now, we need to look at the money.

Speaker 2: 3:26

Always follow the money, right?

Speaker 1: 3:28

Always. What does the capital allocation look like in the first quarter of 2026?

Speaker 2: 3:32

So just in Q1 of 2026, AI and machine learning drug discovery firms secured $1.8 billion in venture funding.

Speaker 1: 3:41

Wait, $1.8 billion in a single quarter?

Speaker 2: 3:43

Yeah. That single quarter represents roughly 25% of all biotech venture funding globally.

Speaker 1: 3:49

That is massive.

Speaker 2: 3:51

It is. And the entities underwriting this transition, they aren't just legacy pharmaceutical companies trying to modernize their pipelines.

Speaker 1: 3:56

Right. It's not just the usual suspects.

Speaker 2: 3:59

No, the capital is heavily concentrated from technology conglomerates. We're talking NVIDIA, Google Ventures, OpenAI.

Speaker 1: 4:07

And why are they pivoting so hard into biology?

Speaker 2: 4:10

Because their core thesis is that biology is fundamentally an information processing problem.

Speaker 1: 4:15

Okay, unpack that a bit. What do you mean by an information processing problem?

Speaker 2: 4:19

Well, if DNA is just a sequence of encoded instructions, and proteins are the three-dimensional mechanical outputs of those instructions.

Speaker 1: 4:27

OK, yeah.

Speaker 2: 4:28

Then the algorithms that these tech companies designed to parse, predict and generate complex data structures, they're uniquely suited to solve human health pathologies.

Speaker 1: 4:38

So they basically view the human body as the ultimate coding environment.

Speaker 2: 4:41

Exactly.

Speaker 1: 4:42

I want to drill down into the computational shift happening here because the phrase AI in medicine has been thrown around for what, over a decade now?

Speaker 2: 4:50

Well, easily a decade.

Speaker 1: 4:51

And usually it refers to an algorithm parsing massive data sets like electronic health records or existing chemical libraries just to flag correlations.

Speaker 2: 5:00

Right. The older generation of AI models deployed in biotech were largely extractive. They were analytical.

Speaker 1: 5:06

But the nature piece emphasizes a structural departure from those older models. We aren't just running sophisticated regressions on existing data anymore, right?

Speaker 2: 5:15

No, not at all. Think about the old models, like a highly efficient sorting mechanism. You would feed it a database of millions of known molecular structures, and it would filter them based on predictive models.

Speaker 1: 5:27

So it would just tell you, hey, out of these existing molecules, this one might be worth testing in a lab.

Speaker 2: 5:32

Exactly. But it was entirely constrained by the boundaries of known chemistry. Couldn't invent anything.

Speaker 1: 5:36

Right. And the paradigm shift we are seeing now relies on generative architectures.

Speaker 2: 5:40

Yes. Similar to how a large language model, you know, like ChatGPT, processes the statistical relationships between words to generate a completely novel paragraph.

Speaker 1: 5:51

So instead of words, it's using the building blocks of life.

Speaker 2: 5:54

Right. These biological foundation models learn the underlying biophysical and thermodynamic rules that govern molecular interactions. They treat amino acids as sequential tokens.

Speaker 1: 6:04

And they can generate entirely new protein sequences that fold into specific targeted 3D shapes, which is huge.

Speaker 2: 6:13

The major inflection point for this was AlphaFold2.

Speaker 1: 6:16

Oh, right. The one that won the 2024 Nobel Prize in chemistry.

Speaker 2: 6:20

Yep. AlphaFold2 demonstrated that a neural network could actually solve the protein folding problem.

Speaker 1: 6:25

And just for context, solving that folding problem was something that stumped structural biologists for half a century.

Speaker 2: 6:33

It's known as Leventhal's Paradox.

Speaker 1: 6:35

Right, which basically stated that a typical protein has so many possible degrees of freedom, so many potential ways it could fold, that it would take longer than the age of the universe to randomly sample all of them.

Speaker 2: 6:46

To find the correct stable shape, yeah.

Speaker 1: 6:48

Yet our bodies fold them in milliseconds.

Speaker 2: 6:50

It's incredible. And AlphaFold II didn't just calculate the physics to figure it out. It used evolutionary data, mapping how sequences co-evolved to predict that final 3D structure.

Speaker 1: 7:02

Okay. So once you have a model that can accurately map a flat sequence of amino acids to a 3D functional structure, what's the next step?

Speaker 2: 7:10

The next logical step in the computation is inversion.

Speaker 1: 7:13

Inversion.

Speaker 2: 7:13

Yeah. You provide the AI with a spatial constraint.

Speaker 1: 7:16

Yeah.

Speaker 2: 7:17

So you define the exact topology of the binding pocket on, say, a pathogenic virus.

