The Biggest AI Breakthrough in Medicine and Drug Discovery?

Every so often, a new AI system appears that feels less like an incremental upgrade and more like a line in the sand.

MAMMAL may be one of those moments.

If the underlying research holds up in broader real-world use, this biology foundation model could become one of the most important developments in medicine and drug discovery in decades. That is not hype for hype’s sake. The claim here is straightforward and massive: an AI model trained across genes, proteins, antibodies, gene expression, and small molecules appears to outperform specialized systems across a wide range of biomedical tasks, including drug toxicity prediction, cancer drug response, antibody binding, and even aspects of protein-target interaction where AlphaFold 3 would seem to have the home-field advantage.

That matters far beyond academic labs.

For pharmaceutical companies, biotech startups, hospitals, research institutes, and public health systems, the traditional drug discovery process is brutally slow, expensive, and failure-prone. For Canadian businesses and institutions trying to compete in a global AI economy, tools like this could reshape how therapies are discovered, validated, repurposed, and personalized.

The big idea behind MAMMAL is simple to describe and very hard to build: instead of using separate AI models for chemistry, genetics, proteins, and cell biology, it tries to understand all of them in one unified system.

And that changes the game.

Why drug discovery is still so broken

To appreciate why MAMMAL is getting so much attention, it helps to start with the uncomfortable truth about modern medicine: most new drugs fail.

A promising therapy can take roughly 10 years and billions of dollars to move through the pipeline, and even then the failure rate remains astonishingly high. Around 90% of drug candidates do not make it through approval.

Think about how absurd that is in the context of everything else technology has achieved.

• We have sequenced the human genome.
• We have AI models that predict protein structures.
• We have models that can analyze DNA and generate biological sequences.
• We have huge chemical databases and advanced clinical analytics.

And yet when it comes to actually developing therapies that work in people, we are still wrong most of the time.

If a civil engineer built bridges with a 90% collapse rate, that would be a scandal. In medicine, it is normal. The reason is not that scientists are careless. It is that biology is staggeringly complex.

A drug does not act inside a vacuum. It enters a living system with layers of interacting mechanisms, feedback loops, unintended targets, safety constraints, and huge variation between patients. A compound might hit the intended protein perfectly and still fail because of toxicity, poor tissue penetration, off-target effects, or downstream biological consequences no one predicted.

This is the central bottleneck MAMMAL is trying to break.

A quick biology primer: DNA, genes, proteins, and disease

At a high level, living systems run on information and molecular machinery.

DNA stores the core instructions. Inside DNA are genes, which can be thought of as instruction sets for building proteins. Those proteins are the molecular workers of the body. They send signals, build structures, accelerate chemical reactions, regulate immune responses, and control countless cellular functions.

But genes are not simply on or off forever. Cells regulate them constantly. Some genes are highly active, some are quiet, and some are altered by mutation. This pattern of activity is called gene expression.

In simplified form, the chain looks like this:

• DNA contains genes
• Genes influence protein production
• Proteins do the work of life
• Disruptions in this chain can lead to disease

That disruption can happen in multiple ways:

• A mutation in DNA changes the instructions
• Gene expression becomes too high or too low
• A protein is produced incorrectly or in the wrong amount
• Cell signalling gets distorted

Cancer is a useful example. In very simple terms, mutations and expression changes can alter the proteins that regulate cell division. Those altered signals can tell cells to keep growing, ignore stop commands, and spread. That is one reason cancer is so difficult: it is not just a lump of bad tissue. It is a systems-level breakdown in regulation.

How drugs are designed today, and why that approach struggles

Traditional drug discovery starts by identifying what went wrong in the disease pathway. Researchers look for the “bad actor” that is driving the condition.

That could be:

• A protein telling cancer cells to keep dividing
• An enzyme helping a virus replicate
• A signalling pathway gone haywire

Once the target is identified, researchers try to create or find a molecule that can interact with it.

The classic metaphor is lock and key. The target is the lock. The drug is the key. If the key fits, it can block the target, activate it, or alter what it does.

