Name: The science of delivering cures straight to your cells | Eric Kelsic
Uploaded: 2025-12-01T17:57:32.000Z
Duration: 15 min 46 s
Description: Thousands of YouTube videos with English-Chinese subtitles! Now you can learn to understand native speakers, expand your vocabulary, and improve your pronunciation...

A life that's full of good experiences, more time with family, with friends, more time to love.

But sometimes genetic illness can cut that short.

At some point, our body breaks down and our bodies are genetic machines.

For many diseases, the cause of the disease is a mutation in the genome.

Gene therapy is a vision that many have had for decades, more than 50 years.

The power of genetic technology is that once you get inside of cells with a DNA molecule, that molecule can stay there for the lifetime of the cell.

So it's the potential for a one time treatment for a disease where you wouldn't otherwise be able to reach the cells and solve for the root cause of the disease.

Today though, for the most part, the genome you're born with is the genome you die with.

Access to this molecular level is out of reach for almost all of this.

have really struggled to be able to get enough of the genetic payload into the cells where they're going to be effective as a therapeutic.

And it's getting inside of the cells that has really been a challenge for many, many years.

I'm Eric Kelsik, CEO and co founder at Diner Therapeutics.

For the past 10 years, I've been working to solve the grand challenge of gene delivery.

How are we going to make gene therapy a mainstream kind of medicine?

We need to solve these grand challenges like delivery, being able to deliver a therapeutic payload to every organ or every cell where there might be some benefit to patient health.

To do that, we're engineering protein shells derived from viruses.

Capsids are the protein shells of Adeno associated virus.

Adeno associated virus is a parasite of other viruses.

AAV naturally isn't known to cause any disease.

The reason why AV gets a lot of attention is because it's one of the smallest viruses and that enables it to get into many places all across the body where we need to deliver therapeutic DNA.

We still don't know a lot about how it functions naturally.

That said, we don't need to understand everything about how the virus works in order to adapt it as a therapeutic technology.

And that's our focus at Dino, engineering the capsid sequence to make capsids a better delivery vehicle for gene therapies.

What's amazing about capsids is they're evolved in nature to do so many different things so they can go through your body, through the blood, find a cell, enter the cell and then be released into the cell cytoplasm, get into the Nucleus through the nuclear pore, break open the capsid and release the genome.

So for a gene therapy going from the blood into cells, into the cytoplasm, into the nucleus, the genetic payload, that's entirely the goal.

And when therapeutic genes are expressed in the nucleus, they can be treating those cells for a patient's entire lifetime.

As a one time treatment, it can be an effective cure.

However, natural capsids, they're not efficient enough for most therapeutic purposes.

So for the past 28 years, protein engineers have been working to modify the capsid to make it better as a therapeutic protein, applying a technique called directed evolution.

Directed evolution is evolution, like occurs in nature, but for a goal that we choose.

And the most common approach there had been to randomly change the capsid sequence to make very large libraries, millions or even billions of different capsid sequence molecules.

a very large one, you have a chance of getting a good hit.

And the reason is because the capsid has many different functions.

And if you break even one of them, then and as a therapeutic, it's essentially useless.

Roughly 80% of the single changes that you could make to the capsid break the most essential function, which is the assembly and packaging of the genome.

What that means is that if by chance you make any mutation four out of five times, it's going to break the function.

And that's a problem for engineering because to get improved function, we're going to need to make multiple changes, maybe even hundreds of changes.

So if every time you make a change, the viability drops down, it gets really hard to have a library of changes that are going to give you a chance of finding an improved capsid.

You need to be able to produce and purify that capsid at scale.

It needs to be stable at low temperature or frozen.

But even when it's in your body, which is a relatively high temperature, it also needs to get into the right cells.

For example, there's a lot of unmet need for gene therapy in the brain because it's very difficult to get therapeutic proteins or other molecules across the blood brain barrier.

At a high dose, you might be able to get into 0.1% or maybe maybe a little bit more of the neurons in the brain.

That's not enough to treat many diseases.

And in addition to that, most of the capsid delivers its payload to the liver.

And at a high enough dose, that can also become Toxic.

We need to improve the efficiency of delivery to the target cell.

Over decades of trying this approach, we just didn't get enough improved variants or variants that were optimized for all the different functions that were needed to make them effective as gene therapies.

I had seen that there was a new wave of technologies coming with the potential to change the way that we engineer proteins completely.

It starts with this DNA multiplexing technology.

So we have an idea of an experiment we want to run testing many different capsids.

They might be designed to bind to a certain receptor, or they might be designed in a neighborhood of sequence space that before we found is promising.

We came up with a way of building very large libraries of capsids in which the sequence was programmed, meaning we had designed it on a computer, synthesized that DNA and then cloned it into the capsid so this could be injected in a few mills.

The best way we have to make a prediction about what's going to be safe and effective for humans is to do an animal experiment.

We do most of our screening in non human primates, especially in cynomologous monkeys.

That's one reason why we developed this technology, because their lives are also very precious.

We want to get as much information as we can from even one experiment.

In this case, we're measuring maybe 100 or 200,000, sometimes even a million different capsid sequences in that one animal.

We'll get all these tissues back from our animal experiments, and

then we want to learn as much as we can from that experiment.

Where did the capsids go or where did they not go?

Extract the nucleic acids, say, purify the DNA, purify the rna.

You can then work all the way back to what was the capsid sequence that this molecule corresponds to.

Is there more or less of that in the library?

And if there's more, that might mean that it was functionally improved for delivery.

If there's less, it might mean that there was a problem and it was broken.

At the end of today, we've got petabytes of data from the DNA sequencing.

I had always thought that proteins, they're too complex for us to understand as humans, certainly too complex for me to understand.

When you look at a string of 735 letters, it's really hard to notice all the differences.

But with all that data, what I could see, even myself, was there's a lot of patterns in that data, patterns about which amino acids work at each position.

My thought was that If I could recognize those patterns, and

the data set is so vast, there's probably a lot more information in them as well.

That's actually the perfect type of problem for a machine learning model.

We can use AI to automate the analysis of all that data and to find even more nuanced patterns to maximize the chances of success.

The expected value of finding an improved variant, we call that AI guided design.

But once we have that data and we've trained models on it, we can now query those models billions and billions of times.

So our ability to scale the computational work is even higher than the molecular side.

We can't possibly test everything in an experiment, but

with machine learning, we can test many different sequences in silico, meaning on a computer, and

the models will tell us which ones they think are better.

Tens or hundreds of different models each have a different insight.

And we compare the opinions of all those different experts to choose the ones that we're very confident are worth investing in as we bring them forward into the nest experiment.

So it's this iterative cycle of making libraries in DNA, measuring their properties, then building models to analyze and understand those properties, then querying the models to know where are the most promising regions of sequence space that we should go next, and

then going back to design a new library, turning that into DNA.

there's always human judgment at some point before we do another round of experiments.

I think that over time, what we want to do is put humans at an even higher point of leverage so that they're able to use their exceptional judgment and shift some of the more routine tasks to AI agents or even just simple scripts that run on the computer.

Being able to collaborate with AIs more effectively is where we'd like to go, so that we can, for example, give instructions to the AI to automate how we analyze the library or how we design it and get back the answers that we expect.

We want to get the results as fast as we can.