Speaker 1: 7:22

Okay. So you show it the shape of the lock you want to pick.

Speaker 2: 7:25

Exactly. And you instruct the model to generate a novel sequence of amino acids that will thermodynamically favor folding into the precise 3D shape required to lock into that pocket.

Speaker 1: 7:38

So it generates the perfect key from scratch.

Speaker 2: 7:41

Yes. It is the transition from predictive modeling to generative design. And this capability is the foundation of what the industry now calls the lab-in-the-loop framework.

Speaker 1: 7:51

OK, I want to break down the mechanics of this lab in the loop thing because this really seems to be the engine powering all those billions in venture capital.

Speaker 2: 7:58

It absolutely is.

Speaker 1: 7:59

Traditional drug discovery to me feels like writing a movie script in total isolation.

Speaker 2: 8:03

That's a good way to look at it.

Speaker 1: 8:04

Like a pharmaceutical company that spends 10 years and a billion dollars writing the script, casting it, filming it without ever showing it to a test audience.

Speaker 2: 8:12

Right. And they just have to launch it into phase three clinical trials and just cross their fingers that the physiological audience actually responds to it.

Speaker 1: 8:19

Right. But the lab in the loop concept sounds much closer to a stand-up comedian refining a set.

Speaker 2: 8:25

Yeah, testing a joke on a live crowd.

Speaker 1: 8:27

Exactly. The comedian writes a premise, tests the joke, instantly reads the reaction, and if a specific punchline fails, they immediately tweak the wording for the next show. They are using rapid, real-world feedback.

Speaker 2: 8:40

That iterative feedback mechanism is exactly how the modern AI platforms are operating, but, you know, at a massive automated scale.

Speaker 1: 8:48

So how does that actually look in practice?

Speaker 2: 8:50

Well, the AI generates a hypothesis for a novel drug molecule in the computer in silico. It predicts how well it will bind, its solubility, all of that. Okay. But the crucial differentiator is that the digital model is physically tethered to an automated physical infrastructure.

Speaker 1: 9:06

So it's not just a simulation.

Speaker 2: 9:08

No. The digital sequence is instantly routed to a wet lab equipped with liquid handling, robotics, microfluidics, high throughput mass spectrometry.

Speaker 1: 9:17

So robots are actually building the molecule the AI just dreamed up.

Speaker 2: 9:21

Right. The novel molecule is synthesized in physical reality and then tested against the biological target to measure the actual real-world binding kinetics.

Speaker 1: 9:31

And when the physical molecule inevitably deviates from the computer's prediction because it's biology, something will always be slightly off, right?

Speaker 2: 9:39

Oh, inevitably. Maybe it binds slightly weaker or it aggregates unexpectedly.

Speaker 1: 9:44

Does that mean it's a failure?

Speaker 2: 9:45

No, that physical failure data isn't discarded. The discrepancy between the computer prediction and the physical reality is actually the most valuable data they generate.

Speaker 1: 9:54

So they feed the failure back into the system?

Speaker 2: 9:56

Instantly. That data is fed back into the neural network to update the model weights. The AI learns precisely where its understanding of the biophysics was flawed.

Speaker 1: 10:05

And then it tries again.

Speaker 2: 10:06

It refines the design and generates a new, optimized sequence. And this cycle happens continuously, compressing what used to be years of medicinal chemistry into literally weeks.

Speaker 1: 10:16

That is mind-blowing. But the Nature article takes this a step further. We aren't just talking about generating standard proteins here.

Speaker 2: 10:24

No, we are talking about engineering at a resolution that frankly sounds more like theoretical physics than pharmacology.

Speaker 1: 10:31

Yeah. The piece mentions Chai Discovery, which is backed by OpenAI, and they are explicitly targeting what they call atomic precision in antibody design.

Speaker 2: 10:40

Atomic precision, yeah.

Speaker 1: 10:41

And I have to push back on this premise a little bit because, you know, precision is one of the most overused marketing terms in biotech. It really is. When we talk about an antibody, we are talking about a massive protein complex. It is a biological behemoth. Tweaking a single carbon atom's position within a structure of that magnitude. I mean, that feels like trying to optimize the aerodynamics of a commercial airliner by moving one single rivet. Does subangstrom precision actually dictate clinical efficacy?

Speaker 2: 11:10

I get the skepticism, but it absolutely dictates efficacy when you are targeting highly dynamic, confirmationally flexible membrane proteins.

Speaker 1: 11:18

Okay, so shape-shifting targets.

Speaker 2: 11:20

Exactly. And that's what Chi Discovery is focusing on. They are deploying this precision against G-protein coupled receptors or GPCRs.