Two major classes of therapies are especially relevant here:

1. Small molecules

These are relatively tiny compounds, often stable and easier to manufacture. Many can be formulated as pills. Because they are small, they can sometimes slip into cells and interact with internal proteins. Tylenol is a familiar example of a small molecule drug class.

2. Antibodies

These are much larger biological molecules. Think of them as highly precise clamps. They bind with extreme specificity to targets, often outside cells or on cell surfaces.

The challenge is that a drug cannot simply bind well. It also has to be safe, selective, manufacturable, and effective at doses humans can tolerate. A molecule that hits the intended target but also disrupts several unrelated systems can fail spectacularly.

This is why the lock-and-key analogy is useful but incomplete. In reality, many “keys” can partially fit multiple “locks,” and the body contains an almost absurd number of locks.

The real problem: biomedical AI has been siloed

One of the most important insights behind MAMMAL is that biology does not happen in separate folders.

Current AI tools are often exceptional at one slice of the problem:

• AlphaFold predicts protein structures
• DNA-focused models analyze or generate genomic sequences
• Chemistry models screen molecules
• Clinical models process trial or patient data

The problem is that disease flows across all these layers. It starts in DNA, changes gene activity, alters proteins, reshapes cell behaviour, and eventually affects whole tissues and organisms.

Most AI systems in biotech are specialized and siloed. They are trained on different datasets, built by different teams, and optimized for different objectives. So instead of one model that can connect the full biological story, we have many models each looking at a fragment.

MAMMAL tries to solve exactly that.

What MAMMAL actually is

MAMMAL is a biology foundation model trained across multiple biomedical domains at once. The model was pre-trained on approximately two billion samples drawn from major biological and chemical databases.

That training mix included data from sources such as:

• Observed Antibody Space, containing billions of antibody sequences
• UniProt, a massive repository of known proteins
• ZINC and PubChem, with millions of small molecule structures
• CellXGene, with large-scale gene expression data

That alone is ambitious. But the real trick is how the system handles such different data types.

How one AI can “read” molecules, genes, and proteins together

Aspirin does not look like a gene. A gene does not look like an antibody. So how do you feed all of that into one model?

The answer is elegant: convert everything into sequences, then use a smart translation layer.

For small molecules, the model uses SMILES strings, which flatten chemical structures into text-like strings of characters.

For gene expression, it represents a cell as a ranked list of genes ordered by activity level. The most active genes come first, the least active come later.

For proteins and antibodies, it reads amino acid sequences.

That still leaves a major challenge. These are different “languages.” Throw them all into a model raw, and the model gets confused. MAMMAL’s solution is a modular tokenizer.

Think of this as an umbrella system with specialized sub-dictionaries:

• One tokenizer for chemistry
• One for genetics
• One for proteins

Each input type is first translated using its own specialized vocabulary. Then all of those representations are mapped into a shared embedding space. That is the crucial step. Once the data lives in one shared mathematical space, the model can learn relationships across domains.

In other words, MAMMAL is not just learning chemistry and proteins separately. It is learning how chemistry interacts with proteins, how proteins connect to gene expression, and how these layers shape disease biology.

That cross-domain understanding appears to be where the model’s power really comes from.

The benchmark results are eye-opening

The researchers evaluated MAMMAL across 11 different benchmarks covering major parts of the drug discovery pipeline. Across these tasks, it reportedly achieved state-of-the-art performance.

That is already notable. But some of the individual results are where things start to look genuinely disruptive.

Drug safety and toxicity prediction

Two especially important benchmarks were used here:

• BBBP, which measures blood-brain barrier penetration
• ClinTox, which predicts clinical toxicity and FDA approval-related outcomes

The blood-brain barrier is a huge issue in pharmacology. If you are treating a brain disease like Alzheimer’s or Parkinson’s, the drug may need to cross into the brain. If you are designing a treatment for another organ, you may want the opposite.

Getting this wrong can derail a therapy.

