🧬 biology

DuoDose: homotypic-aware doublet detection by integrating identity and dosage evidence

DuoDose is a Python-based tool that improves homotypic doublet detection in single-cell RNA sequencing by integrating identity and dosage evidence into a calibrated random forest model, thereby reducing false positives from high-RNA singlets while outperforming existing baselines.

Original authors: Yongqi Huang

Published 2026-08-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yongqi Huang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a detective trying to solve a mystery inside a bustling, microscopic city. This city is a drop of blood or a piece of tissue, and the citizens are individual cells. In recent years, scientists have developed a super-powerful camera called "single-cell RNA sequencing" that lets them take a photo of every single citizen's activity log (their genes) at once. This helps us understand how our bodies work, how diseases start, and how to build better medicines.

But there's a glitch in the camera. Sometimes, the lens gets a little sticky, and two cells get squished together into one tiny bubble. The camera thinks this is just one super-cell, but it's actually a "doublet"—a mix of two different people. If the two cells are from different neighborhoods (like a skin cell and a blood cell), it's easy to spot the mistake; the mixed-up log looks weird and obvious. But what if the two cells are twins? What if they are from the exact same neighborhood and look almost identical? These "homotypic doublets" are the masterminds of the crime. They hide in plain sight, looking just like a normal, healthy citizen, but they have a secret: they have twice as much "stuff" (RNA) as they should.

The problem is that some real, healthy cells are naturally big and busy, carrying a lot of stuff too. If you just look for "lots of stuff," you might accidentally arrest these innocent, hard-working citizens while letting the twin-criminals go free. Scientists have been trying to build a better detective to tell the difference between a naturally busy cell and a fake twin, but it's been a tough case.

Enter DuoDose, a new digital detective created by researcher Yongqi Huang. Instead of just looking at one clue, DuoDose is a smart system that checks two things at once: Identity (who the cell says it is) and Dosage (how much stuff it's carrying). Think of it like a security guard at a club. A normal guard might just check your ID. If you look like a VIP, you get in. But DuoDose is smarter. It checks your ID and counts how many drinks you're holding. If you have a VIP ID but are holding a tray with 50 drinks (which is impossible for one person), DuoDose knows you're a fake twin, even if you look exactly like a real VIP.

The paper shows that DuoDose is incredibly good at solving this specific type of mystery. In a massive test involving 15 different datasets (like 15 different crime scenes) and running the test 75 times with different random seeds, DuoDose caught the "twin" doublets much better than any other existing detective. Specifically, it found 69 out of 75 times that it was better at spotting these tricky twins than the next-best method.

Here is the magic trick: DuoDose doesn't just guess. It was trained to recognize four distinct types of "citizens":

Normal Singlets: The regular, honest citizens.
High-RNA Singlets: The naturally busy, big citizens who are innocent but have lots of stuff.
Homotypic Doublets: The twin-criminals who look like normal citizens but have double the stuff.
Heterotypic Doublets: The obvious mixed-up criminals (different cell types).

By teaching the system to specifically recognize the "naturally busy" citizens as innocent, DuoDose avoids the common mistake of arresting them. In the tests, when the system was set to catch 50% of the twin-criminals, it accidentally flagged only 21.1% of the innocent, busy citizens as suspects. Compare that to other methods, which accidentally flagged between 39.5% and 56.9% of the innocent citizens! DuoDose is much more precise, keeping the innocent safe while catching the twins.

The researchers also tested DuoDose on a real dataset called "cline-ch." They found that DuoDose could spot twin-criminals hiding deep inside the most crowded, normal-looking neighborhoods—places where other detectives missed them completely. While other tools only looked for obvious mix-ups, DuoDose found the subtle ones that were hiding in plain sight.

The paper doesn't claim this is a magic bullet that solves every problem in the world. The "twins" in the test were created by the computer by mathematically adding two cells together, which is a simulation, not a real biological event. However, the results are very strong: the system is reproducible, it runs fast on normal computers (taking about 20 seconds for a large dataset), and it consistently outperforms the competition in these controlled simulations.

In short, DuoDose is a new, highly tuned tool that helps scientists clean up their data. It ensures that when they study how cells behave, they aren't accidentally studying a fake twin. By carefully balancing the clues of "who you are" and "how much you have," it protects the innocent busy cells and catches the sneaky twins, making our understanding of the microscopic world a little bit clearer and a lot more accurate.

Technical Summary: DuoDose

Problem Statement
Single-cell RNA sequencing (scRNA-seq) doublet detection faces a fundamental asymmetry. While heterotypic doublets (formed from transcriptionally distinct cells) often appear as outliers or intermediate states in expression space, homotypic doublets (formed from similar cells) can remain "identity-space inliers" within coherent clusters. Consequently, detectors relying primarily on identity-mixture signals often miss homotypic doublets. Conversely, methods that rely heavily on library size or RNA abundance to identify doublets risk misclassifying natural high-RNA singlets (e.g., large, active, or cycling cells) as doublets. The operational challenge is to distinguish dosage abnormalities indicative of homotypic doublets from naturally high-RNA singlets without discarding biologically informative cells.