Speaker 1: 11:29

PCRs. Remind me what those do again. They are integral membrane proteins characterized by seven

Speaker 2: 11:34

transmembrane alpha helices. Basically, they sit right across the cell membrane and act as the

Speaker 1: 11:40

primary signal receivers for outside stimuli. So they regulate how the cell responds to its

Speaker 2: 11:46

environment. Right. They regulate almost every physiological response in the human body, which is why they are the target of roughly one third of all FDA approved drugs. Wait,

Speaker 1: 11:56

If we already have so many drugs targeting them, why is designing an antibody for them so incredibly difficult?

Speaker 2: 12:02

Because GPCRs are not static structures. Like you said, they are shape shifters. They are incredibly dynamic, constantly shifting between inactive and active states. Right. When a molecule binds to the outside of the receptor, it induces a physical shift in the structure, which opens up a pocket on the inside of the cell to activate a signal.

Speaker 1: 12:19

So it's like a mechanical switch.

Speaker 2: 12:20

Exactly. And traditional small molecule drugs often struggle to be highly selective among different GPCR subtypes because all those binding pockets look very similar.

Speaker 1: 12:32

So a drug might hit the right receptor, but also hit three others by mistake.

Speaker 2: 12:36

Right, causing spite effects. Now antibodies, being much larger, can theoretically cover a wider surface area and achieve exquisite selectivity.

Speaker 1: 12:45

Because they latch onto a bigger, more unique footprint.

Speaker 2: 12:48

Exactly. But generating an antibody that can precisely bind to the outside loops of a GPCR, and doing so in a way that stabilizes a very specific shape of that shifting receptor, that requires an incredibly nuanced understanding of the energy landscape.

Speaker 1: 13:02

So if the target is constantly shifting its shape, the atomic precision of the AI model is necessary to predict those transient intermediate states.

Speaker 2: 13:10

Yes, to design an antibody that acts as a perfect wedge to lock it into the desired position.

Speaker 1: 13:15

Okay, that makes sense. The AI is modeling the exact push and pull of the atoms.

Speaker 2: 13:20

Right. It's modeling the Van der Waals forces, the hydrogen bonding networks, the electrostatic interactions, down to the Angstrom level.

Speaker 1: 13:26

And how successful are they? Because computational design used to struggle just to make small, rigid, mini proteins.

Speaker 2: 13:33

Well, today, Chai claims that one in five of the complex, full-length antibodies designed by their AI demonstrates successful target engagement in the physical lab.

Speaker 1: 13:44

One in five?

Speaker 2: 13:45

Yeah. In the context of antibody discovery, a 20% hit rate for entirely de novo-generated structures against GPCRs is an unprecedented acceleration.

Speaker 1: 13:55

Okay, that is impressive. But even with an unprecedented hit rate, we still run into the fundamental limitations of biologics, right?

Speaker 2: 14:02

You mean the delivery mechanism.

Speaker 1: 14:03

Yeah. An atomically precise antibody is still a massive protein. It cannot survive the proteolytic enzymes and the highly acidic environment of the human stomach.

Speaker 2: 14:12

No, it would get chewed up immediately.

Speaker 1: 14:14

Exactly. So you can't put it in a pill. It requires cold chain logistics. And you, the patient, usually have to travel to a clinic to get it via an IV drip.

Speaker 2: 14:23

The manufacturing overhead is staggering because you have to grow these proteins in massive bioreactors.

Speaker 1: 14:28

Right. So if the AI revolution just gives us slightly more specific IV drips, it doesn't really fundamentally democratize access to the therapy, does it?

Speaker 2: 14:36

And that is exactly the hurdle that Orbis Medicines, a firm based in Denmark, is attempting to bypass.

Speaker 1: 14:42

OK, what are they doing?

Speaker 2: 14:44

They are utilizing their AI infrastructure to design molecules that occupy the chemical space right between a traditional small molecule drug and a massive biological antibody.

Speaker 1: 14:56

So a middle ground.

Speaker 2: 14:57

Yes. They are engineering synthetic macrocycles.

Speaker 1: 15:00

I want to unpack the biochemistry of a synthetic macrocycle because to the listener, it sounds a bit like an oxymoron.

Speaker 2: 15:07

How so?

Speaker 1: 15:08

Well, we are trying to get the sniper-like specificity of a giant antibody, but we want to package it into something small and stable enough to survive the stomach acid. How does that work?

Speaker 2: 15:18

So a macrocycle is essentially a cyclic peptide. Normally, a string of amino acids is linear, like a piece of string with two free ends. Okay. But in a macrocycle, those ends are covalently linked together to form a ring structure.

Speaker 1: 15:30

So it's a circle.

