This might genuinely be one of the biggest breakthroughs in medicine and drug discovery we've seen in decades. A new paper just dropped and it's about
this new AI model called Mammal. And this can completely transform biosciences and medicine. It could potentially come up with a ton of new cures for
diseases like cancer and this can make drug discovery way faster and cheaper and more accurate.
This could also lead to personalized medicine and beyond. Now this paper is super technical and jam- packed with details. So in this video I'm going
to break it down into simple terms so it's easy for anyone to understand. Let's jump right in. Now to understand why this paper is such a big deal,
it's helpful to first start with the problems of medicine and drug design today. Imagine spending 10
years and a billion dollars building something and there's a 90% chance it just doesn't work. 90% That's the failure rate of new drugs. From the
moment scientists come up with a promising new medicine to the moment it goes through clinical trials, roughly 90% don't make it through approval.
That's kind of insane when you think about it. We've already sequenced the human genome, we have
smartphones, we have AI that can predict protein structures or even understand DNA. We have so much amazing technology, yet when it comes to actually
making new medicine, we're still wrong most of the time. Think about this. If I built bridges and nine out of 10 of them collapsed, that would be
horrible. I would be out of business. So why is modern medicine, even with all our insane modern
technology, still hitting this wall? To understand this, we first need to understand some basic biology and how drug design works. You see, every
living thing has DNA. This is the code for living things to function and stay alive. And within DNA, there are sections called genes. These are
basically instruction manuals. They contain the code for making proteins. And proteins are the tiny machines
that actually do most of the work inside your body. They send signals, build structures, speed up chemical reactions, control immune responses, keep
cells alive, and a ton of other stuff. But, genes aren't always active. Some genes can be turned up, some turned down, or mutated. Scientists call
this gene expression. Basically, a measurement of which genes are active and how active they are. And
this controls how many proteins are made. So, we have DNA which contains genes, which code up proteins, which are essential for the function of all
living things. I'm oversimplifying here, but that's essentially the mini lesson on how life works. Now, there are a few possible ways for this pathway
to go wrong, which could cause disease. The DNA might get mutations, which would mess things up. This
could cause proteins to be made incorrectly or not at all. Or the gene expression might be too high or too low, which also affects how proteins are
made. Even subtle errors in this pathway could cause some catastrophic diseases that you and I are familiar with, like cancer. In fact, in really
simple terms, how cancer works is there might be a mutation that damages a gene in the DNA, which controls
for cell division. This changes gene expression or protein function. And subsequently, these proteins then tell the cell to grow, to ignore stop
signals, and spread into a tumor. That's essentially cancer. Now, here's how drugs are currently designed around this process. Scientists usually
start by trying to find the specific part of the disease pathway that has gone wrong. In other words, they
look for the bad guy. Maybe there's a protein telling cancer cells to keep dividing. Maybe there's an enzyme helping a virus copy itself. Anyways,
once the researchers find that target, the goal is to design a drug that can interact with it. Think of the target like a lock and the drug like a
key. If the key fits within the lock, it can either block the protein or change what it does. Now, there
are two main types of drugs that are relevant to this paper. Some drugs are small molecules like Tylenol. They're tiny, stable, and often easy to
manufacture as pills. And because they're so small, they can sometimes just slip inside cells and bind to proteins there. Another type of drug is
called an antibody, which is much larger. You can think of them as highly precise biological clamps. They're
designed to stick to the bad guy with extreme accuracy. Because of their size, they usually work outside cells or on the surface of cells. The hope is
that if you shut down the bad guy, the disease, such as cancer, would also stop growing and die. But this is where it gets incredibly difficult
because the body is not just one simple machine. It's made of a billion moving parts. Even though the
designed drug could successfully bind to the bad guy, it might affect other parts of the body as well and have some side effects. So, it's not as
simple as the analogy of a key and a lock. Maybe that key can also open other locks in the body, which could cause harmful effects. So, drug design is
basically an attempt to find the perfect molecular tool. Strong enough to stop the disease, but precise
enough so that it doesn't also harm the rest of the body, and safe enough to give to a real human. You also have to prove the drug actually works at
safe doses and it doesn't accidentally kill the patient. That's the real challenge of drug design. And this is where modern technology runs into a
weird problem. You see, we have incredibly powerful tools today, but most of them only understand one