What makes MAMMAL’s result striking is that it outperformed Moleformer, a highly specialized model trained specifically on more than a billion small molecule sequences. That is the equivalent of a generalist beating a specialist in the specialist’s own event.

Why would a broader model outperform a chemistry expert on chemistry-related tasks?

Because in biology, chemistry never acts alone. Small molecules exist to interact with proteins and influence gene activity. A model that understands those relationships may make better predictions than one trained on molecular structure in isolation.

That is one of the most important takeaways from this entire paper: multimodal understanding is not a distraction in biology. It is an advantage.

Cell type classification from gene activity

MAMMAL was also tested on the Zeng68K dataset, where the task is to label cell types based on gene expression profiles.

The model has to infer whether a cell is, for example, a CD4+ T cell, an NK cell, or another immune cell type purely from its activity pattern.

This kind of task is foundational in immunology, disease analysis, and treatment monitoring. If you want to understand how a patient’s immune system is responding, you need to know which cells are present and what they are doing.

On this benchmark, MAMMAL improved on the prior state of the art by 7.5%, which is substantial for a mature classification problem.

The cancer drug result may be the biggest signal of all

If there is one section that feels like a genuine biomedical turning point, it is the cancer drug response experiment.

The researchers wanted to test whether MAMMAL could generalize to truly new drugs, not simply retrieve patterns from memory. So they selected four drugs that were excluded from training data and tested the model’s ability to predict how those compounds would perform across more than 800 human cancer cell types.

The question was simple: rank these drugs from most potent to least potent against a broad set of tumour cells.

To make sure the drugs were genuinely unfamiliar, the researchers checked structural similarity using Tanimoto similarity. Three of the four drugs had maximum similarity scores below 0.7 relative to anything the model had seen. In chemistry terms, that means they were effectively strangers.

MAMMAL predicted the following order of potency:

1. Carfilzomib
2. Nintedanib
3. Infigratinib
4. Vemurafenib

Here is where it gets wild.

Carfilzomib is a real FDA-approved drug used for blood cancers. The accepted view has been that it is not useful against solid tumours. But MAMMAL ranked it as the strongest of the four across solid cancer cell types.

That prediction went directly against conventional expectations.

So the researchers tested it experimentally against solid tumour cells. The result reportedly matched MAMMAL’s ranking exactly, and the relative order held across roughly 95% of 805 cancer cell types.

That is not a trivial win. That is an AI system identifying an unexpected therapeutic use for an existing drug and getting the potency ranking right on unseen compounds.

The implications for drug repurposing are enormous.

Why drug repurposing matters for business and healthcare

Drug repurposing means taking an existing drug, or a shelved candidate, and testing whether it can treat a different disease than the one it was originally designed for.

This is attractive because it can compress timelines and reduce risk. Instead of spending 10 to 15 years inventing a molecule from scratch, organizations can search among compounds that already have some safety, manufacturing, or biological data behind them.

If AI can systematically scan libraries of approved and experimental compounds and identify hidden opportunities, the economics of pharma and biotech change fast.

For Canada, that matters on several fronts:

• Biotech startups could pursue more efficient discovery strategies
• Academic labs could validate therapeutic hypotheses faster
• Hospital research networks could connect molecular data to treatment options
• Investors could see lower-cost pathways to value creation

In hubs like Toronto, Montreal, Vancouver, and Waterloo, where AI talent and life sciences increasingly overlap, this kind of multimodal biomedical AI could become a serious strategic asset.

MAMMAL vs. AlphaFold 3: why this comparison is so surprising

Another standout result involved antibodies and protein targets.

AlphaFold is one of the great scientific achievements of modern AI. It transformed protein structure prediction so profoundly that its creators were awarded the Nobel Prize in Chemistry. So when a new model appears to beat AlphaFold 3 on certain binding tasks, people pay attention.

The comparison here was not “which model predicts structure better in general?” It was more specific: given an antibody and a disease target, can the model predict whether they will bind?

Across seven targets, MAMMAL reportedly beat AlphaFold 3 on five.

At first glance, that sounds backwards. AlphaFold can infer 3D structures. MAMMAL is fundamentally a sequence model. How could the text-based system win?