Speaker 2: 15:31

Exactly. And this cyclic architecture severely restricts the flexibility of the molecule. Because it's locked into a rigid structure, it has a much higher binding affinity compared to a linear peptide.

Speaker 1: 15:43

And I imagine the ring structure also protects it.

Speaker 2: 15:45

Yes. It makes it significantly more resistant to exopeptidases, which are the enzymes in your body that degrade proteins by chewing them up from the ends.

Speaker 1: 15:54

Ah, because there are no ends to chew on.

Speaker 2: 15:56

Precisely. So the circular shape itself provides a massive stability upgrade. It mimics the binding footprint of a large biologic, but its total size is vastly smaller.

Speaker 1: 16:07

That is brilliant. So it can survive the digestive tract.

Speaker 2: 16:11

Theoretically, yes, allowing it to be administered as an oral pill. But, and this is where it gets really crazy, the structural geometry is only half of the innovation.

Speaker 1: 16:19

That's the other half.

Speaker 2: 16:20

The composition. They are not constrained by natural biology.

Speaker 1: 16:23

Wait, what does that mean?

Speaker 2: 16:24

Well, every protein in the human body is synthesized using a standard genetic code that translates into 20 canonical amino acids. But Orbis' AI models are generating designs utilizing thousands of unnatural amino acids.

Speaker 1: 16:37

Okay, this is where the underlying chemistry gets fascinating. By unnatural, we mean chemical building blocks that literally do not exist in the human body.

Speaker 2: 16:45

We are talking about expanding the chemical alphabet. Natural biology utilizes L-amino acids. That refers to the left-handed chirality or the specific 3D orientation of the molecules.

Speaker 1: 16:57

Like left-handed versus right-handed gloves.

Speaker 2: 16:59

Exactly. And the digestive enzymes in your stomach have evolved over millions of years to perfectly recognize and break down those natural left-handed amino acids.

Speaker 1: 17:09

Okay, so if the AI uses unnatural ones?

Speaker 2: 17:12

By incorporating de-amino acids, the right-handed mirror images or amino acids, with entirely synthetic side chains, the AI is creating a molecule with a geometry that is fundamentally alien to human biology.

Speaker 1: 17:25

Which means the natural enzymes in the stomach look at this synthetic macrocycle, and the lock and key mechanism completely fails.

Speaker 2: 17:31

Yep. The enzyme doesn't recognize the alien chemical bonds, so it can't break the drug down.

Speaker 1: 17:35

So the pill remains intact.

Speaker 2: 17:37

It remains intact, extending its half-life and ensuring it actually gets into your system. And the AI is doing the heavy lifting here, selecting from thousands of these unnatural blocks, predicting how they will fold into a ring, and ensuring that ring perfectly fits the target disease.

Speaker 1: 17:54

We are using supercomputers to design chemical solutions that the evolutionary timeline of Earth simply never produced.

Speaker 2: 18:00

It's incredible.

Speaker 1: 18:01

But, and there is always a but, the dependency here is data.

Speaker 2: 18:05

Always data.

Speaker 1: 18:06

If the AI is going to invent novel macrocycles using an expanded alien chemical alphabet, it needs an unfathomable understanding of the foundational rules of chemistry. Right. And the Nature article highlights a severe vulnerability in this entire field. If you train a language model on low-quality text, it generates nonsense.

Speaker 2: 18:24

Garbage in, garbage out.

Speaker 1: 18:26

Exactly. And human biological data is limited, highly siloed across different servers, and inconsistently formatted. Aren't we feeding these supercomputers a fractured data set?

Speaker 2: 18:36

That data bottleneck is basically the rate-limiting step in computational biology right now. If a model is trained only on human data, its generative capabilities will be severely biased toward known human biology.

Speaker 1: 18:49

So how do you fix that?

Speaker 2: 18:51

To build a model capable of true biological programming, you have to look beyond human pathology. You have to capture the entire logic of life. And to do that, a company called Basecamp Research is conducting global metagenomic sequencing.

Speaker 1: 19:05

Okay, the scale of the unknown biological landscape they talk about in the piece is almost paralyzing.

Speaker 2: 19:10

Yeah, the CTO of Basecamp is quoted saying that our current understanding of life on Earth is equivalent to five drops of water compared to the Atlantic Ocean.

Speaker 1: 19:19

Five drops of water. That is a sobering assessment of our biochemical ignorance.

Speaker 2: 19:23

It really is. The vast majority of life remains unsequenced. Basecamp's thesis is that the rules of biology, how proteins fold under extreme stress, how viruses bypass defenses, they have already been rigorously beta tested by evolution over billions of years.

Speaker 1: 19:39

So they are going out into the wild to find these beta tests.

Speaker 2: 19:42

Exactly. They're isolating the DNA of extremophiles. Microbes living in deep-sea hydrothermal vents or highly acidic hot springs.

