New Benchmark REACH Exposes How Rare Cancer Cell Detection Methods Collapse Under Pressure

Deep inside a tumor sample, some of the most clinically consequential cells are also the hardest to see. Circulating tumor cells drifting through blood, minimal residual disease cells surviving after treatment, and tiny malignant compartments buried within a heterogeneous tumor microenvironment can each account for one percent or less of a dataset. Single-cell RNA sequencing can, in principle, isolate these populations one cell at a time, but the statistical challenge is brutal: a detector must pull a handful of malignant cells out of a vast non-malignant background, and the label “malignant” itself may rest on copy number evidence, expression signatures, somatic mutations, neighborhood structure, or expert annotation that do not always agree. A team at SRM University-AP in India has now built a benchmark designed to settle, once and for all, how well current computational methods actually perform on this task — and the results are sobering.

The benchmark, called REACH for Rare-cell Evaluation Across Cancer Heterogeneity, is described in Frontiers in Bioinformatics by Moram Venkata Satya Jaswanth, Amruthaluri Gavin, Padarthi Aakash, Muttha Tejesh Kumar, and Abhijit Dasgupta. Rather than asking which method is best in the abstract, REACH formulates rare malignant cell detection as an imbalanced ranking problem. Methods are expected to assign higher scores to high-confidence malignant cells — copy number variation-supported, tumor-intrinsic cells designated P_HC — than to high-confidence background cells designated B_HC, which include immune, stromal, and normal epithelial cells. This malignant-versus-microenvironment contrast deliberately models the circulating tumor cell and minimal residual disease screening regime, and it explicitly excludes the discovery of rare transcriptional substates within an already-malignant population, such as drug-tolerant persisters or cells undergoing epithelial–mesenchymal transition.

What makes REACH unusual is its insistence on reproducibility and multi-track evaluation. The pipeline runs through twelve audited phases, from environment setup to standardized preprocessing, confidence-tiered label construction, benchmark unit generation, method execution under a common wrapper contract, statistical testing, and figure regeneration directly from audited outputs. Every method must produce a continuous cell-level score plus run metadata recording runtime, memory, success or failure status, and degenerate-output flags. Methods that time out or exhaust memory receive fallback random scores that are filtered from aggregation; methods that emit constant or near-constant outputs are retained but flagged. The audited snapshot evaluates ten computational methods on ten curated cancer datasets containing 359,739 cells, producing 11,250 performance measurements across 1,125 benchmark tasks.

The evaluation tracks probe different aspects of method behavior. Track A, the primary endpoint, samples real high-confidence malignant cells into a mock background at four prevalence tiers ranging from 5 percent down to 0.01 percent. Track B applies synthetic Splatter simulations fitted to each real dataset as a secondary stress test. Track C uses background-dominated null controls where no true rare malignant cells are expected, functioning as a specificity pass/fail diagnostic. Track D preserves the natural prevalence of blood and circulating tumor cell datasets. Track E fixes Track A expression matrices while corrupting labels to measure supervised noise sensitivity. A sixth diagnostic track, Track F, masks the global tumor microenvironment contrast entirely and tests whether methods can recover rare malignant epithelial cells hidden among normal epithelial cells of the same tissue origin — a much harder within-lineage problem.

The headline result is a striking gap between supervised and unsupervised performance. A supervised logistic regression trained on 2,000 highly variable genes achieves a median average precision of 1.0 on Track A, indicating that many high-confidence malignant and background labels contain gene expression differences that a supervised model can trivially recover. Current unsupervised methods recover only part of this signal. FiRE, the strongest published inter-lineage detector, reaches a median average precision of 0.313. CaSee, an exploratory cancer/normal cell scorer rather than a dedicated rare-cell detector, reaches 0.512 on all units and 0.606 after excluding 20 degenerate outputs. Statistical testing with a Friedman test and Iman–Davenport correction confirms significant performance differences among methods, and pairwise Wilcoxon tests with Benjamini–Hochberg false discovery rate correction show that FiRE significantly outperforms DeepScena, scCAD, RareQ, and CellSIUS, though not the naive expression threshold baseline or scMalignantFinder.

REACH also reveals strong context dependence. No method performs best across all tumor types. CaSee is the best non-ceiling method on four datasets, the naive expression threshold baseline on two, CellSIUS on one, and scMalignantFinder on one. The supervised ceiling itself collapses on the ovarian cancer ascites dataset, where several unsupervised methods outperform it. This dataset-specific variability justifies reporting method performance in relation to tumor context rather than a single “best method” story, and it cautions against extrapolating leaderboards across cancers.