slice of the puzzle. One AI model might predict protein structures. In fact, that's exactly what Google's AlphaFold does. Another AI might be great at
reading and generating DNA. One is called Evo 2, and in fact, I covered this in a previous video, so see this if you want to learn more. And we have
other tools that can screen chemical compounds, others that can analyze clinical trial data, and so
on and so forth. But, here's the thing, disease does not happen in separate folders. It flows through the entire system, from DNA to gene activity, to
proteins, to cells, and then to entire bodies. The problem is that all these tools are often siloed. They're trained on different data sets, built by
different teams, optimized for different tasks. So, instead of one system that understands the full
chain of biology, we have a bunch of specialized tools looking at disconnected snapshots. It's like trying to solve a crime scene where one detective
only has the fingerprints, another detective only has the security footage, another one only has the final autopsy report. Each clue matters, but they
don't really talk to each other, so they can't connect everything into one story. Well, that's
exactly what Mammal is trying to solve. They've built an AI model that understands everything at once, chemistry, genetics, and protein structure
simultaneously. And the scale of what they did is honestly insane. Mammal was pre-trained on 2 billion samples. They scraped basically every major
biological database out there. They observed antibody space, which contains billions of antibody sequences,
UniProt, which contains almost every protein we know. And then we also have Zinc and PubChem, which contains millions of small molecule structures,
Cell X Gene, which contains a ton of gene expression data. So, the model was trained on everything to do with chemistry, genetics, and proteins. But,
how does the AI actually read all of this? These are in different formats, right? A small molecule
like aspirin looks nothing like a gene, which also looks nothing like an antibody. So, the researchers did something clever. They forced everything
into a single unified format, sequences of characters, but each domain has its own grammar. For molecules, they used something called a SMILES string.
This is basically a way to flatten a 3D chemical structure into just one line of text. For example,
the SMILES for Tylenol becomes this, where every letter is an atom and every symbol, like the equal sign, is a chemical bond. The AI learns chemistry
just by tracking patterns across millions of these text strings. Now, for genes, it's totally different. The model takes every gene in a cell and
ranks them by how active they are. In other words, how much they are expressed. The genes screaming the
loudest go first. The silent ones go last. So, the model reads a cell as a priority list of which genes are doing the most work right now. And then
for proteins and antibodies, which are also proteins, it just reads the chain of amino acids. These are the building blocks of proteins. But, right
now everything is in different formats. It's like you're reading a book, but the first paragraph is in
English, the second paragraph is in sheet music, and the third paragraph is in JavaScript. It gets really confusing, right? This doesn't make sense at
all. And if you throw all this raw data to train a neural network, it also gets super confused. Well, this is where the engineering gets super clever.
Mammal uses what's called a modular tokenizer. Think of a tokenizer as a dictionary the AI uses to
translate this raw data into consistent language that the AI understands. But, Mammal doesn't just use one dictionary. It uses an umbrella tokenizer
with specialized sub dictionaries underneath. So, there's one dictionary for chemistry, one dictionary for genetics, and another dictionary for
proteins. If it sees a small molecule, it uses the small molecule dictionary to convert that into tokens
and embeddings. If it gets a protein, it uses the protein dictionary to convert that into embeddings. And the same logic applies to genes. And here's
the magic part. Once everything is translated and converted into these embeddings, they all get mixed together into a shared multi-dimensional space.
So, the model learns chemistry, proteins, and genetic expression all in one unified space. Because
it merges all these domains together, it can learn the relationships between all these different things. Now, this design sounds great, but how well
does it actually perform in the real world? For example, can it actually predict drug behavior? Well, the researchers put Mammoth through some pretty
intense tests. Specifically, they evaluated it across 11 distinct, highly rigorous benchmarks that
span the entire drug discovery pipeline. They threw everything at it, and it turns out the results across the board were phenomenal. So, here's the
table of results, and across these 11 benchmarks, Mammoth achieved state-of-the-art performance. It outright beat the previous best models in the
world on all of these tasks. Let's dig into some specific benchmarks. Earlier, we discussed the nightmare
of designing a drug that actually works. It's able to reduce the disease, but it also ends up being toxic to the human. So, is Mammoth able to predict
these scenarios? They evaluated Mammoth on two critical safety benchmarks. One is called BBBP, which stands for blood-brain barrier penetration.
