In cancer and many other diseases, the cells that matter most are often present in only tiny numbers. In leukemia, the earliest malignant cells, the last cells surviving after treatment, and the cells responsible for relapse may all represent tiny minorities in a vast background of other cells. Similarly, rare immune cell populations - present at less than 1% of all cells - can determine whether a patient will respond to therapy or succumb to infection. Identifying these rare populations with conventional tools is extremely difficult.
Modern measurement technologies like mass cytometry (also called CyTOF) can simultaneously measure dozens of protein markers on thousands of individual cells at once. This generates enormously complex data - thousands of cells each described by 20 to 40 measurements - that captures the full heterogeneity of a cell population in remarkable detail. However, turning this data into biological and clinical insight requires sophisticated analytical methods.
Existing approaches for analyzing multi-parameter single-cell data typically work in two steps. First, cells are grouped into clusters using unsupervised methods that do not consider disease status. Second, a separate analysis asks which clusters differ between patients and healthy controls. This two-step approach has a fundamental weakness: the clusters are defined without knowing which ones are relevant to disease, so disease-associated rare cell populations may be split across clusters or lumped with unrelated cells and become invisible.
The paper introduces CellCnn, a machine learning method that solves this problem by combining feature learning and disease association into a single step. Instead of first clustering all cells and then asking which clusters differ, CellCnn simultaneously learns what cell populations look like AND whether their presence is associated with the disease phenotype. This allows it to find disease-relevant populations even when they are extremely rare and their characteristics are not known in advance.
CellCnn adapts a type of artificial intelligence called a convolutional neural network (CNN) to analyze single-cell data. CNNs were originally designed for image recognition, where they learn to detect specific visual patterns (like edges or shapes) anywhere in an image. CellCnn repurposes this architecture to detect specific molecular patterns in populations of cells.
The network works through convolutional filters - mathematical templates that each learn to recognize a specific molecular profile. Each filter is applied to every cell in a sample, computing how well that cell's protein expression profile matches the filter's template. A cell that closely matches the filter pattern gets a high 'response score'; cells that do not match get low scores. The learned filter weights, after training, correspond to the molecular profile of the disease-associated cell type.
After computing cell-filter responses, a pooling layer summarizes the responses across all cells in a sample. Using 'max-pooling' records the maximum response - indicating whether any cell in the sample closely matches the disease profile. Using 'mean-pooling' computes the average response - approximating the frequency of cells matching the profile. This pooling step converts the per-cell information into a sample-level prediction of disease status or survival.
The network is trained by showing it samples from patients with and without the disease (or with and without a specific clinical outcome). Training adjusts the filter weights to maximize the accuracy of disease classification. Once trained, the filter weights can be read directly to understand which combination of protein markers defines the disease-associated cell population - making the learned model biologically interpretable, not just a black box.
The researchers first tested CellCnn on a dataset where the correct answer was already known. They used mass cytometry data from immune cells exposed to various chemical signals (paracrine agents), and asked whether CellCnn could identify which cell types responded to each agent. CellCnn correctly identified that monocytes and dendritic cells were the cell types responding to GM-CSF stimulation - a biologically correct answer consistent with the known biology of this immune signaling molecule.
The second test applied CellCnn to a clinically relevant problem: predicting which HIV-infected patients were at highest risk of developing AIDS. Using flow cytometry data from 383 HIV-infected individuals, CellCnn was trained on 256 patients and then tested on 127 held-out individuals. CellCnn identified cell subsets characterized by high levels of Ki67 (a proliferation marker) or naive T cell identity as most predictive of AIDS risk - consistent with prior research. The survival difference between CellCnn's high-risk and low-risk groups was highly significant (p = 0.003), notably more so than the Citrus comparison method (p = 0.030).
Critically, CellCnn achieved this HIV risk stratification in approximately one hour on a single laptop core, compared to three days on 24 server cores for the competing Citrus approach. This dramatic improvement in computational efficiency means CellCnn can be practically applied to large patient datasets without requiring specialized computing infrastructure - an important consideration for translating such methods to clinical or research settings.
CellCnn also identified a rare natural killer (NK) cell population - present in fewer than 1% of cells - associated with prior cytomegalovirus (CMV) infection. This population had a memory-like phenotype (NKG2C+, CD57+) consistent with published immunology. The competing Citrus method completely failed to detect this rare population, illustrating CellCnn's ability to identify genuine biological signals that are invisible to existing tools when those signals come from very rare cell types.