Methodology
DuoDose addresses this by integrating identity and dosage evidence into a calibrated four-class random forest model. The framework operates through three main stages:

Fitted-Reference SafeFeatures: A SafeFeatureTransformer is fitted on a designated reference pool to learn gene selection, normalization, dimensionality reduction (40-component truncated SVD), clustering, and reference-neighbor structures. This state is frozen and applied to validation and test cells to generate a 24-dimensional feature vector. This vector includes:
- Identity Evidence: Measures of whether a cell lies within a coherent reference identity or resembles a mixture.
- Dosage Evidence: Total counts and detected gene complexity.
- Library-Complexity Balance: A feature ( $b_i$ ) describing the balance between count expansion and detected-gene expansion, defined as $\log(1 + g_i) - \frac{1}{2}\log(1 + c_i)$ .
- Artificial-Doublet Bank: An internal reference of simulated doublets.
Controlled Semi-Real Benchmarking: The model is trained without experimental doublet annotations. Instead, "semi-real" doublets are constructed by pairwise summation of raw counts from observed cells. To prevent data leakage, parent cells are partitioned into disjoint fit, validation, and test pools before doublet construction, and the exact parent cells are removed from the background of their respective splits.
- Four-Class Supervision: The model explicitly learns to distinguish between: (O) Ordinary singlets, (R) High-RNA singlets, (H) Homotypic doublets, and (T) Heterotypic doublets.
- Weighted Training: High-RNA singlets are treated as a "hard-negative" class with a weight of 2.0 to penalize confusion with doublets.
Calibrated Inference: A four-class random forest (240 trees) estimates probabilities for each class. These probabilities are calibrated using a sigmoid calibrator on a held-out validation set. The overall doublet score is the sum of the calibrated homotypic and heterotypic probabilities.

Key Contributions

Unified Four-Class Model: DuoDose is the first method to explicitly model high-RNA singlets as a distinct supervised class alongside ordinary singlets and both doublet subtypes, allowing the model to learn the boundary between high-abundance biology and doublet artifacts.
Leakage-Controlled Benchmark: The study employs a rigorous "parent-disjoint" construction where training, validation, and test pools share no parent cells, preventing identity leakage and providing a robust evaluation of generalization.
Identity-Inlier Detection: By combining identity-mixture signals with context-adjusted dosage features, the method targets homotypic doublets that remain embedded within singlet clusters.

Results
Evaluated across 15 datasets and 75 independent runs (5 seeds per dataset), DuoDose demonstrated superior performance in homotypic detection while maintaining competitive heterotypic detection:

Homotypic Sensitivity: DuoDose achieved a homotypic AUPRC of 0.635 ± 0.105, significantly outperforming the strongest external baseline (scDblFinder, 0.367; scds, 0.414) with a mean paired gain of +0.208. It won in 69/75 runs against the best external method.
High-RNA Protection: At a matched 50% homotypic recall, DuoDose produced the lowest false-positive rate (FPR) for high-RNA singlets (0.211 ± 0.073) compared to baselines ranging from 0.395 to 0.569. It achieved the strictly lowest FPR in 63/75 runs.
Overall Performance: The overall AUPRC was 0.715 ± 0.101, outperforming all comparators (Scrublet: 0.466; scDblFinder: 0.631; DoubletFinder: 0.536; scds: 0.641).
Heterotypic Detection: While not uniformly the best for heterotypic doublets (AUPRC 0.519 vs. scDblFinder's 0.604), DuoDose maintained competitive performance, preserving an informative heterotypic score while expanding homotypic sensitivity.
Real-Data Application: In the cline-ch dataset, DuoDose identified experimentally annotated doublets within identity-dense regions that were under-ranked by other methods, confirming its ability to detect doublets that lack strong mixture signals.

Significance and Claims
The paper claims that DuoDose provides a practical route to recover homotypic-aware doublet candidates that are systematically under-ranked by identity-geometry alone. By unifying identity-mixture and identity-inlier dosage evidence in a single calibrated model, it substantially extends homotypic sensitivity without resorting to indiscriminate library-size thresholds that remove biologically informative high-RNA cells. The author emphasizes that the model's ability to separate homotypic doublets from high-RNA singlets is driven by the explicit supervised learning of the high-RNA class and the integration of library-complexity balance, rather than simple count thresholds. The method is implemented in Python, released under the MIT License, and designed for reproducibility with frozen reference transformations.

Technical Summary: DuoDose

More like this