QuantCell: machine learning based cell annotation integrating qualitative and quantitative imaging profiles

QuantCell is a machine learning framework that integrates quantitative imaging profiles with qualitative data to significantly improve cell annotation accuracy and coverage in complex spatial omics datasets, as demonstrated by its ability to increase annotated cells from 33.1% to 90.2% in mouse bone marrow while maintaining high confidence.

The Gist

Imagine you are walking through a massive, crowded city square at night. This square is filled with thousands of people (cells), but they are all wearing very similar dark coats. Your job is to identify who is a police officer, who is a doctor, who is a baker, and who is a rare, elusive street artist.

In the world of biology, this "city square" is a piece of tissue (like bone marrow), and the "people" are cells. Scientists use special cameras to take pictures of these cells, looking for tiny glowing badges (proteins) that tell them what job each cell does.

The Problem: The "Guessing Game" is Too Hard
Traditionally, scientists tried to identify these cells by looking at the pictures and saying, "Okay, if a cell has Badge A and Badge B, it's a doctor." But this is like trying to identify people in a crowd by only looking at their hats.

• It's slow: A human has to look at every single person one by one.
• It's incomplete: If a person is wearing a hat but no badge, or if the badge is faint, the human can't tell who they are. In the study, this method only identified about 33% of the people in the crowd.
• Rare people get missed: If the street artist only shows up once in a million people, you might never find them.

The Solution: Meet QuantCell (The Super-Intelligent Detective)
The authors of this paper created a new tool called QuantCell. Think of QuantCell as a super-smart AI detective that doesn't just look at what badges a cell has, but how bright those badges are, where they are on the cell, and the shape of the cell itself.

Here is how QuantCell works, using a simple analogy:

1. The "Ground Truth" Training Camp

First, the human experts (the detectives) look at a small, manageable group of people in the crowd. They carefully check their badges and say, "Yes, this one is definitely a doctor," or "This one is a baker." They label about 33% of the crowd with high confidence.

• The Twist: Instead of just stopping there, QuantCell takes these labeled people and uses them as a "training camp" for a computer.

2. Learning the "Quantitative" Clues

The computer doesn't just learn "Doctor = Badge A." It learns the nuance.

• It learns that a "Doctor" badge might be very bright in the center of the cell but dim on the edges.
• It learns that a "Baker" badge might flicker slightly differently than a "Police" badge.
• It looks at the whole picture: the size of the cell, the texture, and the background noise.

It's like teaching a child to recognize a cat. You don't just say "It has ears." You show them thousands of cats and explain, "See how the fur texture changes in the light? See how the tail moves?" The AI learns these subtle, quantitative details that humans might miss.

3. The "Safety Net" (False Discovery Rate)

This is the most clever part. The AI is allowed to say, "I'm not sure."

• If the AI sees a cell that looks mostly like a doctor but has a weird badge pattern, it doesn't guess. It leaves the cell blank.
• The researchers set a strict rule: "We will only label a cell if we are 95% sure we aren't making a mistake."
• This ensures that the final list of identified cells is incredibly accurate (96.5% accuracy!).

4. The Result: Filling in the Gaps

Because the AI is so good at spotting these subtle patterns, it can now look at the remaining 67% of the crowd that the humans couldn't identify.

• Before: Only 33% of the crowd was identified.
• After: QuantCell successfully identified 90% of the crowd!
• The Rare Findings: It even found the "street artists" (rare stem cells) that were previously invisible because they didn't fit the simple "hat" description perfectly.

Why This Matters

Imagine you are trying to fix a broken machine. If you only know what 33% of the parts are, you can't fix it. But if you know what 90% of the parts are, and you know exactly what the rare, critical parts are doing, you can solve the mystery.

In summary:

• Old Way: A human looking at a crowd, guessing who is who based on simple rules. Misses a lot, especially the rare people.
• QuantCell: A super-smart AI that learns from the human's best guesses, studies the fine details of light and shape, and then confidently identifies almost everyone in the crowd, while strictly avoiding mistakes.

This tool helps scientists understand complex diseases (like cancer) and how our bodies work by finally being able to "see" the rare and hidden cells that drive these processes.

Technical Summary

1. Problem Statement

Spatial omics technologies (e.g., PhenoCycler/CODEX, MIBI) enable high-resolution, multiplexed imaging of RNA and protein expression. However, cell annotation (assigning cell identities based on molecular profiles) remains a significant bottleneck, particularly in complex tissues like bone marrow (BM) containing diverse and rare cell populations.

• Limitations of Current Methods:
  • Manual/Semi-supervised approaches: Rely on labor-intensive manual identification of "ground truth" cells based on qualitative marker combinations. This is unfeasible for rare cell types (e.g., stem cells) or cells defined by complex marker combinations.
  • Existing ML Models: Many existing tools (e.g., Astir, AnnoSpat, MAPS) are often optimized for specific datasets, lack generalizability, or rely on clustering (unsupervised) which fails to detect rare populations in imbalanced datasets. Some require extensive manual annotation for training, which is a major bottleneck.
  • Data Underutilization: Conventional methods often discard cells that do not fit strict qualitative marker definitions, leaving a large fraction of cells unannotated.

