Cellector: A tool to detect foreign genotype cells in scRNAseq data with applications in leukemia and microchimerism.

This paper introduces Cellector, a computational tool designed to accurately detect rare foreign genotype cells in single-cell RNA sequencing data, enabling the identification of measurable residual disease in leukemia patients post-transplant and the analysis of microchimerism with high sensitivity.

The Gist

Imagine you are walking through a massive, bustling city made entirely of tiny, glowing cells. In this city, almost everyone looks exactly the same—they wear the same "genetic uniform" and speak the same "genetic language." This is your body's normal state.

But sometimes, a few "foreign agents" sneak into the city. These agents wear a slightly different uniform. They might be:

• Leaking in from a transplant: Like a new citizen moving into a neighborhood after a move.
• Hiding in a cancer patient: Like a spy (cancer cells) trying to blend in with the good guys after a surgery.
• Leftover from pregnancy: Like a tiny bit of a child's DNA lingering in a mother's body years later (microchimerism).

The problem? These "foreign agents" are incredibly rare. We are talking about finding one specific person in a crowd of 2,000 people.

The Problem with Old Tools

Previously, scientists tried to find these agents using two main methods, both of which had flaws:

1. The "Clustering" Method: Imagine trying to sort a crowd by asking everyone to stand in groups based on their clothes. If 99.9% of the crowd wears blue and only 0.1% wears red, the computer gets confused. It thinks the red group is just a weird variation of the blue group, not a separate team. It fails when the groups are so uneven.
2. The "ID Card" Method: Some tools require you to know the exact genetic ID card of the "bad guys" before you start looking. But in many medical cases (like leukemia relapse), we don't have that ID card yet, or getting it is too expensive and slow.

Enter: Cellector (The "Genetic Detective")

The paper introduces Cellector, a new digital detective tool designed specifically to find these rare "foreign agents" without needing to know their ID cards in advance.

Here is how Cellector works, using a simple analogy:

1. The "Voice Print" Check
Every cell has a unique genetic "voice print" based on tiny variations in its DNA. Cellector listens to the "voice prints" of every cell in the sample. It builds a profile of what the "majority voice" sounds like (the normal cells).

2. The "Outlier" Alarm
Instead of trying to sort everyone into groups, Cellector asks a simple question: "Does this cell's voice sound weird compared to the crowd?"

• If a cell sounds 99% like the crowd, it's ignored.
• If a cell sounds slightly different (even just a little bit), Cellector flags it as a "potential anomaly."

3. The "Iterative Cleanup"
This is the magic step. Cellector doesn't just flag one cell and stop. It says, "Okay, I found a weird cell. Let's pretend that cell doesn't exist for a moment and rebuild the 'normal crowd' profile without it." Then, it scans the crowd again.

• By removing the "weird" cells one by one, the definition of "normal" gets sharper and sharper.
• This allows it to find the "foreign agents" even when they are hiding in a crowd of 2,000 cells (0.05% of the total).

4. The Final Verdict
Once it has isolated the "weird" cells, it creates two distinct profiles: one for the "Good Guys" (Majority) and one for the "Foreign Agents" (Minority). It then gives every cell a probability score: "I am 99% sure this cell is a foreign agent."

Why This Matters in Real Life

1. Catching Leukemia Relapse Early
For leukemia patients who have had a bone marrow transplant, the goal is for their new marrow to be 100% from the donor. If even a tiny handful of the patient's old cancer cells remain, they can grow back and cause a relapse.

• Old way: You might miss these few cells until the cancer is huge and dangerous.
• Cellector way: It can spot those few "spy" cells when they are still just a tiny speck (0.05%). This allows doctors to treat the cancer before it becomes a full-blown emergency, potentially using gentler, safer drugs.

2. Understanding Pregnancy and Transplants

• Mothers and Babies: Cellector can find the tiny traces of a baby's cells that stay in a mother's body for decades, helping us understand how our immune systems evolve.
• Organ Transplants: It can track exactly which immune cells from a patient are attacking a new kidney, helping doctors understand why a transplant might be failing.

The Bottom Line

Think of Cellector as a super-sensitive metal detector at an airport. Old detectors might miss a tiny piece of metal if there are too many people around. Cellector is so sensitive it can find a single paperclip in a stadium full of people, even if that paperclip looks almost exactly like the other metal in the crowd.

This tool is free, open-source, and ready to help doctors and scientists find the "needles in the haystack" that could save lives.

Technical Summary

1. Problem Statement

The detection of rare, genetically distinct cells within a single-cell RNA sequencing (scRNA-seq) dataset is critical for several biological and clinical applications, including:

• Leukemia Relapse Detection: Identifying Measurable Residual Disease (MRD) in patients post-hematopoietic cell transplant (HCT). If patient-derived leukemia cells persist or return amidst donor-derived cells, it signals relapse.
• Microchimerism: Studying the presence of foreign cells (e.g., maternal cells in fetal tissue or vice versa) in natural biological contexts.
• Transplant Biology: Monitoring immune cell infiltration in solid organ transplants.

Limitations of Existing Tools:
Current computational methods for genotype-based cell demultiplexing face significant challenges in these scenarios:

• Clustering-based tools (e.g., souporcell, vireo, scSplit): These rely on clustering cells by genotype. They perform poorly when the minority cell population is extremely small (highly skewed cluster sizes), as the minority cluster is often indistinguishable from noise or merged into the majority cluster.
• Genotype-requiring tools (e.g., Demuxlet): These require prior knowledge of donor genotypes, which may be unavailable, costly to obtain, or impossible to acquire in certain clinical settings.
• Sensitivity vs. Specificity: Existing tools often trade high sensitivity for low positive predictive value (PPV) when detecting rare cells, leading to false positives that are unacceptable for clinical decision-making.

