When a treated patient fails to respond, two very different things may have happened. The drug may have reached the tissue and failed biologically, or the patient may simply not have taken it. Clinical records cannot tell these apart, because pharmacy refills record that pills were possessed, not swallowed (Zimolzak et al.). Pathway enrichment analysis is the standard way to turn a long list of genes that differ between responders and non-responders into a short list of biological processes, by testing which curated gene groups ("pathways") are overrepresented among the top genes. It takes that non-responder label at face value. Our claim is that this mislabeling does more than add noise. Non-adherent "non-responders" carry an untreated transcriptome, so contrasting them against responders mainly recovers the drug's own on-target response pathways, and these are then reported as "resistance mechanisms." The more patients skip doses, the more the drug's mechanism of action leaks into the resistance pathway list.
We propose to treat each patient's actual drug intake as a hidden yes/no variable. Its prior probability comes from adherence evidence in the records: refill-based proportion of days covered, and adherence statements extracted from clinical notes. Its likelihood comes from a small, held-out set of genes that read out whether the drug engaged its target. Every downstream gene ranking and enrichment test is then computed with posterior intake weights, and the remaining uncertainty is carried through by repeated imputation of the hidden intake labels. In two sentences: enrichment pipelines currently mistake "didn't take the drug" for "drug didn't work," which plants the drug's own pathways into resistance findings. We correct the patient labels probabilistically before enrichment, using record-based adherence priors and target-engagement genes kept strictly separate from the genes being tested.
Core Claim and What Is New
The file's review found no work connecting adherence uncertainty to enrichment analysis, and the closest neighbours each stop at one side. Cousins et al. built gene set proximity analysis (GSPA), which expands gene sets through protein-network embeddings. They validated drug predictions against Medicare Advantage claims, but the enrichment itself ran on in-vitro CRISPR screens, and claims-based exposure was taken as given. Chen et al. model intake explicitly as a mediator between treatment assignment and outcome (the conditional front-door adjustment and their LobsterNet multi-task network). Their outcomes are scalar clinical endpoints, never high-dimensional expression or pathway scores. Biral & Dave cluster enrichment output with a text-and-gene-set contrastive model. That is pure post-processing, and it assumes the input rankings are clean. Kalka et al. name the prescription–consumption gap as a confounder of pharmacogenomic heritability but do no molecular analysis.
Our contribution differs from all four in three ways. It targets the upstream labels rather than the enrichment statistic or its output. It makes a specific, falsifiable prediction about the direction of the bias: on-target pathways leak into resistance lists. It treats intake as a latent class over a whole transcriptome rather than a mediator of one outcome.
Method
For patient $i$ recorded as treated, let $z_i \in \{0,1\}$ indicate actual intake. The prior $\pi_i = P(z_i = 1 \mid \text{records})$ comes from a calibrated adherence model. In its simplest form this is a logistic model on early-window proportion of days covered, following Zimolzak et al.. In its richer form it adds sentence-level adherence labels from notes using the ClinicalBERT classifier setup of the COLING 2020 adherence-extraction work.
Next, fix a target-engagement gene set $D$: genes known to respond to the drug, for example from a drug-signature database as used by Cousins et al. Compute a scalar engagement score $s_i$ on $D$, and fit a two-component mixture in which the $z_i=0$ component is anchored to untreated controls matched on covariates. The posterior is
$$ w_i = \frac{\pi_i\, f_1(s_i)}{\pi_i\, f_1(s_i) + (1-\pi_i)\, f_0(s_i)}, $$
where $f_1$ and $f_0$ are the engagement-score densities under intake and non-intake.
Differential expression for the resistance contrast (true non-responders versus responders) is then run only on genes outside $D$. We draw $M$ samples of $z$ from the posteriors, recompute gene rankings and enrichment for each draw, and combine pathway statistics across draws. The combination accounts for between-draw variance, so pathways whose significance depends on who was really treated are reported with wider intervals rather than false confidence. Excluding $D$ from testing is essential: without it, the engagement score would trivially decide the enrichment of on-target pathways. Pathways overlapping $D$ are reported separately as "not testable for resistance under this design." The enrichment engine itself is interchangeable, so we run Fisher's exact test, GSEA, and GSPA on the corrected rankings.
Research Program
The work has three parts.
The first establishes the leakage effect empirically. Using the semi-synthetic benchmark described below, we measure how often the drug's own pathways rank in the top resistance pathways under naive labels, as a function of the non-adherence rate.
The second builds and stress-tests the correction across prior quality. We vary the prior from uninformative, to refill-only, to refill plus notes, to oracle. We also vary the size and noise of $D$, which tells us whether the method degrades gracefully when adherence evidence is weak. That case matters in practice: Butte et al. describe outpatient adherence as close to a black box.
The third is an honesty check on circularity, comparing results when $D$ is chosen from external signatures versus learned from the same cohort.
Baselines
The naive reference point takes every recorded-treated non-responder at face value, runs standard differential expression, and then runs GSEA and Fisher's exact test, as implemented in standard tools or funkea. Every method sees the same expression matrices, the same gene-set library (KEGG via the GEO2KEGG mapping), and the same multiple-testing correction.