2. Methodology: The QuantCell Framework

QuantCell is a machine learning (ML) framework designed to bridge the gap between qualitative (binary positive/negative) and quantitative (intensity-based) imaging data to improve annotation accuracy and coverage.

Workflow Steps:

1. Data Acquisition & Segmentation:
   • Tissue sections are imaged using multiplexed platforms (tested on PhenoCycler).
   • Cell segmentation is performed using QuPath (nuclei-based via StarDist or membrane-based via Cellpose-SAM).
2. Qualitative Ground Truth Generation:
   • A small subset of cells is manually annotated in QuPath based on predefined positive/negative marker combinations (qualitative profiles).
   • Cells matching these combinations exactly are labeled as "Ground Truth." Cells with ambiguous profiles are labeled "Other" or left unannotated.
3. Feature Extraction:
   • QuantCell extracts quantitative imaging features from the segmented cells.
   • Features include statistical descriptors (mean, median, min, max, standard deviation) of fluorescence intensity across multiple cellular compartments (nucleus, cytoplasm, membrane, whole cell) for each protein marker.
4. Model Training & Selection:
   • Training Set: 80% of the conventionally annotated cells are used to train ML models.
   • Testing Set: 20% is reserved for validation.
   • Model Evaluation: Multiple classifiers (Random Forest, MLP, Logistic Regression, etc.) are evaluated using Macro Average Precision (AP) via 10-fold cross-validation.
   • Optimization: The top-performing models undergo hyperparameter tuning. In the study, Random Forest emerged as the optimal model due to high performance and low runtime.
5. Prediction & Thresholding:
   • The best model predicts cell types for the remaining unannotated cells.
   • FDR Control: A user-defined False Discovery Rate (FDR) threshold (default 5%) is applied. Cells with prediction probabilities below this confidence threshold remain unannotated to ensure high-confidence results.

3. Key Contributions

• Integration of Qualitative and Quantitative Data: QuantCell uniquely leverages qualitative marker definitions to generate training data, then uses quantitative intensity features to classify cells that do not fit strict qualitative rules.
• Automated Ground Truth Generation: Unlike methods requiring manual annotation of thousands of cells, QuantCell automatically generates a training dataset based on predefined marker logic, eliminating the labor bottleneck for rare cell types.
• Robustness to Rare Populations: By using supervised learning on high-confidence ground truth, the framework avoids the "absorption" of rare cell types into larger clusters, a common failure mode of unsupervised clustering methods.
• User-Defined Confidence Control: The inclusion of an FDR-based thresholding strategy allows researchers to balance annotation coverage with classification confidence.
• Platform Agnosticism: The framework is designed to be compatible with various spatial omics platforms (2D/3D imaging, RNA/protein) and segmentation methods.

4. Key Results

The framework was validated using mouse bone marrow (BM) imaged with PhenoCycler (25 protein markers, 21 cell types, ~245k cells).

• Annotation Coverage: QuantCell increased the percentage of annotated cells from 33.1% (conventional method) to 90.2% at a 5% FDR.
• Accuracy: The model achieved 96.5% accuracy when compared to the conventional ground truth on the held-out test set.
• Feature Importance:
  • Whole-cell intensity provided the best predictive power, outperforming membrane-specific signals (likely due to overlapping signals in dense BM tissue).
  • Mean and Median intensity were more predictive than Minimum or Maximum.
  • Standard Deviation (signal variability) was more predictive than peak intensity, suggesting spatial heterogeneity carries biological signal.
• Benchmarking:
  • Compared against Astir, AnnoSpat, and MAPS, QuantCell demonstrated superior precision, particularly for rare cell types.
  • Runtime: QuantCell was 45 times faster than the second-best model (MAPS) while maintaining higher accuracy.
  • External Validation: Tested on three independent Hodgkin lymphoma datasets (MIBI and CODEX), QuantCell maintained high precision and significantly reduced runtime compared to MAPS.
• Cell Type Performance: Performance (F1 score) correlated with cell frequency and circularity but was robust across cell types defined by varying numbers of markers.

5. Significance

QuantCell addresses a critical gap in spatial biology analysis by providing a scalable, automated, and high-confidence solution for cell annotation.

• Enables Rare Cell Discovery: It allows for the robust detection of rare populations (e.g., hematopoietic stem cells) that are often missed by manual or clustering-based methods.
• Efficiency: It drastically reduces the manual labor required for data analysis, making large-scale spatial proteomics studies more feasible.
• Reliability: The FDR control mechanism ensures that biological conclusions are drawn from high-confidence annotations, reducing the risk of false positives in downstream analyses (e.g., cell-cell interaction studies).
• Generalizability: The framework's ability to adapt to different imaging modalities and marker panels makes it a versatile tool for the broader spatial omics community.

In conclusion, QuantCell represents a significant advancement in spatial data analysis, transforming qualitative marker definitions into a robust, quantitative machine learning pipeline that maximizes data utility while maintaining rigorous statistical confidence.