Speaker 1: 19:50

Places where human proteins would instantly melt.

Speaker 2: 19:53

Right. They have compiled this trillion gene atlas, processing the sequence data from over a million newly discovered microbial species.

Speaker 1: 20:00

Okay, gathering the DNA of an extremophile living in a boiling deep-sea vent is a remarkable logistical feat. But how does that translate into a drug for a human patient?

Speaker 2: 20:11

The immediate high-impact application of this massive data set is the identification of large serene recombinases, or LSRs.

Speaker 1: 20:19

Okay, we need to unpack LSRs carefully because the entire biotech industry has been obsessing over CRISPR for a decade.

Speaker 2: 20:25

Oh, CRISPR has been everything.

Speaker 1: 20:27

Right. The prevailing narrative is that CRISPR is the ultimate gene editing tool. So why are AI companies suddenly pivoting to search the deep ocean for LSRs?

Speaker 2: 20:36

Because CRISPR, despite its revolutionary impact, is fundamentally a destructive tool.

Speaker 1: 20:41

Destructive?

Speaker 2: 20:41

Yeah. The Cas9 enzyme locates a specific sequence in the genome and induces a double-strand break. It physically severs both strands of the DNA double helix.

Speaker 1: 20:50

And the cell does not like that.

Speaker 2: 20:52

Not at all. The cell perceives a double-strand break as catastrophic damage and panics. It immediately initiates error-prone repair mechanisms, just haphazardly jamming the broken ends back together.

Speaker 1: 21:04

Which causes random mutations.

Speaker 2: 21:06

Exactly. So if you are trying to use CRISPR to insert a large, healthy therapeutic gene, you are basically causing localized genetic trauma and just hoping the cell repairs it using the template you provided.

Speaker 1: 21:20

It's like trying to fix a typo in a manuscript by setting off a small explosive charge on the page and then hoping you can tape the new paragraph over the blast radius before the paper degrades.

Speaker 2: 21:30

That is a very visceral analogy, but it's medically accurate. It causes immense cellular stress.

Speaker 1: 21:35

Which makes regulatory agencies nervous.

Speaker 2: 21:37

Very. And this is why LSRs are the focal point now. Large serene recombinases are enzymes naturally used by viruses that infect bacteria. When one of these viruses needs to integrate its DNA into the host, it doesn't cause a double-strand break. The LSR creates a transient covalent bond, nicks the strands, rotates them, swaps the genetic material, and then seals the strands back together.

Speaker 1: 22:01

It's a seamless integration. It acts like an insert cursor on a word processor.

Speaker 2: 22:05

Exactly. The DNA strands are never fully dropped or exposed as a break, meaning the cell's panic-induced damage response is never triggered.

Speaker 1: 22:12

And you can insert massive amounts of code this way.

Speaker 2: 22:15

You can theoretically drop an entire multi-gene pathway into the human genome with high fidelity and zero double-strand breaks.

Speaker 1: 22:22

That is incredible, but what's the catch?

Speaker 2: 22:24

The engineering challenge is targeting. Natural LSRs evolved to recognize specific landing sites in bacterial genomes. If you put a wild microbes LSR into a human cell, it won't know where to land.

Speaker 1: 22:36

Or it might land randomly and disrupt a tumor suppressor gene and cause cancer.

Speaker 2: 22:40

We need to program it to recognize a human DNA sequence.

Speaker 1: 22:43

Which brings us right back to the AI data race.

Speaker 2: 22:46

Yep. And it is a fascinating three-way race to solve this. Each company is feeding their neural networks fundamentally different types of data.

Speaker 1: 22:54

Right. So we have Basecamp Research relying on their trillion gene atlas from the extremophiles.

Speaker 2: 22:59

Right. They are banking on the sheer evolutionary volume of the natural world. Then you have ProFluent, who just announced a major partnership with Eli Lilly. And what's their approach? Purely computational. They are using a database of over 3 billion known protein sequences and betting that if the AI processes enough data, it will infer the grammar of protein DNA interactions without needing to go to a volcano.

Speaker 1: 23:23

And the third competitor?

Speaker 2: 23:24

A German biotech called Seamless Therapeutics. They're basically ignoring the massive AI data sets entirely.

Speaker 1: 23:31

Really? What are they doing instead?

Speaker 2: 23:33

They're relying on empirical manual lab data. They physically mutate a recombinase, test how it alters binding, and feed that hyperspecific cause and effect data into their models.

Speaker 1: 23:43

So it's a stress test of modern AI methodologies. Do you train the model on infinite nature, billions of sequences of text, or highly controlled lab data?

Speaker 2: 23:51

Exactly. And whoever wins that race won't just own a novel drug. they will own the foundational operating system for writing human biology.