Another one is ClinTox, which is a data set used to predict clinical toxicity and FDA approval. Let's
pause on this blood-brain barrier for a second. This is actually a massive hurdle in pharmacology. You see, the brain has this incredibly tight
security system. Of course, it doesn't want any external stuff flowing into the brain. But if you're designing a therapeutic for something like
Alzheimer's or Parkinson's, your drug must be able to penetrate this barrier to reach the brain. Or conversely,
if you're designing a really heavy chemotherapy drug for the liver, you absolutely do not want it crossing this barrier and getting into the brain. So
predicting that penetration is super important. Now, for this benchmark, the reigning champion for predicting this was a model called MoleFormer. This
is a highly specialized model. It was trained exclusively on over a billion small molecule
sequences. So think of it as like a hyper-focused master of small molecule chemistry. On the other hand, this new Mammal is a generalist. But the
insane thing is, on this blood-brain barrier penetration test, while the hyper-specialized MoleFormer achieved a very good score, Mammal was able to
beat it. And then same for this ClinTox safety benchmark, which predicts FDA approval. Mammal also beat
MoleFormer by a few percentage points. Let's take a second to highlight why this is so impressive. In almost every field, a specialist will beat a
generalist in their specific domain, right? For example, if a decathlete races a professional swimmer, the decathlete will likely lose since he's not
really specialized in swimming. Well, in this case, Mammal is the decathlete. It spreads its knowledge
across chemistry, genetics, and proteins. It's not just specialized in small molecule sequences, but yet it was able to beat the specialist pro
swimmer. It beats MoleFormer that was highly specialized in chemistry. That's actually very insightful. Why does knowing about genes and proteins also
make it better at chemistry? Well, that's because everything is interconnected in biology. There's a
billion moving parts, and this proves that being multimodal, being able to understand all these different domains, is not a distraction, but instead
an advantage. Small molecules exist to interact with proteins and alter gene expression. Because mammal was forced to learn the relationships between
all these different modalities during training, it developed a much deeper understanding of the
overall biology of the molecules. And this makes it much better at predicting outcomes compared to a specialized model. But the researchers didn't
stop there. If you make videos or content online, you should definitely check out Runway, the sponsor of this video. They just released Runway Agent.
Think of it as like your creative partner that can autonomously turn your idea into a publish-ready
video. I mean, full multi-shot videos with voice-over, music, scene transitions, and actual narrative structure already assembled for you. You start
by simply describing what you want. Maybe it's a product launch, an ad campaign, a YouTube intro, or even just a vague idea in your head. So, instead
of staring at a blank timeline in editing software for hours, you're basically working with an AI
creative partner that helps shape the idea with you. And here's the crazy part. Before the video is generated, you can see the entire plan first.
