What if you could tell, before someone received a vaccine, whether it would work well for them? Researchers report an AI model doing exactly that — by reading the antibodies a person already carries.

A team at Arizona State University’s Biodesign Institute, led by Joshua LaBaer, trained a deep-learning model on antibody fingerprints across 185 antigens to identify “sentinel antibodies.” The study analysed 8,687 blood samples from 4,089 people — 2,445 healthy and 1,644 immunosuppressed — focused on COVID-19 vaccination, and was published in Cell.

The counterintuitive core of the idea

The sentinel antibodies are markers reflecting a person’s overall antibody-producing capacity rather than fighting the vaccine target directly.

That distinction is the whole insight, and it is easy to misread. The model is not finding antibodies that cross-react with SARS-CoV-2. It is finding antibodies whose presence and level report on how well the immune system works — the way a mechanic infers engine condition from how an unrelated component sounds.

An immune system that has mounted robust, durable antibody responses to the pathogens a person has encountered is an immune system likely to respond robustly to a new vaccine. The specific antibodies are the readout, not the mechanism.

What the model found

Higher pre-existing antibodies to certain common microbes — including Staphylococcus aureus, respiratory syncytial virus and a human respirovirus — correlated with stronger COVID-19 vaccine responses.

Those choices make sense as indicators. They are organisms nearly everyone encounters, so the relevant variation is not whether someone was exposed but how well they responded — which is precisely what a capacity marker requires. An antibody to something rare would confound exposure with response.

“Certain biomarkers, when analysed with AI, can predict who is likely to respond well to a vaccine, even before they receive it,” the researchers said.

Why the immunosuppressed cohort matters

Including 1,644 immunosuppressed people alongside 2,445 healthy participants is what makes this clinically interesting rather than merely elegant.

People on immunosuppressive therapy — after transplant, for autoimmune disease, during cancer treatment — respond variably to vaccination, and the variation is poorly predicted by which drug they take or at what dose. Some mount good responses; others essentially none.

They are also the people for whom the answer matters most, since they face higher risk from infection and cannot rely on a standard schedule protecting them.

What it could change in practice

The approach could help identify patients needing extra doses or alternative protection, guide vaccine development, and personalise care for immunocompromised people who often respond poorly.

Currently, poor response is discovered after vaccination, if at all — by measuring antibodies afterwards, which is not routine, or by the patient becoming infected. Knowing in advance changes the options: an additional dose, a different vaccine, or protection through monoclonal antibodies rather than relying on the patient’s own response.

For vaccine development, being able to characterise responders and non-responders before enrolment would let trials stratify participants rather than discovering afterwards that efficacy varied by a factor nobody measured.

The obvious limitations

The work is anchored on COVID-19 vaccination, and whether sentinel antibodies predict responses to other vaccines is untested. If the markers genuinely reflect general capacity they should generalise; if they capture something specific to this vaccine platform, they will not.

Prediction is also correlational. The model identifies who is likely to respond, and it does not establish why those individuals differ, which limits what can be done about it — knowing someone will respond poorly is useful only where an alternative exists.

And a model trained on this population may perform differently in others. Antibody profiles reflect regional pathogen exposure, so a model built in one setting may need recalibration elsewhere.

Why the approach is interesting regardless

Immunology has long lacked a practical measure of general immune competence. Clinicians can count immune cells and measure specific antibody levels, and neither answers the question of how well a person’s immune system will handle something new.

Why measuring immune competence has resisted solution

The gap this work addresses is older and wider than vaccination, and worth stating.

Clinicians routinely need to know how well a patient’s immune system functions — before starting immunosuppression, when deciding how aggressively to treat an infection, when assessing whether someone can safely receive a live vaccine. The available measures are crude: counts of immune cell populations, immunoglobulin levels, occasionally functional tests of specific pathways.

Those measure components rather than performance. Someone can have normal cell counts and mount poor responses, or abnormal counts and function adequately, because immune competence emerges from coordination among many parts rather than from the quantity of any one.

Reading the accumulated antibody repertoire sidesteps that by measuring outcomes rather than components — what the system has actually achieved against everything it has met. Whether that proves a robust general index or an artefact of this particular dataset is the question, and it is a more interesting one than the vaccination application alone.

Using existing antibody repertoire as that measure is a genuinely different idea — treating the accumulated record of past immune responses as a functional test rather than a history. It is a research finding for now, and a practical step toward more personalised immunisation. Research news, not medical advice.