Speaker 1: 23:59

That is wild, because if we cross that threshold, if we can seamlessly write large blocks of code into the genome, the application space just explodes.

Speaker 2: 24:08

Oh, completely.

Speaker 1: 24:09

We transition from merely replacing defective genes to actively programming the genome to execute complex logic, like turning a cell into a biological sensor.

Speaker 2: 24:20

And the most aggressive application of this programmable logic is in oncology, cancer.

Speaker 1: 24:25

Right. Because cancer is fundamentally a disease of corrupted logic. The genome has accumulated mutations that dysregulate its growth pathways.

Speaker 2: 24:34

And a company called Erlia is attempting to leverage that specific dysregulation by using AI to design synthetic genetic switches.

Speaker 1: 24:41

Let's break down this concept of a genetic switch, because the central challenge of treating cancer is toxicity, right?

Speaker 2: 24:47

Absolutely. Because cancer cells originate from our own tissue. They don't have a distinct bacterial wall for the immune system to easily recognize.

Speaker 1: 24:54

So traditional chemotherapy is essentially just systemic poison. You administer a toxic compound and rely on the fact that the fast-dividing cancer cells will absorb it and die slightly before the healthy cells do.

Speaker 2: 25:06

It's a brutal race.

Speaker 1: 25:08

Yeah.

Speaker 2: 25:08

But Early's approach is attempting to completely eliminate that systemic toxicity.

Speaker 3: 25:13

How?

Speaker 2: 25:13

By mapping the transcription factors of malignancies. Their AI identified that regardless of the originating mutation, cancer cells almost universally exhibit hyperactive transcription factors.

Speaker 1: 25:25

And transcription factors are the proteins that bind to DNA to control the rate of transcription, right? Like the regulatory volume knobs of the genome.

Speaker 2: 25:32

Great way to put it. In a healthy cell, these knobs are tightly regulated. In a cancer cell, they are cranked up to 11, continuously driving the cell to divide.

Speaker 1: 25:40

Okay, so what does Early do with that information?

Speaker 2: 25:43

Early's AI designed synthetic promoters, basically specific sequences of DNA that act as a docking site. But these are heavily biased. They are engineered to only be activated by those specific hyperactive transcription factors found in a tumor.

Speaker 1: 25:57

So they attach this synthetic promoter to a payload gene, creating the switch. And they deliver it systemically throughout the patient's body.

Speaker 2: 26:05

Right. But the elegance is the conditional activation. When it enters a healthy cell, the transcription factors are normal. They don't recognize the synthetic promoter. The switch stays off. The healthy cell is completely unaffected.

Speaker 1: 26:18

But when it enters a cancer cell...

Speaker 2: 26:20

The environment is saturated with those hyperactive factors. They blindly bind to the synthetic promoter, flip the switch to the on position, and force the cancer cell to transcribe the payload gene.

Speaker 1: 26:31

And what is the payload?

Speaker 2: 26:33

It's a cytokine called interleukin-12, or IL-12.

Speaker 1: 26:36

Oh, wow. Okay, for the listener, systemic administration of IL-12 has been the white

Speaker 2: 26:41

whale of immunotherapy for decades. It is incredibly potent. It essentially screams at the immune system to attack the area. But you can't just inject it into a patient's

Speaker 1: 26:49

bloodstream because it causes a massive cytokine storm. It's too toxic.

Speaker 2: 26:53

Right, which is why there are no FDA-approved systemic IL-12 therapies. But Early's genetic switch changes the game. The IL-12 is not in the bloodstream. It is only synthesized internally by the cancer cells that flip the switch. It's a molecular booby trap.

Speaker 1: 27:09

The tumor is hijacked by its own dysregulated machinery, forced to manufacture a localized

Speaker 2: 27:15

concentration of IL-12. It essentially flags itself for destruction, recruiting T-cells that obliterate the tumor without exposing the rest of the body. That is profound. It's like a burglar

Speaker 1: 27:26

picking a lock, and the exact motion of them picking the lock is what automatically handcuffs

Speaker 2: 27:31

them. I love that analogy. And in their preclinical models, they have demonstrated robust tumor regression. They are pushing for human clinical trials slated for 2027. Wow. But while Early is

Speaker 1: 27:43

programming the cell to manufacture a weapon, another company profiled in the NaturePiece is

Speaker 2: 27:47

taking the exact opposite approach, right? Yes, Anodia Therapeutics. They are using machine learning to shut down the manufacturing process of the disease entirely. And they are targeting the endoplasmic reticulum. Right. Specifically, a protein channel known as the Sec61 translocon.

Speaker 1: 28:03

The Sec61 channel. This is basically the central toll booth for roughly one-third of the entire human proteome. Over 4,000 different proteins must pass through this doorway to mature.