The strongest existing enrichment method on this compendium is GSPA. Cousins et al. report that it beat GSEA in area under the precision–recall curve (AUPRC) on 75.8% of GEO2KEGG datasets. We run GSPA with its published embedding radius on naive labels. Network expansion is the most plausible competitor that might absorb label noise without any adherence modeling. To beat it, our corrected pipeline must achieve higher AUPRC for planted resistance pathways and a lower leakage rate on a majority of datasets at non-adherence rates of 20% and above. We also report GSPA on corrected labels, to show the correction is complementary rather than a substitute.
Two covariate-based competitors run on identical data with the same hyperparameter search budget. The first is the clinical-filter baseline, which keeps only patients with proportion of days covered at or above 0.80. The second is covariate adjustment, which adds the adherence prior $\pi_i$ as a continuous covariate in the differential-expression model. The latter is the analogue of the standard backdoor adjustment that Chen et al. beat with a 20% reduction in heterogeneous-effect error (PEHE). Their number is for scalar outcomes. For enrichment under adherence contamination we found no reported result on Lacuna, so all enrichment comparisons will be new measurements.
Two ablations isolate the key ingredients. The prior-only variant uses $w_i = \pi_i$ and ignores expression. The expression-only variant uses a flat prior and thresholds the engagement score. Beating both shows that combining record evidence with molecular evidence is what matters.
Why not just…?
Why not just filter to patients with proportion of days covered ≥ 0.80? Possession is not ingestion, so filtered patients can still be untreated (Zimolzak et al.). Filtering also discards patients and therefore power. Even the best refill model in the file has 78.2% specificity and 86% positive predictive value at day 90, which leaves a meaningful contaminated fraction. The filter baseline runs at matched contamination levels, and we expect its leakage rate to fall but not to zero. If it matches our method when priors are refill-only, we will report that the expression likelihood adds nothing in that regime.
Why not just add adherence as a covariate in the expression model? This is the most serious threat. A covariate shifts mean expression linearly with $\pi_i$. Non-adherence, however, produces a mixture: a patient is either on-drug or off-drug, not 60% on-drug. A linear term cannot remove the off-drug component from the non-responder group. Our simulations generate mixture data, which favours our model by construction, so we also run a generator in which drug effect scales continuously with dose fraction. In that setting covariate adjustment should be competitive, and we will report where the crossover lies.
Why not just score the drug signature and drop low scorers? That is the expression-only ablation. It fails when target-engagement genes are noisy or few, and it gives hard labels with no uncertainty propagation. If it matches the full method, the honest conclusion is that record-based priors are dispensable whenever a good engagement signature exists. That would still be a useful finding, but it would shrink the contribution to the leakage diagnosis and the gene-splitting rule.
Why not just trust network-expanded enrichment to smooth over label noise? GSPA's expansion pulls in network neighbours of whatever genes rank highly. If those genes are drug-response genes, expansion should amplify leakage, because on-target pathways are dense network neighbourhoods. The GSPA-on-naive-labels baseline directly tests this prediction.
Proposed Experiments
The file contains no cohort with both transcriptomes and adherence records, so the main benchmark is semi-synthetic and built from real expression data. From each GEO2KEGG dataset used by Cousins et al., plus the four COPD cohorts (GSE5058, GSE8545, GSE11906, GSE20257) analysed by Lv et al., we take control samples as "untreated." We create "treated" samples by perturbing a designated KEGG pathway, which plays the role of the drug-response pathway. Within a subset of treated samples we then perturb a second, distinct pathway, the resistance pathway, so that these samples are true non-responders.
Non-adherent non-responders are injected by relabeling untreated samples as treated, at rates from 0% to 50%. Adherence priors are drawn from a model calibrated to the reported operating point of Zimolzak et al. (69.5% sensitivity, 78.2% specificity), with weaker and stronger variants around it.
We use two metrics. The first is AUPRC for ranking the planted resistance pathway. The second is the leakage rate, defined as the fraction of runs in which a pathway overlapping the drug-response pathway enters the top ten resistance pathways.
The first decisive check is whether naive GSEA and GSPA leakage rises monotonically with the non-adherence rate. If it does not, the core claim fails and the project stops there. The second is whether the full method keeps leakage near its 0% contamination level while matching or beating GSPA's AUPRC at 20–50% contamination. The third is the circularity test: results must hold when $D$ comes from an external signature rather than from the cohort itself.
A real-data stage would need a treated cohort that has both expression profiles and pharmacy or note records. We name this as the required next resource, not something the file supplies. The natural setting is metformin non-response, where Gao et al. and Kalka et al. document large unexplained, non-genetic variation in response.
References
- Early Detection of Poor Adherers to Statins: Applying Individualized Surveillance to Pay for Performance
- Gene set proximity analysis: expanding gene set enrichment analysis through learned geometric embeddings
- Conditional Front-door Adjustment for Heterogeneous Treatment Assignment Effect Estimation Under Non-adherence
- Contrastive Learning for Gene Set Enrichment Analysis Post-Processing
- Estimating heritability of glycaemic response to metformin using nationwide electronic health records and population-sized pedigree
- Extracting Adherence Information from Electronic Health Records
- Advancing evidence-based treatment of infectious diseases in children with real-world data: Opportunities and challenges
- funkea: Functional Enrichment Analysis in Python
- m6A RNA methylation regulators could contribute to the occurrence of chronic obstructive pulmonary disease
- The effect of oral diabetes medications on glycated haemoglobin (HbA1c) in Asians in primary care: a retrospective cohort real-world data study