Ten ways AI will help research on aging, cancer and drug delivery

I have written code for most of my life and done biology for most of my career, and the two have never been closer. Each new model announcement claims to change everything. In the fields I work in, aging, cancer biology and drug delivery with extracellular vesicles, the gains are more specific than that, and some of them are already in the lab. Here are ten places where I expect AI to help the most, roughly in the order I expect them to arrive.

1. Designing proteins instead of finding them

For decades, getting a protein that binds a target meant screening large libraries and hoping. AlphaFold changed the first half of the problem in 2021 by predicting structure from sequence with near-experimental accuracy, and AlphaFold 3 extended prediction to complexes of proteins with DNA, RNA and small molecules. Design tools such as RFdiffusion then turned the question around: start from the shape you want and generate a protein that folds into it. The 2024 Nobel Prize in Chemistry went to David Baker for protein design and to Demis Hassabis and John Jumper for structure prediction.

This is the area I know best from the computing side. As SEVO Bioscience's liaison with Mila, the Quebec AI Institute, I worked on an AI algorithm that designs proteins for downstream applications.

2. Giving drug carriers an address

Extracellular vesicles are good at being accepted by the body and poor at going anywhere in particular. Most of what is injected ends up in the liver and spleen. One way to steer them is to display a protein on their surface that binds a receptor on the target cells.

Protein design makes that practical. Instead of borrowing a natural ligand that also binds a dozen other things, you can design a small binder for one receptor, check in silico that it folds and binds, and fuse it to a protein that sits in the vesicle membrane. The same goes for skin, where a binder could help a vesicle stick to the cells it is meant to reach. I expect designed targeting proteins to become a standard part of building a carrier within a few years.

3. Predicting what the immune system will do

A carrier that triggers the immune system can't be given twice, which is why immune testing is part of the carrier work from the start. Today that testing happens late and costs a lot: you make the carrier, then expose human immune cells to it and measure the response.

Models trained on known immune epitopes can already flag protein sequences likely to be presented to T cells, and they are improving. For designed proteins on a vesicle surface, that means removing the riskiest sequences before anything is made. They will not replace testing on human cells, but they can decide which candidates are worth testing.

4. Reading microscopy images at scale

Much of cell biology still runs on someone looking at images and counting. Deep learning has changed that. Cellpose, a general model for finding cell boundaries in microscopy images, works across many cell types without retraining, and similar tools now classify cell states from shape and staining.

For my work this matters in two places. Senescent cells look different: larger, flatter, with changes in the nucleus. A model that spots them reliably makes senescence assays faster and less subjective. And in carrier studies, uptake is measured by imaging labelled vesicles inside cells, which a model can quantify across thousands of cells instead of a few dozen.

5. Measuring biological age

In 2013, Steve Horvath built an epigenetic clock that estimates age from DNA methylation across many human tissues. Since then, machine learning has produced clocks from gene expression, proteins and blood tests, and newer clocks try to predict mortality rather than age.

The value for aging research is a measurement that moves within months. A treatment meant to slow aging can't wait decades for lifespan data. A clock that responds to treatment, and can be broken down into separate processes such as inflammation or mitochondrial function, gives a trial something to measure. The clocks still need validation in people over time, and I am cautious about any clock that has only been tested against age.

6. Making sense of single-cell data

A single experiment can now measure gene expression in hundreds of thousands of individual cells, and spatial methods add where each cell sits in the tissue. No one reads that by eye. Machine learning sorts the cells into types and states, finds the rare populations and lines them up across experiments.

In cancer, that is how researchers find the small group of cells that survive a treatment and regrow the tumour. In aging, it shows which cell types age first in a tissue. In skin, it can show what happens to fibroblasts in the dermis as collagen production falls. The bottleneck has moved from generating the data to asking good questions of it.

7. Characterizing vesicles properly

Vesicle preparations are messy. Each one is a mix of particles of different sizes carrying different proteins and RNA, plus contaminants from isolation. The field's guidelines, MISEV2023, list what should be measured, and much of that is now done with mass spectrometry and particle-by-particle imaging that produce large datasets.

Machine learning can compare a new preparation against reference profiles, flag contamination, and check that one batch looks like the last. For a company making vesicles for skin products, batch-to-batch consistency is the core quality question, and it is exactly the kind of pattern recognition these models do well.

8. Planning experiments

Loading a drug into vesicles depends on many variables: the method, the ratio of drug to vesicle, temperature, time, buffer. Testing every combination is impossible, so labs usually change one variable at a time and miss the interactions.

Optimization methods that choose the next experiment from the results of the last ones can find a good recipe in far fewer runs. Paired with automated liquid handling, they become the so-called self-driving lab. The same approach applies to formulation for skin, where stability, texture and delivery pull in different directions.

9. Reading the literature

More papers are published on aging, cancer and extracellular vesicles every week than anyone can read. Language models are already useful for finding relevant studies, pulling methods and results into tables, and spotting when two fields have answered the same question in different words.

They also make things up. A model that summarizes a paper can invent a result or a citation that sounds right. I use these tools to find papers and then read the papers. Used that way they save a great deal of time.

10. Choosing patients and doses

The last step is the slowest. Many cancer drugs work well in a minority of patients and poorly in the rest, and trials fail because the right patients were mixed with the wrong ones. Triple-negative breast cancer, which I studied as a postdoc, is a good example: it is a group of diseases with different biology under one name. Models that read pathology slides, gene expression and clinical data together can help sort patients into groups that respond differently, and help pick doses from early data.

This is also where the stakes and the regulation are highest, so it will move slowest.

What AI will not do

None of these removes the experiment. A designed protein is a hypothesis until it is made and tested. A clock is a correlation until a treatment moves it and outcomes follow. A model trained on published data inherits its gaps, and the most interesting biology tends to sit in those gaps.

What changes is the shape of the work. Fewer experiments are wasted, more questions get asked, and the loop between an idea and a result gets shorter. For someone who has always had one foot in code and one at the bench, that is the most exciting part.