Speaker 2: 28:15

Exactly. And you can't just plug the channel entirely, because if you do, you induce catastrophic cellular stress and the cell dies.

Speaker 1: 28:22

So how do you stop just the bad proteins?

Speaker 2: 28:25

Well, Enodia recognized that the proteins passing through aren't identical. They all have a signal peptide, a specific sequence of amino acids that acts like a molecular barcode.

Speaker 1: 28:34

A VIP pass that tells the toll booth to open up.

Speaker 2: 28:37

Right. And Enodia drew inspiration from a really nasty pathogen. Mycobacterium ulcerans, the bacteria responsible for a flesh-eating disease.

Speaker 1: 28:45

Oh, lovely.

Speaker 2: 28:46

I know, but it's present biology. To evade the immune system, this bacteria secretes a toxin called mycolactone. This toxin binds to the Sec61 channel and acts as a highly selective bouncer. It reads the barcodes and denies entry only to the host's immune signaling proteins while letting normal housekeeping proteins pass.

Speaker 1: 29:03

So the bacteria naturally evolved a highly selective Sec61 inhibitor.

Speaker 2: 29:08

Exactly. And Anodia use machine learning to decode how that works. Their AI is generating small molecules that mimic the mechanism of that toxin, but are programmed to block entirely different proteins.

Speaker 1: 29:20

So if an oncogene is producing a protein that drives tumor metastasis, Anodia's AI designs a molecule that binds to SEC-61 and selectively blocks only that specific metastatic protein based on its barcode.

Speaker 2: 29:34

The disease-causing protein is trapped, fails to mature, and is rapidly degraded by the cell. It neutralizes the threat before it's even fully constructed.

Speaker 1: 29:43

We have covered so much ground here. AI designing atomic precision antibodies for shape-shifting targets, Alien macro cycles surviving stomach acid, extremophile recombinases editing genes seamlessly, synthetic promoters triggering localized cytokine storms. It's staggering.

Speaker 2: 29:58

It really feels like science fiction, but it's happening.

Speaker 1: 30:00

But as we synthesize this material, we have to impose a rigorous reality check here.

Speaker 2: 30:04

We do.

Speaker 1: 30:05

The theoretical output of these neural networks is phenomenal. But if I am a patient navigating a complex chronic illness today, I need to know the actual translation timeline. Because despite all these billions of dollars, there is currently no purely AI discovered drug available on the open market.

Speaker 2: 30:21

That is a fair critique. While a fully de novo AI drug hasn't cleared commercial regulation yet, the impact of these models is already massively compressing the research timelines.

Speaker 1: 30:32

Give me an example.

Speaker 2: 30:33

Recursion pharmaceuticals, they heavily leverage AI for morphological profiling. The traditional discovery phase usually takes four to six years. Recursion compressed that, taking a project from target identification to an advanced candidate in just 17 months.

Speaker 1: 30:47

17 months down from five years. That is a massive acceleration.

Speaker 2: 30:51

And we are seeing AI used to optimize existing drugs, too. Generate biomedicines, apply their platform to Tespire, an existing severe asthma drug.

Speaker 1: 30:58

Tespire is usually given via injection every four weeks, right?

Speaker 2: 31:02

Generate use their AI to engineer specific amino acid substitutions, massively increasing how tightly the drug binds to its target.

Speaker 1: 31:10

And they also tweaked how the body recycles it, extending its half-life.

Speaker 2: 31:14

Exactly. By using AI to optimize those kinetics, they created a vastly more potent version of the drug, accelerating it into phase three trials in just four years. And they're looking at a dosing regimen of just twice a year instead of monthly.

Speaker 1: 31:29

Going from 12 injections a year to two is a profound reduction in patient burden.

Speaker 2: 31:34

It really is.

Speaker 1: 31:35

But this brings me to a structural critique of the entire industry. The CEO of InSilico Medicine had a really sobering quote in The Nature Piece.

Speaker 2: 31:43

Oh, about moving with the traffic.

Speaker 1: 31:44

Yes. He stated that AI can drastically accelerate the discovery phase, but after that, you're moving with the traffic.

Speaker 2: 31:51

It's true.

Speaker 1: 31:52

We can compress five years of discovery into 17 months, but we are essentially building a high-speed bullet train that just slams into a massive, unmoving bottleneck human clinical trials.

Speaker 2: 32:03

The clinical trial infrastructure is the immovable object. You can use a supercomputer to engineer a perfect antibody, but a computer cannot yet simulate the systemic complexity of a human being.

Speaker 1: 32:13

It can't predict if a drug is going to randomly cause liver toxicity in a specific subpopulation.