The null-control Track C results are equally important for interpretation. All eight unsupervised, exploratory, and naive scoring methods control their false-positive rate near the expected 3 percent floor set by the random baseline, and pairwise tests show they are statistically indistinguishable from it. The supervised ceiling is the sole outlier, with a false-positive rate of 0.788 due to in-sample overfitting on all-negative data. This convergence shows that the moderate Track A average precision values of outlier-based methods do not arise from indiscriminate over-flagging of background cells, but from an inability to rank malignant cells above the tumor microenvironment subpopulations these methods legitimately surface as outliers.

The most consequential finding comes from Track F. When the background is restricted to normal epithelial cells of the same lineage, nine of ten methods fall to near-chance average precision at 0.1 percent prevalence. The supervised ceiling itself drops from 1.000 on Track A to 0.001, outputting constant scores because the inter-lineage contrast on which it was trained has been masked. Cluster-decomposition methods such as scCAD show inverted rankings, with mean AUROC significantly below 0.5, indicating that normal epithelial background cells are being ranked systematically above the malignant positives. The sole exception is scMalignantFinder, a pretrained pan-cancer classifier that retains perfect performance — but this result is colorectal-cancer-specific and cannot be generalized. The practical conclusion is stark: methods whose Track A ranking rests on the inter-lineage malignant-versus-microenvironment contrast lose essentially all discriminative power once that contrast is removed.

The authors are careful to frame the supervised ceiling as an inter-lineage separability reference rather than a target that unsupervised rare-cell finders should reach. The near-perfect Track A performance reflects massive inter-lineage expression variance; a lineage-masking scorer comparing epithelial markers against immune and stromal markers would be expected to approach this bound. Three independent lines of evidence — the ceiling’s dataset-specific collapse, its Track F collapse, and its degradation under label noise — indicate that biological separability is the primary driver, with conservative label construction playing a secondary role. The gap between the ceiling and unsupervised methods therefore reflects a real capability gap, not an artifact of label design.

REACH’s limitations are acknowledged openly. The current snapshot relies on infercnvpy as the sole copy number inference arm, though a circularity check confirms that CNV is used purely for labeling and not as a scored comparator, and CNV burden itself discriminates malignant from non-malignant cells with an overall AUROC of 0.782 across 57,811 cells. Track F is restricted to a single colorectal cancer dataset because it is the only one carrying both epithelial cell-type annotations and normal/tumor tissue origin labels. Most datasets use 10× Chromium technology, and spatial transcriptomics and multi-omics data are not yet included. Method stability remains a practical issue, with 249 fallback units and 377 degenerate units recorded across the benchmark. The authors also note that public leaderboards carry overfitting risk and recommend hidden test datasets and containerized execution in future versions.

For clinicians and computational biologists, the message is clear. Inter-lineage leaderboards should not be extrapolated to intra-lineage rare-cell discovery. Detecting drug-tolerant persisters, epithelial–mesenchymal transition intermediates, or emergent resistant subpopulations within an already-identified malignant compartment remains a distinct and currently unsolved problem that will require methods modeling within-lineage structure explicitly rather than global rarity or anomaly. REACH provides a transparent platform for future method comparison and a practical reference for developing more robust malignant cell detection tools — and, on the evidence of this benchmark, extending that reference from the inter-lineage regime to the intra-lineage regime is the more consequential open problem.

Subject of Research: A reproducible benchmark for rare malignant cell detection in single-cell RNA sequencing

Article Title: REACH: a reproducible benchmark for inter-lineage rare malignant cell detection in single-cell RNA sequencing

Article References: Satya Jaswanth, M. V., Gavin, A., Aakash, P., Tejesh Kumar, M., & Dasgupta, A. (2026). REACH: a reproducible benchmark for inter-lineage rare malignant cell detection in single-cell RNA sequencing. Frontiers in Bioinformatics, 6, Article 1878538. https://doi.org/10.3389/fbinf.2026.1878538

Image Credits: AI Generated

DOI: 10.3389/fbinf.2026.1878538

Keywords: single-cell RNA sequencing, rare malignant cell detection, REACH benchmark, computational biology, cancer bioinformatics, tumor microenvironment, circulating tumor cells, minimal residual disease, reproducible benchmarking, machine learning, REACH, reproducible