2. Methodology

Cellector is a computational tool designed to identify rare foreign genotype cells using a sparse beta-binomial anomaly detection framework. It operates under the assumption that the majority of cells in a sample share a single genotype, while a minority (<20%) may possess a foreign genotype.

Key Algorithmic Steps:

1. Variant Calling & Allele Counting:
   • The tool identifies expressed regions overlapping common human genetic variants (MAF >1%).
   • It counts alternative and reference alleles for each cell barcode using vartrix.
2. Beta-Binomial Modeling:
   • A beta-binomial distribution is constructed for each genetic variant based on the aggregate data of the sample.
   • Parameters $\alpha$ and $\beta$ represent the total counts of alternative and reference alleles, respectively, with a uniform prior added (Bayesian conjugate prior).
   • This distribution models the expected allele expression for the majority genotype.
3. Anomaly Scoring:
   • For each cell, the tool calculates the log-likelihood of its observed alleles given the majority genotype distribution.
   • This value is normalized to account for varying sequencing depths (number of reads covering variants) per cell.
4. Iterative Refinement:
   • Cells identified as statistical outliers (anomalous) are temporarily removed from the majority distribution parameters.
   • The anomaly detection is repeated iteratively until the set of detected foreign cells converges.
5. Final Classification:
   • Two distinct beta-binomial distributions are created: one for the majority cells and one for the identified minority cells.
   • A posterior probability is calculated for each cell to belong to either distribution, providing a final call.
6. Hybrid Strategy for Higher Prevalence:
   • If the minority population exceeds ~20%, anomaly detection loses sensitivity. In such cases, Cellector runs alongside souporcell (a clustering tool).
   • The tool selects the output that maximizes the log-likelihood difference between the two resulting distributions, ensuring optimal separation regardless of the minority fraction size.

3. Key Contributions

• Ultra-Low Detection Limit: Cellector can accurately detect foreign cells at frequencies as low as 0.05% (or even a single cell in a 10,000-cell dataset), a threshold where clustering methods fail.
• Related Individual Detection: It is specifically validated to distinguish cells from related individuals (e.g., parent-child, siblings), which is the most common scenario in HCT but the most difficult for genotype-based tools due to shared alleles.
• No Prior Genotype Required: Unlike Demuxlet, Cellector does not require pre-existing genotype data for the donors or recipients.
• High Specificity: The method prioritizes high Positive Predictive Value (PPV), which is crucial for clinical applications where false positives could lead to unnecessary aggressive treatment.

4. Results and Validation

The authors validated Cellector using both experimental mixtures (cell hashing) and in-silico mixtures with ground truth.

• Cell Hashing Validation (Real Mixtures):
  • Haploidentical (Mother/Child): In mixtures of 10,000 child cells spiked with 10, 50, or 250 maternal cells, Cellector achieved 100% PPV and sensitivities between 93–100%. It successfully detected all 3 maternal cells in the 10/10,000 spike-in.
  • HLA-Identical (Brothers): In mixtures of ~8,000–10,000 cells, Cellector maintained 100% PPV and 84–100% sensitivity, detecting all 4 cells in a 10/10,000 spike-in.
• In-Silico Validation (Ground Truth):
  • Tested across a wide range of spike-in levels for both related and unrelated individuals (using HipSci data).
  • Performance: PPV remained near 100% across thousands of runs. Sensitivity stayed above 90% for most spike-in levels, dropping to ~70% only when the number of spiked-in cells was extremely low.
• Comparison to State-of-the-Art:
  • Compared against souporcell, vireo, and demuxlet.
  • Cellector maintained high PPV (>90%) down to a single cell, whereas other tools showed low PPV until the minority fraction reached 1.5–3.5%.
  • Other tools often sacrificed PPV for sensitivity in low-frequency regimes, making them less suitable for clinical MRD detection.
• Robustness to Data Depth:
  • Downsampling experiments showed Cellector maintains near-perfect PPV and >85% sensitivity until the median UMIs per cell drop below 904 (for related individuals) or 2000 (for unrelated).

5. Applications Demonstrated

The paper demonstrates Cellector's utility in three distinct biological contexts:

1. Maternal/Fetal Cell Atlas: Successfully identified maternal macrophages and decidual stromal cells in placental samples (experimental contamination), distinguishing them from fetal cells despite different cell types.
2. Neuronal Microchimerism: Detected microchimeric cells in a neuronal sample where cells were expected to have differentiated similarly to the host tissue, confirming the presence of foreign genotypes via anomaly plots.
3. Kidney Transplant Biopsies: Analyzed samples at 5 and 28 days post-transplant. Cellector identified the infiltration of patient-derived immune cells (monocytes, macrophages, T cells) into the donor kidney, tracking the evolution of the immune response over time.

6. Significance

Cellector represents a significant advancement in scRNA-seq analysis for clinical and research settings involving rare cell populations.

• Clinical Impact: By enabling the detection of MRD at extremely low levels with high confidence, it facilitates earlier detection of leukemia relapse, potentially allowing for less invasive therapeutic interventions.
• Research Impact: It provides a robust method for studying natural microchimerism and transplant rejection mechanisms without the need for prior genotype knowledge or biochemical tagging.
• Accessibility: The tool is open-source (MIT license) and available on GitHub, promoting reproducibility and adoption in the community.