Every scene, the visual direction, the flow of the story, and you can tweak it, redirect it, or completely reshape it before it gets rendered. It
doesn't feel like gambling on prompts anymore. And once everything looks right, Runway Agent builds the
actual finished video for you. Multiple scenes, voice-over, music, editing, all assembled and ready to publish. Most AI video tools still you
fragments, but this feels much closer to an actual end-to-end workflow. That means more videos, faster campaigns, smaller teams producing bigger
content, and way less time stuck in post-production. Try Runway Agent today using the link in the description
below or the QR code on the screen, and get 50% off your first 3 months with the code agent50. Next, they tested it on labeling cell types using the
Zang 68K data set. This consists of thousands of gene activity data from a ton of different immune cell types in the blood. The AI's job is to look at
the genetic activity of the cell and correctly label it. For example, this is the CD4+ T cell or
this is the NK cell, etc. etc. Basically, label the cell given its genetic activity. And this is incredibly important for figuring out how a patient's
immune system is responding to a disease or treatment. It's a fundamental classification task. And on this highly complex task, MAML achieved a 7.5%
improvement over the state-of-the-art model. Again, this is a huge leap. And it gets even more
impressive. I think the highlight of this paper is this cancer drug response section. This is proof of a genuine biomedical breakthrough. So, here's
the setup for this test. The researchers wanted to prove that MAML wasn't just regurgitating information it had memorized from training. They wanted
to see if it could predict how completely new and unseen drugs would perform against human cancer
cells. So, they selected four drugs that were definitely not included in its training data. So, this includes carfilzomib, nintedanib, infigratinib,
and vemurafenib. What a mouthful. Anyway, they took the chemical strings for these four drugs and they gave them to MAML in addition to the genetic
profiles of over 800 different types of cancer tumor cells. This includes like lung cancer, breast
cancer, colon cancer, and so on. A huge diversity of human cancer cells. And they asked MAML one question. How effective are these four new drugs
across these 800 tumor cell types? Rank them from the most deadly down to the least deadly to the cancer. Now, you might be wondering if these drugs
weren't in the training data, what if they still looked very similar to something in the training data
that the AI memorized. That's a very valid concern. So, to prove these drugs were truly new to the model, to prove that it has never seen these types
of drugs before, the researchers calculated something called the Tanimoto similarity. This is basically used to measure how structurally similar two
molecules are based on their chemical structure. And the researchers confirmed that three out of the
four drugs had a maximum Tanimoto value of less than .7 when compared to anything Mammal had ever seen. And in the rigorous world of chemistry, a
score below .7 means that these drugs are basically structurally different. So, these drugs are like strangers to the AI. They look completely new to
the AI and like nothing it has seen before. Now, back to the task. It needs to rank how potent each drug
will be against 805 different variations of cancer tumors. And here's its prediction. Mammal returned a very confident and specific ranking. It
predicted that carfilzomib is the most potent across the majority of tumors. It then ranked nintedanib second, and then this one third, and this one
fourth, the least potent of the four. Now, here's the plot twist. If you actually look up carfilzomib, it's
a real FDA-approved drug. But it's currently approved and only used for blood cancers. All this time, doctors believed that carfilzomib was useless
against solid tumors. They just thought it didn't work. That's the currently accepted view from expert oncologists. But here, the AI put this as the
number one strongest drug against solid cancer types. Remember, we don't have this knowledge before.
Experts never thought this would be the case. If we put this to the test in real life, we would expect it to fail miserably, right? But here, Mammal
seems so confident. It looked at the structure of this drug and confidently said that this would be very strong against solid tumor cells, completely
contrary to what doctors think. So, the researchers put this to the test. They conducted a real-life
physical experiment where they tested these drugs against solid cancer cells, and what they found was shocking. The results perfectly matched mammals'
predictions. It got the precise potency ranking exactly right. Carfilzomib was the most potent, cascading down to vemurafenib as the least potent.
Mammals successfully preserved this relative order across approximately 95% of all 805 cancer cell
types. Let's take a second to process how insane this is. An AI that merely reads text strings looked at a blood cancer drug it had never seen before.
It looked at the genetics of hundreds of solid tumors, which is a different thing, and it correctly deduced that this drug would be effective against
these solid tumors, even though for decades human experts assumed these drugs would not work. The
implications of this are massive. What we're witnessing is an AI that can generalize to entirely new chemical compounds and accurately predict their
uses for treating diseases, which is the entire field of drug discovery, and especially for drug repurposing. If you're not familiar with drug
repurposing, it means taking a drug that already exists and testing whether it can treat a different
disease. Instead of starting from zero and designing a brand new molecule, scientists look at approved drugs or failed drug candidates and ask, "Could
this work somewhere else?" And this is really important because discovering or inventing a new drug from scratch takes like 10 to 15 years and
billions of dollars. But if you have an AI that can scan the entire library of existing drug candidates
and see if it can treat another disease, that opens up massive opportunities. But as incredible as this cancer drug story is, remember that small
molecules are just one type of drug. Another type is the antibody. These are larger proteins that clamp onto the problematic protein very precisely.
