🧬 biology

Blast percentage confounds transcriptome to drug-response associations in Beat AML: DHODH and sorafenib as a worked example

This study demonstrates that failing to adjust for blast percentage in the Beat AML dataset significantly confounds transcriptome-drug response associations, as evidenced by the substantial loss of statistical significance for markers like DHODH and sorafenib when blast content is included in the analysis.

Original authors: Cumhur Gokhan Ekmekci

Published 2026-11-01

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Cumhur Gokhan Ekmekci

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the fight against acute myeloid leukemia, a type of blood cancer, scientists often look for clues within the patient's own cells to predict which drugs will work best. They examine the transcriptome, which is the complete set of genetic instructions being read and used by the cells at a specific moment. By comparing these genetic patterns to how the cancer cells react to different medicines in a lab dish, researchers hope to find a "signature" that tells them which treatment to choose. This approach relies on the idea that the genetic makeup of the cancer cells themselves holds the key to their vulnerability. However, a sample of bone marrow or blood is rarely made up of only cancer cells. It is a mixture containing the cancerous blasts, healthy blood cells, and other tissue. The proportion of cancer cells in that mix can vary wildly from one patient to another, and this variation has long been a source of confusion in interpreting genetic data.

A recent study by independent researcher Cumhur Gokhan Ekmekci investigates a specific problem that has been overlooked in a major dataset known as Beat AML. This dataset pairs genetic information from leukemia patients with detailed records of how their cells responded to 125 different drugs. While previous analyses of this data adjusted for factors like how fast the cells were dividing or how mature they were, they did not account for the percentage of cancer cells, or "blasts," present in the sample. The researcher set out to see if ignoring this simple fact was distorting the results, leading scientists to believe certain genes were important drug targets when they were actually just markers of how many cancer cells were in the dish.

The investigation began with a simple question: does the amount of cancer in a sample change the way we read the genetic data and the drug test results? The answer turned out to be a definitive yes. The study found that the percentage of blasts in a sample correlates with both the genetic signals being measured and the strength of the drug's effect. When a sample has a high number of cancer cells, the genetic signals associated with rapid growth appear stronger, and the drug appears more effective simply because there are more target cells to kill. This creates a false connection. A gene might seem to predict a good response to a drug not because it controls the drug's action, but simply because it is more active in samples that happen to have more cancer cells.

To measure how big this problem was, the researcher ran a massive comparison involving 12,000 different combinations of genes and drugs. They took the standard way of analyzing the data, which ignores the blast percentage, and compared it to a new way that included the blast percentage as a factor. The results were striking. Under the standard method, nearly 1,200 gene-drug pairs appeared to be significant matches. When the blast percentage was added to the calculation, more than 400 of those matches disappeared. In other words, about 35 percent of the connections that looked strong were actually illusions created by the composition of the sample. Conversely, only a tiny fraction of new connections appeared when the adjustment was made. This suggests that the previous list of promising drug targets was significantly inflated by the unmeasured variable.

The study used a specific example to show exactly how much a single candidate marker could change. Researchers had previously identified a gene called DHODH as the strongest predictor of how well a drug called sorafenib would work. In the standard analysis, this gene ranked first among all 125 drugs tested, with a very strong statistical link to the drug's success. However, once the percentage of blasts was included in the model, the strength of that link dropped noticeably. The gene fell from the top rank to fourth place, and the statistical certainty of the connection weakened considerably. While the link did not vanish completely, it became much less impressive. Furthermore, when the researchers looked at other drugs designed to target the same pathway as sorafenib, none of them showed a similar connection to the DHODH gene. This indicates that the original strong result was likely a side effect of the sample composition rather than a true biological mechanism.

The researcher also checked to ensure that this change was not caused by other factors, such as the type of sample taken or the way the drug response was measured. None of those other variables moved the results significantly. The only factor that altered the outcome was the inclusion of the blast percentage. The study also looked at whether this finding held up in a different group of patients from a separate database, but the results there were not strong enough to confirm the pattern, likely because that data lacked the specific clinical details needed for the same kind of analysis. This does not disprove the finding but highlights that the initial result needs careful re-evaluation before it can be trusted.

The conclusion of this work is that the percentage of cancer cells in a sample is a critical piece of information that must be included when analyzing genetic data against drug responses. It is a factor that is already recorded in medical records and costs nothing to include in the analysis. Ignoring it leads to a high rate of false discoveries, where scientists might chase genes that are merely reflecting the amount of cancer present rather than the cancer's true sensitivity to a drug. By adjusting for this variable, the list of potential drug targets becomes shorter and more reliable. The study does not claim to have found a new cure or a definitive marker for sorafenib; instead, it serves as a necessary correction, showing that the path to finding real biomarkers requires cleaning the data of these hidden confounding factors first. The cost of including this simple adjustment is negligible, but the cost of leaving it out is a significant distortion of the scientific picture.

Technical Summary: Blast Percentage Confounds Transcriptome to Drug-Response Associations in Beat AML

Problem Statement
The Beat AML project provides a unique resource pairing ex vivo dose-response data for 125 compounds with transcriptomic and genomic data from the same patients. This dataset is widely mined to identify transcriptional markers of drug sensitivity. Standard analytical conventions in this field adjust for general drug sensitivity (using residual response), differentiation state, proliferation, and genotype. However, these analyses typically omit the blast percentage (the proportion of leukemic blasts in the specimen) as a covariate.