The most demanding test was detecting leukemic cells in a simulation of minimal residual disease (MRD) in acute myeloid leukemia. MRD refers to small numbers of leukemia cells that remain in a patient's bone marrow after treatment - invisible to standard microscopy but capable of causing relapse. The researchers simulated MRD by mixing known numbers of AML blast cells into samples of healthy bone marrow at precisely controlled, very low frequencies.
CellCnn was asked to classify samples as healthy, CN-AML (cytogenetically normal AML), or CBF-AML (core-binding factor AML) based on mass cytometry data measuring 13 surface and intracellular markers. At a blast frequency of 0.1% (1 in every 1,000 cells being a leukemic blast), CellCnn correctly identified the leukemia subtype in test samples it had never seen during training. The healthy and two leukemia classes were clearly separated in CellCnn's learned representation space.
When the blast frequency was reduced to 0.01% (1 in every 10,000 cells), CellCnn continued to identify leukemic blasts accurately. For the personalized medicine scenario - where the same patient's samples are compared before and after treatment - CellCnn could detect leukemic blast populations down to frequencies of 0.005% (1 in every 20,000 cells). These sensitivity levels are in a range clinically relevant for MRD detection in leukemia.
CellCnn also correctly characterized the biological differences between CN-AML and CBF-AML blasts. CN-AML blasts were identified as CD7+, CD38+, and CD117+ (c-KIT+), while CBF-AML blasts showed CD15+ and CD38mid expression. These distinctions match the known immunophenotypes of these two leukemia subtypes from the clinical literature, providing biological validation that CellCnn was detecting genuine leukemic populations and not spurious signals.
Minimal residual disease detection is one of the most important challenges in leukemia management. Patients who achieve a complete remission - meaning no leukemia cells are visible by standard testing - can still relapse because of small numbers of residual cells that survive treatment. Early detection of MRD has been shown to predict relapse in multiple leukemia types, and increasingly patients with detectable MRD receive intensified therapy to prevent relapse.
Current MRD detection methods for AML include flow cytometry (measuring protein markers on individual cells) and molecular methods (detecting specific gene mutations or fusion proteins by highly sensitive PCR). Flow cytometry-based MRD requires expert operators to identify abnormal populations visually - a process that is subjective, time-consuming, and requires extensive training. CellCnn could enable automated, standardized MRD detection that does not depend on manual gating and could be deployed more broadly across clinical laboratories.
The ability to distinguish cytogenetically normal AML from core-binding factor AML at the single-cell level from protein expression data alone has additional value. CN-AML and CBF-AML require different treatment intensities and different follow-up protocols. An automated tool that can reliably identify these subtypes from a bone marrow sample could supplement or in some settings reduce the need for specialized cytogenetic testing.
Beyond leukemia, the detection of rare immune cell populations has implications for cancer immunotherapy. Predicting which patients will respond to immune-based treatments - whether checkpoint inhibitors, cellular therapies, or vaccine approaches - may depend on identifying specific rare immune cell subsets in the tumor microenvironment or peripheral blood before treatment. CellCnn's ability to detect rare populations associated with disease status positions it as a tool for developing new immune biomarkers.
CellCnn represents a conceptual advance in how single-cell data is analyzed. The key insight is that separating the steps of feature extraction and disease association - as all prior methods did - is fundamentally limiting when disease-relevant cell populations are rare. By learning these simultaneously through a single neural network, CellCnn efficiently extracts the biological signal relevant to disease while ignoring the vast majority of single-cell variation that is not disease-related.
The tool is computationally efficient, scaling linearly with the number of measured protein markers. This makes it applicable not just to flow and mass cytometry but also to emerging technologies including single-cell RNA sequencing, which measures gene expression in thousands of individual cells simultaneously, and single-cell imaging data. As these technologies generate ever-larger datasets, scalable analysis methods like CellCnn will be essential.
CellCnn is implemented as open-source software freely available for download, written in Python and using established neural network libraries. The authors validated it across multiple datasets with very different measurement platforms and disease contexts, demonstrating generalizability. The software's outputs are biologically interpretable - filter weights describe the molecular characteristics of identified cell populations in terms of the measured markers - making results accessible to biologists without deep machine learning expertise.
The researchers envision that as cancer patient cohorts grow and are analyzed with single-cell technologies through initiatives like The Cancer Genome Atlas (TCGA), representation learning approaches like CellCnn will uniquely enable discovery of rare disease-causing or disease-tracking cell populations that shape both basic research understanding and personalized treatment decisions.