And predicting the structure and application of this is super complicated. You see, a protein is just a
long chain of amino acids that fold into complex 3D shapes. That shape determines what they stick to and what effect they will have in the body. But
in theory, a protein has a ginormous number of possible ways it could fold. If it tried every possible shape randomly, it could take longer than the
age of the universe to find the correct one. But in real life, of course, proteins fold almost
instantly because physics guide them towards the most stable shape. So, to create a protein-based drug, like an antibody, scientists need to
understand not just the protein sequence, but how it folds. How does it move and what does it bind to? Now, at least in terms of predicting how
proteins fold, that's exactly what Google's legendary AlphaFold does. But since Mammal was also trained on
proteins, what happens if we compare Mammal with the latest AlphaFold 3 head-to-head? Keep in mind, AlphaFold is super accurate. It's like the
industry standard for predicting proteins. It's considered like one of the greatest modern scientific achievements, and the team even won a Nobel
Prize in chemistry for this. Now, we just established that Mammal is just a sequence model, right? It reads
text in one dimension. How can it possibly beat AlphaFold, which is specialized at predicting the 3D structure of proteins? Well, the researchers put
this to the test. They showed both Mammal and AlphaFold an antibody and a disease target. And they asked, will these two stick together? Yes or no.
They tested both models across seven different well-documented targets, and again, the results were
shocking. Mammal actually beat AlphaFold 3 on five of those seven targets. Again, this is pretty crazy. AlphaFold 3 can literally see 3D shapes of
these proteins. So, shouldn't it easily see if it fits with the target disease? How on Earth could a one-dimensional text reader like Mammal actually
win? To understand why, we have to understand a very flawed assumption about how proteins work at the
microscopic level. You see, in textbooks or in other content, they often describe proteins as just rigid, static structures. Every textbook
illustration shows proteins as these solid interlocking shapes, because, well, that's the only way we could historically draw them or photograph them.
But, in reality, at least for human proteins, around 30 to 40% of the sequence consists of intrinsically
disordered regions, or IDRs for short. Now, what on Earth are these? These sections of the protein actually don't have a stable 3D shape. Instead,
they're floppy and highly flexible. An IDR kind of looks and behaves like a piece of wet spaghetti, constantly wiggling, folding, and shifting around.
So, actually, protein structures are quite dynamic, not static. They're not just solid blocks, but
they're actually fluid, and they can change shapes or move around. And, this is crucial because some of the most important protein targets, like EGFR
and HER2, which are notoriously aggressive drivers of breast cancer and stomach cancer, have a ton of these floppy intrinsically disordered regions.
If you give these floppy regions to AlphaFold 3, well, it struggles to actually predict the structure
of this because it's trained on predicting static and solid 3D structures. AlphaFold only learned from those frozen snapshots of proteins during
training. So, if you give AlphaFold 3 the sequence of these highly flexible and chaotic targets and ask if a specific antibody will bind to it, it
actually does not answer this well. But, if you give the same question to Mammal, it had much better
predictions. Mammal actually wins on the floppy proteins because it isn't trying to draw a static picture of them. Because Mammal is a large language
model operating on sequences, it doesn't try to force the protein into any specific shape. Instead, it seems to understand the underlying grammar or
the rules of proteins, including floppy parts. It relies on the math of the sequence and not just a
snapshot of the structure. And this is a brilliant distinction. It's far more robust at predicting whether an antibody will bind to a disease target.