Speaker 2: 32:19

Exactly. You still have to do phase one safety trials, advance to phase two, and conduct massive phase three trials with thousands of patients over years. The AI gets us to the starting line of clinical trials faster, but it doesn't change the temporal reality of the FDA approval process.

Speaker 1: 32:36

And the bottleneck isn't solely clinical, is it? There's a massive engineering challenge that I feel like these digital-first AI companies often underestimate.

Speaker 2: 32:45

The analog hurdle.

Speaker 1: 32:46

Right, the physical manufacturing of the drug. Chemistry, manufacturing, and controls, or CMC.

Speaker 2: 32:51

It is the harsh reality of synthetic chemistry. An AI can dream up the most elegant, hyper-efficient synthetic macrocycle. But if physically building that molecule in a lab requires 30 discrete chemical steps using rare catalysts and yields almost nothing?

Speaker 1: 33:06

It's commercially useless.

Speaker 2: 33:07

Completely.

Speaker 1: 33:08

Yeah.

Speaker 2: 33:08

If you can't mass produce it, the drug will never reach a patient.

Speaker 1: 33:11

It's like using a supercomputer to design a hyper-efficient aerodynamic vehicle. But the factory only has cast iron and standard rivets. The physical supply chain dictates the reality.

Speaker 2: 33:24

Which is exactly why we are seeing all these massive partnerships. The AI startups have the algorithms, but they lack the physical infrastructure.

Speaker 1: 33:32

So they partner with the legacy giants. ProFluent partnering with Eli Lilly, recursion with Sanofi.

Speaker 2: 33:39

The AI firms act as the digital discovery engine, and the big pharma conglomerates provide the massive chemical synthesis factories and the decades of regulatory experience to navigate the FDA.

Speaker 1: 33:49

To crystallize everything we have analyzed today, generative AI and biology has fundamentally transitioned from just sorting data into an active engineering platform.

Speaker 2: 33:58

We have completely shifted the paradigm from the serendipitous discovery of naturally occurring compounds to the deterministic programming of biology.

Speaker 1: 34:06

The mechanics are solidly in place, even with the clinical and manufacturing bottlenecks. But as we wrap up, we have to consider the ultimate logical endpoint here.

Speaker 2: 34:14

Right. And this is where it gets incredibly profound.

Speaker 1: 34:17

Because throughout this entire analysis, we have framed these AI models entirely within the context of pathology. Fixing broken systems, curing cancer, resolving inflammation.

Speaker 2: 34:28

But the underlying capability extends far beyond just repair.

Speaker 1: 34:32

Right. If a neural network can generate a synthetic macrocycle using alien D-amino acids that is far more stable and efficient than any naturally occurring human enzyme, we are confronted with a massive biological question.

Speaker 2: 34:46

Because the trajectory implies that we are not just matching human biology, we are computationally exceeding its natural limits.

Speaker 1: 34:52

Exactly.

Speaker 2: 34:53

Evolution is a slow, reactive process. It optimizes for survival, not for physiological perfection. If we master the grammar of molecular interactions and we are no longer constrained by the 20 natural amino acids, the application of AI will inevitably shift.

Speaker 1: 35:08

The horizon is not simply about returning a sick patient to a baseline state of health.

Speaker 2: 35:12

No. If we can write sequences that fold into physical structures capable of superior metabolic efficiency, enhanced cellular longevity, or total resistance to viral integration, the definition of what constitutes an optimal human biological system will be fundamentally rewritten.

Speaker 1: 35:31

The transition from therapeutic medicine to directed programmable evolution is simply a matter of computational scale at this point.

Speaker 2: 35:39

It really is.

Speaker 1: 35:40

The concept that the molecular tools we are designing to cure disease today will become the exact mechanisms used to systematically rewrite our own biological limitations tomorrow. It's a striking realization.

Speaker 2: 35:51

From relying on a random fungal spore on a petri dish to engineering novel biology at the sub-Angstrom level.

Speaker 1: 35:58

It's quite a leap. Well, thank you for joining us on this deep dive. Keep questioning the mechanics and continue exploring the complex, programmable architecture of the world around you.

Speaker 2: 36:10

heliox is produced by michelle bruecher and scott bleakley it features reviews of emerging research and ideas from leading thinkers curated under their creative direction with ai assistance for voice imagery and composition systemic voices and illustrative images of people are representative tools not depictions of specific individuals thanks for listening today four recurring narratives underlie every episode. Boundary dissolution, adaptive complexity, embodied knowledge, and quantum-like uncertainty. These aren't just philosophical musings, but frameworks for understanding our modern world. We hope you continue exploring our other episodes,

Speaker 3: 36:51

responding to the content, and checking out our related articles at helioxpodcast.substack.com.