MANAscore: How an AI-Powered 3‑Gene Model Could Transform Cancer Immunotherapy

The Problem With Finding Tumor-Reactive T Cells

Kellie Smith, Ph.D., a researcher at the Bloomberg Kimmel Institute for Cancer Immunotherapy, spent years trying to isolate these elusive immune cells using sophisticated research strategies. The results were real — but painfully limited.

“It was very expensive. It took a really long time, and it used a lot of patient samples, and we were still only able to find these cells in three patients,” Dr. Smith explained.

That’s the core challenge in cancer immunotherapy research. You can have the right treatment strategy, but without a reliable way to identify which patients have tumor-reactive T cells — and in what quantity — you’re essentially flying blind.

Traditional methods required massive datasets, enormous sample volumes, and weeks of analysis. Scaling that to routine clinical use was never realistic. Something had to change.

How AI Turned Terabytes Into Three Genes

Dr. Smith and her team turned to AI to do what human analysis couldn’t — process immense biological datasets and extract the signal from the noise.

The result was MANAscore, a predictive model built on just three well-known genes:

• CD39 — marks T cells that recognize the tumor but are worn down from sustained immune activity
• CXCL13 — recruits other immune cells and helps build local hubs of immune activity within the tumor
• IL-7 receptor (IL-7R) — supports T cell survival and long-term immune memory

Together, these three markers identify T cells that are most engaged with the cancer — even in tiny numbers — and capable of initiating a meaningful immune response.

What makes this genuinely remarkable is the compression. Other predictive models require 200 or more genes to achieve similar results. MANAscore does it with three. That’s not a minor efficiency gain — it’s the difference between a research-only tool and something a clinic can actually use.

Why Simplicity Is the Real Innovation

In AI and machine learning, more features don’t always mean better predictions. Overfitted models with hundreds of variables often perform worse on real-world data than lean, well-chosen models built on the right signals.

MANAscore demonstrates this principle in a high-stakes medical context.

Because the model relies on just three genes, it can be applied to archival tissue samples — biopsies already sitting in hospital storage from patients treated years ago. That opens up a massive retrospective research opportunity and, more importantly, a practical path to clinical deployment.

Early results show that the frequency of three-gene-positive cells correlates strongly with how well patients respond to immune checkpoint inhibitors. That’s a direct, measurable link between the biomarker and treatment outcome — exactly what clinicians need to make better decisions.

What MANAscore Is Being Used For Right Now

The team isn’t waiting to see if the model holds up. They’re actively validating it through two parallel tracks.

Retrospective validation involves looking back at pre-treatment biopsies from patients whose outcomes are already known. This tests whether MANAscore’s predictions would have matched real-world results — a critical step before any clinical tool earns trust.

Prospective testing means applying the model to new patients as they begin treatment, tracking predictions against outcomes in real time.

Beyond response prediction, the team is also using MANAscore to study how tumor-reactive T cells interact with their cellular neighbors — including regulatory T cells that can suppress the immune response. Understanding those interactions could point toward entirely new treatment strategies.

This is where AI’s value compounds. It doesn’t just answer the question you asked. It opens doors to questions you hadn’t thought to ask yet.

The Clinical Timeline: Closer Than You Think

MANAscore is still classified as a research tool. But Dr. Smith believes it could be ready for clinical use within two years.

“By applying MANAscore to a single biopsy slide, we may soon be able to quickly predict which patients are most likely to benefit from immunotherapy,” she said.

That’s a significant statement. A single biopsy slide. A three-gene panel. A prediction that could determine whether a patient receives immunotherapy — or is spared from a treatment unlikely to help them while being directed toward alternatives that might.

For oncologists managing complex treatment decisions under time pressure, that kind of rapid, evidence-based signal is transformative.

What This Looks Like for a Real Patient

The numbers and gene panels matter. But so does the human reality behind them.

REBB, a pancreatic cancer patient treated through a Kimmel Cancer Center immunotherapy clinical trial, was so weak before treatment that she needed a wheelchair for any prolonged distance. She lost her daughter Valerie to breast cancer. She knows exactly what’s at stake in this research.

“Almost immediately, I started feeling better. My immune system was destroying the cancer. By April, I was running through the airport, and by December 2022, I was cancer-free. I want this for all cancer patients.”

She added: “The idea that we can move so much faster using AI is extremely exciting.”

Her story is a reminder of what’s actually being optimized here. Not a model’s accuracy score. Not a publication in a peer-reviewed journal. A person’s life — and the speed at which the right treatment reaches them.

What MANAscore Tells Us About AI in Healthcare

MANAscore isn’t just a cancer research breakthrough. It’s a case study in how AI in Healthcare should be applied to complex medical problems.

The approach — use AI to process massive datasets, identify the minimal viable signal, then build a tool simple enough to deploy in real clinical settings — is a template worth paying attention to.

Most AI healthcare tools fail not because the underlying models are wrong, but because they’re too complex, too expensive, or too dependent on data that most hospitals don’t have. MANAscore sidesteps all of that. Three genes. One biopsy slide. Actionable prediction.

That’s the kind of AI application that actually makes it out of the lab.

The Bigger Picture for Precision Oncology

Immunotherapy has already changed cancer treatment. Checkpoint inhibitors have produced remarkable outcomes for patients who respond to them. The problem has always been predicting who will respond — and who won’t.

MANAscore directly attacks that uncertainty. If validated at scale, it could shift immunotherapy from a treatment oncologists try and hope works, to one they prescribe with confidence based on a patient’s specific immune profile.

That’s precision oncology in practice. Not in theory. Not in a press release. In a clinic, on a biopsy slide, in the next two years.