And again, this unleashes some insane possibilities. We can now plug all antibody candidates through this AI and ask if it can bind to certain
diseases. And this could lead to a ton of huge discoveries and new drugs. Now, so far we've only asked
Mammal questions about existing candidates and whether they respond to certain diseases. What if we want to generate a completely new drug from
scratch? The crazy thing is they actually designed Mammal to be able to do this. It's not just an analyst, but it can also act as the inventor. So,
next the researchers tested its ability to design new antibodies from scratch. To understand the
significance of this, we first need to understand how antibodies work. An antibody is basically a massive Y-shaped protein used by our immune system
to identify and deactivate things like viruses or cancer cells. Picture it as a giant microscopic claw floating through the bloodstream. Now, the main
body of that Y-shaped claw is quite stable and consistent. But, the very tips of the claw, the parts
that actually reach out and physically grab the disease antigen, these are highly variable. These tips are called CDRs or complementarity determining
regions. And the CDRs are kind of like the fingers at the end of the claw. These fingers are the most important part of the antibody because they are
highly specific. The sequence of amino acids that code up these fingers determine exactly what
disease the antibody can grab onto. In fact, the entire field of modern antibody design involves trying to engineer new and better fingers to grab a
specific target. So, how did they set up this test for Mammal? They used a huge dataset called SabDab, which has a ton of different antibodies. They
took thousands of known and successful antibodies, but they artificially erased the CDRs or the
fingers so that the AI couldn't see it. They then fed Mammal the disease it needed to target, and essentially they gave it a fill-in-the-blank test.
Based on this target disease, predict the exact sequence of amino acids the fingers should have to bind onto this disease. Remember, Mammal only reads
one-dimensional text sequences. It's just a language model. It wasn't allowed to look at any 3D
structural maps of the antibody or the disease. It had to infer the physical shape of both of these just by understanding the grammar or the rules of
proteins. And guess what? The results were shocking. Mammal was able to generate the correct sequences of all these fingers even better than other
state-of-the-art methods that were specialized at this. But, here's where it gets even crazier. There's
one specific region where it completely blew away the competition, and that's the CDRH3 region. Now, what makes this specific finger so special? You
see, this CDRH3 region is notoriously the longest and most complex and chaotic and variable region of the entire antibody. Because of its length and
flexibility, it's almost entirely responsible for determining what the antibody can bind to. And
consequently, it's also the hardest piece for any AI to predict. And for this region, Mammal absolutely crushed the competition. It achieved a massive
19% improvement compared to the previous leading models. A 19% leap in predicting the hardest, most variable part of the antibody using nothing but
its underlying understanding of the grammar of proteins. That's pretty astonishing. It means the
model isn't just memorizing common patterns, it truly understands the rules of biology. So, as we zoom out, what does this all mean? We potentially
have the first true foundation model for biology. One that doesn't just read papers or just look at molecules or analyze genes, but it does all of it
at once. It's one unified model that understands everything. If Mammal actually works the way the
paper claims, we could be looking at a future where drug discovery isn't a 10-year, billion-dollar gamble that fails 90% of the time. Instead, we can
now design drugs way faster and cheaper and more accurately. We can also explore existing known drug candidates and see if they can target any other
disease. I mean, we can expect a massive acceleration in new drugs in the coming months if this is
true. And because this model also understands DNA and gene expression, personalized medicine could become mainstream. We can just take a DNA and blood
sample from the patient and plug it through this AI to figure out what's causing their specific disease, what's wrong with their DNA, or which
proteins in their body are messed up. And then this AI can also figure out what are the best drugs to
prescribe to this specific person. Or it can even custom design a new antibody just for this patient. And this This bold prediction, but in the paper,
they've already shown that it's really good at predicting how drug candidates bind to cancer cells. It already found a new genuine breakthrough. So,
you know, if we keep using this model, or if we keep improving this model even further, we could
potentially see some new cures for cancer very soon, in addition to other diseases. This is definitely one of the most impactful papers this year so
far. It's going to have huge impacts across medicine, biosciences, and drug discovery if what they claim is indeed true. Anyways, that sums up my deep
dive on this paper. It's very technical. There's a ton of complex biology stuff here, so I hope I
simplified it enough for you to understand. Let me know in the comments what you think of this. As always, I will be on the lookout for the top AI
news and tools to share with you. So, if you enjoyed this video, remember to like, share, subscribe, and stay tuned for more content. Also, there's
just so much happening in the world of AI every week. I can't possibly cover everything on my YouTube
channel. So, to really stay up-to-date with all that's going on in AI, be sure to subscribe to my free weekly newsletter. The link to that will be in
the description below. Thanks for watching, and I'll see you in the next one.