Disrupting cellular memory to overcome drug resistance

Gene expression memory refers to how long a specific transcriptional state persists within an individual cell or lineage and can span hours to weeks. In cancer, intermediate-timescale memory underlies phenotypes such as stemness, differentiation, metastasis, and drug resistance, where states persist across divisions yet remain switchable. In melanoma, drug-naive cells fluctuate between a drug-susceptible state and a primed state that predisposes to resistance upon targeted therapy. Preventing resistance may be possible by driving primed cells into a drug-susceptible state, but the cues that trigger switching and the ability to track switching over longer timescales are limited. Existing bulk and single-cell approaches either lack temporal resolution or cannot infer state memory over days to weeks. The authors aim to develop a robust single-cell lineage-tracing method to quantify memory and identify regulators of state switching in melanoma, with the goal of therapeutically modulating state transitions to reduce resistance.

Prior work has associated intermediate gene expression memory with key cancer phenotypes, including drug tolerance and resistance. In melanoma, rare primed states characterized by low MITF and high AXL correlate with early resistance and worse progression-free survival. Bulk RNA-seq-based memory measurements captured persistence but not switching drivers, while scRNA-seq captures heterogeneity without temporal persistence information. Short-timescale methods (e.g., RNA velocity, time-resolved scRNA-seq) resolve hours, whereas longer timescales require lineage tracking. High-throughput cellular barcoding paired with scRNA-seq has enabled linking lineage to transcriptomes. Multiple studies implicate EMT-like programs and TGF-β signaling in melanoma phenotype switching and resistance; PI3K signaling intersects with receptor tyrosine kinase pathways and non-canonical TGF-β signaling, and co-targeting MAPK and PI3K pathways reduces resistance but is limited clinically by toxicity.

The authors developed scMemorySeq, combining high-complexity lentiviral lineage barcoding (transcribed 100 bp semi-random barcodes in GFP 3′ UTR) with scRNA-seq to quantify gene expression memory and detect state switching. In WM989 (BRAF V600E) melanoma cells, they transduced barcodes, sorted primed cells (EGFR+/NGFR+) and mixed controls, expanded ~4 doublings (12–14 days), and performed 10x Genomics 3′ scRNA-seq with parallel PCR-based barcode recovery to assign lineages. They identified drug-susceptible (SOX10/MITF-high) and primed (EGFR/AXL/NT5E-high) clusters, validated heterogeneity, and proposed NT5E as a marker capturing the entire primed cluster. Functional validation used NT5E flow sorting followed by vemurafenib to assess resistant colony formation. They profiled in vivo PDX tumors using HCR RNA FISH for NT5E and SOX10, and analyzed additional melanoma line lines (WM983B) and patient scRNA-seq datasets for primed signatures. They used lineage composition to classify lineages as maintaining a state or switching between drug-susceptible and primed. Differential expression between drug-susceptible cells from non-switching lineages vs. those in lineages that later switch identified an intermediate state; they computed intermediate state scores and applied GSEA to implicate EMT and TGF-β pathways. A stochastic two-state model estimated proliferation and switching rates from barcoded scRNA-seq. To causally test modulators, they conducted a 4-condition split experiment: after barcoding and ~7–8 doublings, cells were split and treated 5 days with TGFB1, TGFBRi (LY2109761), PI3Ki (GDC-0941), or untreated, followed by scRNA-seq with barcode linkage. They quantified state fractions and lineage-specific switching across matched barcodes. They further constrained models with live-cell imaging of proliferation (H2B-GFP IncuCyte tracking). Flow cytometry (NT5E-APC) quantified primed fraction changes across treatments in WM989 and WM983B; additional ligands (EGF, BDNF, IL6) tested PI3K activation effects. Finally, pretreatment scheduling studies applied 5-day TGFB1, TGFBRi, or PI3Ki, followed by 4 weeks of BRAF/MEK inhibition (dabrafenib + trametinib) to quantify resistant cells and colonies, with normalization by post-pretreatment cell number. Computational analysis used Cell Ranger, Seurat (SCTransform, UMAP, Louvain clustering), UCell for signature scoring, custom barcode processing (STARCODE clustering, 10x Feature Barcode), and custom stochastic modeling.