The hidden problem with static protein thinking

The answer lies in a flawed intuition many people have about proteins.

We often imagine proteins as rigid, stable, interlocking objects because that is how textbooks and diagrams traditionally depict them. But many proteins are not rigid at all. Large portions of human proteins are made up of intrinsically disordered regions, or IDRs.

These regions do not settle into one neat stable structure. They are floppy, flexible, and dynamic, more like wet spaghetti than machine parts.

This matters because some medically important targets, including proteins like EGFR and HER2, contain significant disordered regions. Models built around static structural snapshots can struggle in these situations.

AlphaFold is incredibly strong at predicting stable 3D conformations. But if the biology of binding depends on flexibility and dynamic sequence-level behaviour, a sequence-native model may sometimes hold the advantage.

MAMMAL appears to benefit from this. Rather than trying to force a static picture, it seems to learn the underlying “grammar” of proteins, including sequences associated with flexible and disordered behaviour.

That makes it potentially powerful for antibody discovery, especially against difficult disease targets.

Designing entirely new antibodies from scratch

This is where the paper moves from analysis to invention.

MAMMAL was also tested on its ability to generate new antibody components. To understand why that is significant, it helps to know how antibodies work.

Antibodies are Y-shaped proteins used by the immune system. Most of the antibody structure is fairly stable, but the tips of the Y contain highly variable regions that determine what the antibody can grab onto. These regions are called CDRs, or complementarity-determining regions.

They are effectively the fingers of the antibody.

The researchers used a large antibody dataset and removed the CDR sequences, then asked MAMMAL to predict them based on the target antigen. This is a very hard fill-in-the-blank problem. The model was not given 3D structural maps. It had to infer the likely amino acid sequences from sequence relationships alone.

It reportedly outperformed specialized state-of-the-art methods, with its most dramatic gain in the notorious CDR-H3 region.

Why is CDR-H3 special? Because it is typically the longest, most flexible, most variable, and most functionally important part of the antibody. It often plays a central role in determining target specificity. It is also one of the hardest regions for any model to predict well.

On this region, MAMMAL achieved a 19% improvement over prior leading methods.

That is a serious leap.

What this means for Canadian healthcare, biotech, and enterprise AI

This is where the conversation gets especially relevant for a Canadian Technology Magazine audience.

MAMMAL is not just a science story. It is a business technology story.

If biology foundation models become practical tools, they could influence:

• Pharmaceutical R&D by reducing failed candidate selection
• Healthcare delivery through more personalized treatment matching
• Biotech venture creation by lowering barriers to therapeutic design
• Public-private research partnerships by accelerating translational medicine
• AI infrastructure demand across cloud, compute, and model deployment

For Canada, there is a particularly interesting opportunity at the intersection of AI strength and life sciences ambition. The country already has recognized AI research depth and growing biomedical innovation clusters. A model like MAMMAL hints at what happens when those worlds truly merge.

Imagine the practical possibilities:

• A Toronto hospital integrates genomic and expression data to identify better therapy options for specific cancer patients.
• A Montreal biotech uses multimodal AI to repurpose existing compounds for rare disease indications.
• A Vancouver life sciences firm screens antibody candidates against difficult targets without relying exclusively on slow wet-lab iteration.
• A GTA startup builds workflow products on top of biology foundation models for pharma clients.

None of that is guaranteed by one paper. But the direction is hard to ignore.

The long-term vision: personalized medicine at scale

Perhaps the boldest implication is personalized medicine.

If a model can understand DNA, gene expression, proteins, antibodies, and small molecules together, then in principle it could help map an individual patient’s biology to individualized treatment options.

The future scenario looks something like this:

1. A patient provides DNA and blood samples.
2. The AI analyzes gene expression and molecular abnormalities.
3. It identifies the disrupted pathways and likely disease mechanisms.
4. It recommends the best existing therapies for that specific patient.
5. In more advanced settings, it could even help design a tailored antibody candidate.