While a recent four-cohort consistency study concluded that adjusting for blast percentage was unwarranted due to a lack of observed association, this paper argues that blast percentage is a critical confounder. In acute myeloid leukemia (AML), the non-blast fraction dilutes both the transcriptome signal and the ex vivo assay readout. Consequently, transcripts associated with the proliferating blast compartment may appear to correlate with drug response simply because drugs kill blasts, not because of a specific molecular mechanism. The paper posits that failing to adjust for this composition inflates false-positive marker discovery.

Methodology
The study utilized data from Beat AML waves 1–4, comprising 520 samples with both expression and drug data.

Models: The author compared a "base model" (adjusting for median drug response, differentiation axis, proliferation score, and FLT3-ITD status) against an "extended model" that added blast percentage as a covariate.
Scale Comparison: To assess the generalizability of the confounding effect, the author tested 12,000 gene-by-drug pairs (800 randomly sampled protein-coding genes against the 15 most tested compounds). Crucially, the sample set was fixed per drug for both models to ensure differences were not driven by changes in sample size ( $n$ ).
Worked Example (DHODH & Sorafenib): The analysis focused on DHODH (dihydroorotate dehydrogenase) as a candidate marker for sorafenib sensitivity. The author performed a genome-wide calibration using expression-matched null distributions to determine the empirical significance of the DHODH-sorafenib association under both models.
Statistics: Associations were calculated using Spearman correlations on ranks. Significance was determined via Benjamini-Hochberg False Discovery Rate (FDR) control. Survival analyses used Cox proportional hazards.

Key Results

Blast Percentage as a Confounder: Blast percentage correlated significantly with both DHODH expression ( $\rho = +0.353$ ) and residual sorafenib response ( $\rho = -0.285$ ). This mechanical relationship arises because specimens with higher blast content contain more target cells and higher levels of proliferation-linked transcripts.
Impact on Association Discovery: When blast percentage was added to the model:
- 35% of the gene-drug associations that were significant under the base model (FDR < 0.1) lost significance.
- Only 3.7% of the associations gained significance.
- This asymmetry (loss vs. gain) was more pronounced than in solid tumor purity analyses (which showed ~14% loss and ~11% gain).
- The effect varied by compound; for example, significant associations for sorafenib dropped from 121 to 42, while trametinib remained relatively stable (238 to 222).
The DHODH-Sorafenib Case Study:
- Base Model: DHODH expression was the strongest correlate of sorafenib response ( $\rho = -0.259$ , FDR $6.5 \times 10^{-7}$ ), ranking sorafenib first among 125 drugs. The empirical $p$ -value against a matched null was $< 0.001$ .
- Extended Model (+ Blast): The correlation attenuated to $\rho = -0.155$ (using marrow blast) or $-0.186$ (using specimen-matched blast). Sorafenib's rank dropped to 4th, and the empirical $p$ -value rose to 0.012.
- Selectivity: Despite attenuation, the association remained selective for sorafenib among FLT3-directed compounds; other FLT3 inhibitors (e.g., quizartinib, gilteritinib) showed no significant correlation after adjustment.
Ruling Out Alternatives: The author tested other potential confounders (specimen type, relapse status, raw vs. residual AUC, mitochondrial content scores) and found that only blast percentage significantly altered the estimates.
Replication Failure: The prognostic association of DHODH with overall survival observed in Beat AML did not replicate in the TCGA-LAML cohort, likely due to differences in treatment (standard induction vs. sorafenib) and missing covariate data.

Significance and Claims
The paper claims that blast percentage belongs in the adjustment set for transcriptome-to-drug-response analyses in Beat AML.

Cost of Omission: The omission of this covariate costs approximately one-third of the associations reported as significant, creating a high rate of false-positive candidate markers.
Nature of the Artifact: The confounding is bidirectional: blast percentage affects the transcriptome (by increasing the abundance of proliferation-linked genes) and the assay readout (by increasing the number of killable target cells).
Specificity of the DHODH Finding: While the DHODH-sorafenib association is weakened by adjustment, it is not eliminated. The author suggests the residual signal may be pharmacological (related to de novo pyrimidine synthesis competition with cytarabine or mitochondrial priming via MCL1) rather than purely compositional, though they explicitly state that a mechanism is not established and the effect size is small.
Practical Implication: The author argues that including blast percentage is low-cost (as it is already in diagnostic records) but high-value, as it prevents the propagation of inflated markers. They propose that future marker claims in Beat AML should be recomputed with this covariate, and that deconvolution-based estimates of blast fraction could serve as a more robust alternative where morphological data is missing.

The paper concludes modestly, emphasizing that the reported attenuation is a statement about how the data behaves under different models, not a demonstration of causality, and that DHODH should not be proposed as a clinical selector based on these findings alone.

Technical Summary: Blast Percentage Confounds Transcriptome to Drug-Response Associations in Beat AML

More like this