Can single-cell transcriptomics enable massive parallel assessment of genetic variants?

SCRAMseq replaces the bespoke assay with a transcriptome readout

Multiplexed assays of variant effect (MAVEs) test many variants at once, but most versions need a purpose-built phenotypic readout for each gene, such as cell survival or the expression of one surface marker. That requirement is why MAVEs have spread faster in cancer and common-disease genetics than in rare disorders, where the number of diagnosed patients is small and assay development is hard to justify. The paper's premise is that a single readout also risks collapsing a protein's full activity into one number.

SCRAMseq (Single Cell RNAseq Associated with MAVEs by sequencing) addresses both problems. A balanced library of missense variants is introduced so that each cell carries no more than one gene copy, cells expand for more than 72 hours, and single cells are then processed by full-length scRNA sequencing. Each cell yields two linked records: RNA genotyping identifies which variant it carries, and its transcriptome describes the phenotype. The variant's effect is summarized as a tolerance score derived from a gene-expression signature, so the assay does not depend on any preselected marker.

A 2,265-variant TP63 library read through keratinocyte conversion

The authors focused on TP63, a transcription factor that drives stratified epithelial development. ClinVar variants in TP63 are 35% variants of unknown significance, and 73% of those are missense; known pathogenic missense changes cluster in the DNA binding domain (DBD) and the sterile alpha motif (SAM) domain. The team mutagenized amino acids 260–319 of the DBD and 513–568 of the SAM domain, producing 2,265 variants, or 99.69% of all possible substitutions in those regions.

To give the variants a biologically meaningful context, they optimized conversion of human BJ-T fibroblasts into induced keratinocyte-like cells by co-expressing P63 and KLF4 from a lentiviral vector. Conversion efficiency exceeded 70% within 14 days, and flow sorting on ITGβ4, a surface marker of basal keratinocytes and a P63 target, measured it at 75%. Bulk RNA sequencing found more than 1,400 differentially expressed genes enriched for keratinocyte, skin, and epidermis terms. This conversion setup is the gene-specific reference against which the generalizable SCRAMseq readout is judged.

Single-mutant assays confirm the neutral and pathogenic split

The authors picked individual mutants from each predicted class and tested them with independent methods. Four neutral variants (G293H and G310Q in the DBD; T527I and H556A in the SAM) and fourteen pathogenic variants were analyzed by quantitative immunofluorescence of KRT14, flow cytometry of ITGβ4, RNA sequencing, native PAGE, and a luciferase displacement reporter.

The separation was consistent. Only neutral mutants activated KRT14 and converted fibroblasts; no pathogenic variant did. Every pathogenic SAM mutant showed aggregation, with aggregates fourfold more abundant than monomers, while DBD mutants stayed monomeric like wild type. Neutral mutants reached 85–90% KRT14-positive cells, comparable to wild-type P63, except G310Q at about 60%. G310Q also stood apart transcriptomically: it lacked whole blocks of genes induced by wild-type P63, including KRT5, KRT14, and ITGβ4, and showed roughly sevenfold higher luciferase expression than wild type in the displacement assay. The paper treats G310Q as a neutral-class variant with measurably reduced activity rather than as a clean negative.

SCRAMseq agrees with the gene-specific assay at 84.2%

The central comparison is between SCRAMseq and MITE, the established workflow that sorts converted cells by ITGβ4 and reads variants by bulk amplicon sequencing. Using MITE as a functional reference rather than a clinical ground truth, the authors report 84.2% overall concordance, made up of 53.8% concordant pathogenic-pathogenic calls and 30.4% concordant neutral-neutral calls. Classification agreement reached an AUC of 0.91 in the conversion setup and 0.92 in the non-conversion setup.

The remaining discordance is attributed to what the two methods measure: MITE separates cells on ITGβ4 expression, whereas SCRAMseq uses a broad transcriptomic signature. The paper also reports that the experimental scores outperformed in silico predictors, including large language models, on Spearman correlation, AUC-based correlation, and concordance. That comparison supports the value of functional data, but it is a benchmark against predictors, not evidence that SCRAMseq replaces clinical classification.

Score distributions separate variant classes across the library

Across all classified variants, the tolerance-score distribution for neutral variants differed significantly from that of pathogenic variants (Mann–Whitney U, P < 2.2e-16). Neutral variants had a median score above zero; pathogenic variants had a median below zero. The authors then took the five best-represented neutral variants (L568A, W559F, Y535C, D524K, D543N) and five pathogenic variants (I562G, F513P, I537P, I558S, L545N) and compared their score distributions, again finding a significant difference (P = 1.6e-15).

The pattern was asymmetric. Pathogenic mutants were almost entirely unable to activate the signature, while neutral mutants activated it across a wide range of scores. The authors link that spread among neutral variants to the limited time allowed for overexpression and to the unbiased design of the setup. In practical terms, a low score is a strong signal of loss of function, whereas a high score does not guarantee wild-type-equivalent activity.

MYOD1 shows the workflow is not TP63-specific

To test whether SCRAMseq depends on the TP63 conversion system, the authors built a MITE library covering the first stretch of MYOD1, a developmental transcription factor, for about 1,300 variants. They deliberately chose an N-terminal region outside known functional domains and ran it in the same non-selective overexpression conditions used for the TP63 non-conversion setup.

In the single-cell experiments, genotyped cells numbered 2,331 for P63 converted, 1,216 for P63 non-converted, and 1,836 for MYOD1 non-converted. Detected variants were 1,108 of 2,228 for P63 converted, 719 of 1,056 for P63 non-converted, and 790 of 1,306 for MYOD1 non-converted. The MYOD1 run therefore demonstrates feasibility in a second gene and a domain without a known functional assay, but it is a demonstration of applicability rather than a clinical validation of MYOD1 variant interpretation.

What the evidence supports and where it stops

The strongest claim the data support is methodological: a full-length single-cell transcriptome can serve as a general phenotypic readout for MAVEs, and for TP63 it reproduces the gene-specific MITE classification closely enough to be useful. The two-gene, two-context design supports portability across experimental settings and genes better than a single-gene study would.

Several boundaries matter for interpretation. MITE is a functional reference, not clinical ground truth, so the 84.2% concordance measures agreement between assays rather than diagnostic accuracy. Variant recovery was partial in every run, and the number of genotyped cells per experiment ranged from 1,216 to 2,331, so coverage differs by condition. The authors also note that current MAVEs, including this one, still rely on relatively simple phenotypic readouts and that the neutral-class spread reflects the short overexpression window. The paper's own framing is that SCRAMseq broadens where MAVEs can be applied, not that it removes the need for orthogonal validation of individual variants.

The priority the paper establishes is a trade: give up the gene-specific selection assay and gain a transcriptome-wide readout that can be pointed at any disease-driving gene. For TP63, that trade costs about 16% classification agreement with MITE; for a gene with no established functional assay, such as the MYOD1 region tested here, there is no MITE baseline to lose. That asymmetry, rather than the headline concordance figure, is the practical threshold readers should carry away.

Sources used in this answer