That is still an aspirational vision. Clinical translation, regulation, validation, liability, privacy, and implementation all remain major hurdles. But the paper points to a world where that vision feels materially closer rather than purely theoretical.

A note of caution: breakthrough does not mean instant replacement

It is important to stay grounded.

MAMMAL’s reported results are extraordinary, but biomedical AI is notorious for looking impressive in papers and much messier in deployment. Models need external validation, reproducibility across labs, and performance under real clinical and commercial constraints.

There is also a difference between ranking candidates well and delivering approved medicines at scale. Even a highly capable model does not eliminate the need for experiments, toxicology, regulation, manufacturing, and human oversight.

Still, none of that weakens the core significance here. If MAMMAL truly generalizes the way these results suggest, then we may be looking at one of the first genuine foundation models for biology, not just another narrow AI tool.

The bottom line

MAMMAL matters because it tackles the deepest structural problem in medicine and drug discovery: biology is connected, while our tools have mostly been fragmented.

By training one system across chemistry, genetics, proteins, antibodies, and gene expression, the model appears to capture relationships that siloed systems miss. That may explain why it can beat specialist models on specialist tasks, identify unexpected cancer drug potential in unseen compounds, outperform AlphaFold 3 on certain antibody-target binding problems, and generate difficult antibody regions with striking gains.

If these findings continue to hold, the implications are massive:

• Faster drug discovery
• Cheaper candidate screening
• More effective drug repurposing
• Stronger antibody engineering
• Better path toward personalized medicine

For Canadian business leaders, healthcare innovators, and tech entrepreneurs, this is exactly the kind of AI development worth tracking closely. It sits at the intersection of commercial opportunity, national research strength, and real human impact.

The future of medicine may not belong to one model alone. But MAMMAL makes a compelling case that the next era of biomedical innovation will be built by AI systems that understand biology as one connected language.

FAQ

What is MAMMAL in AI and medicine?

MAMMAL is a multimodal biology foundation model designed to understand several core domains of biology at once, including genes, proteins, antibodies, small molecules, and gene expression. Its goal is to support drug discovery, drug repurposing, antibody design, and broader biomedical research.

Why is MAMMAL considered a breakthrough in drug discovery?

It is considered a breakthrough because it appears to outperform specialized models across multiple benchmarks, rather than just one narrow task. The most striking result was its ability to predict the effectiveness of unseen cancer drugs against hundreds of tumour cell types and correctly identify an unexpected strong candidate for solid tumours.

How is MAMMAL different from AlphaFold 3?

AlphaFold 3 is primarily focused on structural prediction, especially 3D protein-related modelling. MAMMAL is a broader foundation model that works across multiple biological data types. In some antibody-target binding tasks, MAMMAL reportedly outperformed AlphaFold 3, particularly where protein flexibility and intrinsically disordered regions made static structure assumptions less effective.

What is drug repurposing, and why does MAMMAL matter for it?

Drug repurposing means finding new uses for existing drugs or previously developed candidates. This is attractive because it can save years of development time and reduce cost. MAMMAL matters because it appears capable of identifying useful relationships between drugs and diseases even when the compounds are structurally new to the model.

Could MAMMAL lead to personalized medicine?

Potentially, yes. Because it can connect DNA, gene expression, proteins, and drugs, a model like MAMMAL could eventually help match individual patients to more precise therapies. In more advanced future use cases, it might also assist in designing custom therapeutic antibodies for specific patients.

Why should Canadian businesses and healthcare leaders care about MAMMAL?

Canadian organizations working in AI, life sciences, healthcare, and business technology should care because models like this could reshape pharmaceutical R&D, clinical decision support, biotech startup formation, and translational research. For Canada’s growing innovation hubs, this kind of technology could create major opportunities in both research and commercialization.

Is your organization ready for the next wave of biomedical AI?

As biology foundation models begin moving from research papers toward practical deployment, the winners may be the companies and institutions that start building expertise early. Is your business, lab, or healthcare organization prepared for this shift in AI-powered medicine? The answer may define who leads the next chapter of Canadian tech.