Understanding what makes one patient respond to cancer treatment while another does not often comes down to the cells inside a tumor. But tumors are complex communities with many different cell types, and figuring out that mix can be difficult, especially across large groups of patients. Researchers from the University of Texas MD Anderson Cancer Center have developed a new computational method called scPER to tackle this challenge.

Single cell RNA sequencing can reveal the unique genetic activity of individual cells, but working with very large datasets or combining information from different research groups can introduce noise and confusion. scPER was designed to overcome these issues by building strong reference panels from many single cell datasets and filtering out misleading signals. It then applies advanced machine learning tools to estimate what kinds of cells are present in a tumor, even when only bulk RNA sequencing data is available.

Overview of scPER and the function of each stage

Stage 1. Reference integration and preprocessing. scRNA-seq datasets from multiple studies or tissues are integrated to form a reference matrix R (rows = genes, columns = cells). After QC and feature selection, denoising and imputation are applied to stabilize gene signals across datasets. Output: preprocessed single-cell expression R with cell-type labels and confounder labels (study/platform/tissue). Stage 2. Confounder-aware representation learning. An adversarial autoencoder learns a low-dimensional latent space Z from R. The encoder E maps expression x to z; the decoder D reconstructs x from z. An adversary V tries to predict designated confounders c (e.g., study/platform/tissue) from z. Training minimizes reconstruction while implicitly maximizing adversary loss with weight λ (i.e., L_rec − λL_adv), yielding Z that preserves biology but reduces nuisance variation. Output: cell embeddings Z and a trained encoder E for embedding new data. Stage 3. Bulk embedding and proportion estimation. Bulk RNA-seq samples T are normalized like the single-cell data and passed through the trained encoder E to obtain bulk embeddings in the same latent space. An XGBoost regressor is trained on latent features from single-cell references to predict cell-type compositions P for each bulk sample (columns = samples, rows = cell types). Output: estimated cell-type proportions P across all bulk samples. Stage 4. Downstream biology and clinical association. The estimated compositions P enable downstream analyses: (i) identify critical genes modulating the cell proportions, (ii) test clinical endpoints such as survival or response to immunotherapy, and (iii) identify phenotype-associated changes in cell types or subclusters. R = reference single-cell matrix; T = bulk expression; Z = scPER latent embeddings; P = estimated cell-type proportions; λ = adversarial weight; c = confounder label. Arrows indicate data flow from reference construction to embedding, deconvolution, and phenotype analyses.

The researchers showed that scPER performs better than several widely used tools when estimating the cellular makeup of tumors. More importantly, it helped identify specific cell groups linked to treatment outcomes. For example, in metastatic melanoma, scPER highlighted a rare T cell subcluster that strongly predicts how patients respond to immunotherapy. In metastatic urothelial cancer, it revealed how TGFβ signaling may suppress CD4 naïve T cells and reduce the effectiveness of PD L1 checkpoint inhibitors.

By improving how scientists interpret complex tumor samples, scPER may help researchers uncover the cellular clues that explain why some patients benefit from cancer therapies while others do not. Its success suggests that stronger computational tools can bring us closer to truly personalized cancer treatment.

Availability – scPER is an available tool on GitHub (https://github.com/BrianLlll/scPER).