Untreated controls reveal that pre-treatment prediction of cancer cell fate mostly reflects growth fitness

Technical Summary: Untreated controls reveal that pre-treatment prediction of cancer cell fate mostly reflects growth fitness

Problem Statement
A central question in cancer biology is whether cells that survive chemotherapy possess a pre-existing, drug-specific "tolerance" state encoded in their transcriptome, or if survival is largely a stochastic outcome. Recent studies using sister-cell lineage tracing (specifically the ReSisTrace experiment) suggest that pre-treatment gene expression can predict which sister cell will survive treatment. However, a critical confounding factor remains: cells that survive treatment may simply possess high "growth fitness," allowing them to recover from the mechanical stresses of laboratory handling (synchronization, splitting, and re-plating) regardless of drug exposure. Without a properly controlled comparison, it is unknown how much of the predictive signal in pre-treatment transcriptomes reflects true drug tolerance versus general proliferative fitness.

Methodology
The authors re-analyzed the ReSisTrace dataset, which includes four experimental arms: carboplatin, olaparib, natural killer (NK) cells, and an untreated control. Crucially, the untreated arm undergoes the exact same synchronization, splitting, and re-plating procedures as the drug arms but lacks the therapeutic agent.

1. Experimental Design: In the ReSisTrace design, barcoded cells divide once; one sister is sequenced pre-treatment, and the other is exposed to treatment. The sequenced cell is labeled "pre-resistant" if its sister survives, and "pre-sensitive" otherwise.
2. Modeling Approach: The authors trained logistic regression models on 50 principal components (PCs) of gene expression to predict sister fate.
   • Drug-Arm Models: Trained on labels from specific drug arms.
   • Fitness Model: Trained exclusively on the untreated control arm (predicting recovery from handling stress only).
   • Comparison: The authors compared the performance of drug-arm models against the fitness model. They ensured fair comparison by matching training set sizes (sub-sampling drug arms to match the control arm's cell count) and using lineage-grouped cross-validation to prevent data leakage between related cells.
3. Advanced Validation:
   • Complex Models: Tested graph neural networks (GNNs), gradient boosting, and single-cell foundation models (Geneformer, scGPT) to see if non-linearities could extract more drug-specific signal.
   • Label-Free Scoring: Developed a "density score" based on the distance to nearest neighbors in expression space to test if the signal is a general property of cell state distribution.
   • Ceiling Estimation: Calculated the theoretical maximum predictability (AUROC ceiling) by measuring the correlation between pre-treatment sisters, acknowledging that sisters are not identical.
   • External Validation: Applied findings to three other datasets: a melanoma multi-treatment study (Schaff et al.), a lung cancer study (Watermelon), and a triple-negative breast cancer study. Additionally, analyzed bulk tumor data (TCGA-OV) and patient pre/post-chemotherapy single-cell data to check for clinical consistency.

Key Results

1. Predictive Power of Fitness: A simple logistic regression model trained only on the untreated control arm (the "fitness model") achieved an AUROC of 0.70–0.73 when applied to drug-arm labels. This is nearly identical to the performance of models trained directly on the drug arms (AUROC 0.72–0.75).
2. Minimal Drug-Specific Gain: The incremental gain in AUROC provided by training on drug-specific labels over the fitness model was small (0.014–0.025) and, in many cases, statistically indistinguishable from zero (95% confidence intervals included zero). The upper bound of this drug-specific signal was estimated at ~0.05 AUROC.
3. Model Complexity: More complex models (GNNs, transformers) did not improve performance, suggesting the information limit lies in the data (sister discordance and noise), not the algorithm.
4. Nature of the Signal:
   • Shared Signal: The shared predictive signal corresponds to a proliferative, low-stress state (high MYC, mTORC1, E2F; low stress/apoptosis pathways).
   • Drug-Specific Signal: The small drug-specific component points to a slower cycling state among otherwise fit cells, rather than a distinct resistance mechanism.
   • Clone Size Dominance: In a second lineage system (melanoma), clone size (a proxy for fitness) predicted survival better than gene expression under all six treatments.
5. Clinical Consistency: In patient samples receiving chemotherapy, the "fitness signature" (proliferative state) was depleted in post-treatment malignant cells, not enriched. This contradicts the hypothesis that the pre-treatment fitness state predicts clinical survival, suggesting that states predicting recovery in a dish do not necessarily predict survival in a patient.

Key Contributions

• Re-evaluation of Pre-treatment Predictors: The study demonstrates that most of the signal in pre-treatment transcriptome-based fate prediction is driven by general growth fitness (recovery from handling stress) rather than specific drug tolerance.
• Methodological Benchmark: The authors introduce a rigorous evaluation framework requiring an untreated control arm processed identically to drug arms. They provide a Python package, persist.fitadj, to calculate the "fitness-adjusted" gain (ΔAUROC) for any lineage-tracing design.
• Ceiling Analysis: The paper quantifies the theoretical limit of predictability due to sister-cell discordance, estimating that even a perfect model of the sequenced cell cannot exceed an AUROC of ~0.84 in drug arms due to biological noise.
• External Generalization: The findings are validated across multiple cancer types and lineage tracing designs, showing that clone size often outperforms expression-based predictors.

Significance and Claims
The authors conclude that claims of pre-existing drug-tolerant states must be tempered by the realization that "fitness" is a dominant confounder. Without an untreated control arm, models trained on drug-arm data likely learn a "fitness predictor" rather than a "drug-resistance predictor."

The paper argues that:

1. Benchmarks are flawed: Current benchmarks for single-cell drug response (e.g., scDEAL) that lack untreated controls may reward models for learning growth fitness rather than drug tolerance.
2. Experimental Design: Future studies must include an untreated control arm processed through the same protocol to isolate drug-specific effects.
3. Biological Interpretation: The "persister" state identified in culture may largely reflect a robust, proliferative state that recovers from stress, rather than a specialized, drug-resistant phenotype. The small drug-specific component identified is real but minor and difficult to resolve without larger sample sizes.

The study does not claim that drug tolerance does not exist, but rather that the predictable component of cell fate in these specific lineage-tracing experiments is overwhelmingly driven by growth fitness, and that separating the two requires the specific control design proposed.