• scMemorySeq resolved two robust transcriptional states in WM989 melanoma: a drug-susceptible state (SOX10, MITF high) and a primed state (EGFR, AXL, NT5E high), with memory over multiple divisions and rare state switching. • NT5E identified the entire primed cluster; NT5E-high sorted cells produced 5.5-fold more resistant cells than NT5E-low upon vemurafenib treatment. • Primed-like cells (NT5E-high/SOX10-low) were detected in WM989 PDX tumors in mice, in WM983B cultures, and in 5/7 analyzed drug-naive human melanoma biopsies; primed signatures overlap published resistance-associated states (e.g., MITF-low/AXL-high) linked to shorter PFS. • Lineage modeling estimated proliferation and switching: primed cells divide at ~half the rate of drug-susceptible cells; drug-susceptible cells switch to primed once every ~135–233 divisions; primed cells revert to drug-susceptible once every ~5–8 divisions, indicating higher stability of the drug-susceptible state. • An intermediate transcriptional state precedes priming: 575 DE genes distinguished drug-susceptible cells that later switch vs. those that do not (Bonferroni-adjusted p<0.05, |log2FC|≥0.25). GSEA implicated EMT, hypoxia, UV response down, IFNγ/TNFα signaling, and TGF-β. Intermediate cells retained SOX10/MITF yet upregulated primed markers (e.g., FN1, SERPINE2); unique genes included NFATC2, MGP. • TGF-β signaling modulates entry into the primed state: TGFB1 (5 days) increased primed fraction in WM989 from 1.98% to 19.15% and in WM983B from 10.04% to 81.56%. TGFBRi alone had minimal effect on primed maintenance but blocked TGFB1-induced priming. • PI3K pathway constrains the primed state: PI3Ki (GDC-0941, 2 µM, low-toxicity) reduced primed fraction in WM989 from 1.98% to 0.31% and in WM983B from 10.04% to 3.66%, and blocked TGFB1-induced priming; other PI3K-activating ligands (EGF, BDNF, IL6) did not induce priming, suggesting TGFB1’s unique dual activation of PI3K and SMAD pathways. • Barcode-matched lineage analysis showed TGFB1 increases switching into primed among lineages drug-susceptible at baseline; PI3Ki drives switching out of primed among lineages primed at baseline (6/8 fully primed lineages switched; overall, 93% of lineages reduced primed fraction under PI3Ki). • Pretreatment strategy: 5-day PI3Ki before BRAF/MEK reduced resistant colonies by 62% and resistant cells by 57% versus BRAF/MEK alone; normalized to post-pretreatment cell number, resistant cells decreased by 36%. Co-treatment with PI3Ki and BRAF/MEK nearly eliminated resistance but is clinically limited by toxicity.

The study addresses the central question of how to quantify gene expression memory, identify regulators of state switching, and exploit these dynamics to mitigate drug resistance. scMemorySeq integrates lineage barcodes with scRNA-seq to directly observe persistence and transitions between drug-susceptible and primed states in melanoma. By focusing on lineages that switch, the authors identified an intermediate EMT-like state enriched for TGF-β signaling, implicating these pathways as early drivers of priming. Causal perturbations confirmed that TGFB1 accelerates switching into primed, whereas PI3K inhibition promotes reversion to the drug-susceptible state. Modeling constrained by proliferation measurements indicates that switching rates, not differential growth or death, best explain treatment-induced shifts. Translationally, brief PI3K inhibitor pretreatment reduced resistance to subsequent MAPK-targeted therapy, suggesting a practical alternative to concurrent combination therapy that may alleviate toxicity. The findings establish that global signaling modulation can tune cell-state memory to sensitize heterogeneous tumor populations and highlight a generalizable framework for rational scheduling to precondition cell states prior to therapy.

This work introduces scMemorySeq, a lineage-tracing single-cell framework to quantify gene expression memory and detect state switching. In melanoma, the method delineated drug-susceptible and primed states, discovered an EMT/TGF-β–enriched intermediate state, and identified TGF-β and PI3K as key modulators of switching. Functionally, brief PI3K inhibitor pretreatment reduced resistance to subsequent BRAF/MEK inhibition, supporting a strategy of state modulation before targeted therapy. Future research should extend scMemorySeq to other cancers and microenvironmental contexts, refine dosing schedules and durations for state-switching agents, validate biomarkers (e.g., NT5E) prospectively in patients, and test clinically whether pretreatment regimens can balance efficacy with toxicity better than concomitant combinations.

The study is preclinical and primarily utilizes melanoma cell lines and a PDX model; clinical efficacy and safety of pretreatment regimens were not tested. Some lineages did not respond to PI3K inhibition within the 5-day window, suggesting variable kinetics of switching or dosing dependencies. The fully primed baseline lineage set was small (n=8), limiting statistical power for some switching analyses. TGFBR inhibition did not reduce primed fraction, indicating pathway redundancy or compensatory signaling. Co-treatment with PI3K and MAPK inhibitors, although effective ex vivo, is limited by clinical toxicity; the extent to which brief pretreatment mitigates toxicity while sustaining benefit remains to be established. Generalizability beyond melanoma and across diverse tumor microenvironments requires further validation.