Immune checkpoint inhibitors release a patient's own T cells against the tumor, and the same release frequently attacks healthy tissue: these immune-related adverse events (irAEs) affect 56.2% of patients within a year in a 6,526-patient real-world cohort, and their onset is associated with a 19-fold increase in mortality hazard on combination therapy (A real-world cohort study of immune-related adverse events in patients receiving immune checkpoint inhibitors). The biology of this toxicity is sex-biased in a puzzling way: females develop more irAEs overall (odds ratio 1.15 in that cohort), while males develop more of the often-lethal cardiac ones (univariate odds ratio 1.17, with roughly 40% fatal outcomes, in FAERS pharmacovigilance data on dual checkpoint blockade; Cardiac adverse events associated with dual immune checkpoint inhibitors: a pharmacovigilance analysis from the FDA adverse event reporting system). Nobody can explain this split, because the two relevant data worlds never meet: safety databases like FAERS carry adverse event reports but no tumor biology, and tumor genomic repositories like TCGA carry microenvironment measurements but no adverse event outcomes.
This proposal builds the missing bridge, and the bridge is deliberately simple: Pearson correlation — the standard measure of straight-line association between two continuous variables — computed separately in males and females. The core move is to treat sex not as one binary column in a pooled regression, but as the axis along which two continuous measurements are taken: on one side, organ-system-level irAE signal strengths mined from FAERS with standard disproportionality statistics; on the other, quantitative immune infiltration scores and myeloid-suppression gene axis activities deconvolved from bulk tumor transcriptomes. If the correlation structure between toxicity and microenvironment genuinely differs by sex, then pooling the sexes should destroy it — and that destruction, which we can measure, is the falsifiable heart of the proposal.
Core Claim
The claim is threefold. First, within non-reproductive solid tumors (lung, hepatocellular, colorectal, melanoma, head and neck), the linear association between immunosuppressive myeloid axes — the MYBL2–CCL2 macrophage recruitment axis established in ovarian cancer (The MYBL2–CCL2 axis promotes tumor progression and resistance to anti-PD-1 therapy in ovarian cancer by inducing immunosuppressive macrophages) and the tumor-intrinsic PPAR$\gamma$/MDSC axis established in hepatocellular carcinoma (IDDF2020-ABS-0201 Targeting hepatoma-intrinsic pparγ signaling overcomes immune checkpoint therapy resistance by inflaming the tumor microenvironment) — and macrophage or suppressor-myeloid infiltration differs in strength between male and female patients. Second, this sex-differential microenvironment wiring is detectably aligned, at the cancer-type level, with the organ-specific sex bias of irAE signals in FAERS: we predict the female overall-irAE excess tracks T-cell-inflamed correlation fingerprints, while the male cardiac excess tracks myeloid-inflammatory ones. Third, sex-aggregated analysis of the same data dilutes both associations below detection, in the same way that pooling male and female rats hid a female-only anemia signal in preclinical toxicology (Acute and subacute oral toxicity assessment of Gancao Xiexin decoction in Sprague-Dawley rats).
The mechanistic premise comes from sex-differential cancer genomics: roughly 15% of autosomal genes show sex-biased copy number alterations, and transcription factor networks wire into distinct pathways in males versus females, which means the slope of any gene-to-cell-type relationship need not be shared across sexes (Genome-Wide Sex and Gender Differences in Cancer). A pooled covariate adjustment assumes a shared slope with an intercept shift, which is exactly the assumption this proposal drops.
Research Program
The program has three arms that share data partitions, so every method sees identical patients, reports, and strata.
Arm A builds the toxicity layer. We extend the consensus disproportionality design of the donepezil and eculizumab FAERS studies — where a signal counts only if it clears thresholds on the reporting odds ratio (ROR), proportional reporting ratio (PRR), Bayesian confidence propagation neural network, and the Multi-Item Gamma Poisson Shrinker simultaneously (Comprehensive safety assessment of donepezil: pharmacovigilance analysis based on the FDA adverse event reporting system; Adverse drug events (ADEs) risk signal mining related to eculizumab based on the FARES database) — to every organ system, every single and dual checkpoint regimen, and every report quarter from 2015 onward, computed separately within male and female report strata. The output is a matrix of sex-specific signal strengths indexed by organ system and cancer type, in the spirit of the dual-ICI cardiac analysis but exhaustive rather than cardiac-only.
Arm B builds the biology layer. On TCGA bulk transcriptomes for the non-reproductive tumors above, we run TIMER-style purity-corrected deconvolution, which estimates abundance of six immune subsets while controlling the collinearity and the "purity paradox" — infiltrating immune transcripts misattributed to tumor cells — that corrupt uncorrected bulk analysis (Digitally deconvolving the tumor microenvironment). We then score each tumor for the MYBL2–CCL2 axis and the PPAR$\gamma$ pathway signature, residualize expression and infiltration against age and tumor purity, and compute Pearson correlations within each sex. The sex difference is tested with a Fisher transformation:
$$z_s = \tfrac{1}{2}\ln\frac{1+r_s}{1-r_s}, \qquad Z = \frac{z_m - z_f}{\sqrt{\tfrac{1}{n_m-3}+\tfrac{1}{n_f-3}}}$$
where $r_s$ is the sex-specific partial correlation between axis score and infiltration fraction, $z_s$ its Fisher-transformed value, and $n_m$, $n_f$ the male and female sample sizes. This is the standard normal approximation for comparing two independent correlations, and it converts our claim into a per-tumor, per-axis hypothesis test with a $p$-value we can correct for multiplicity.
Arm C is the bridge, and it is honest about being ecological: FAERS reporters and TCGA patients are different people, so the linking correlation runs across cancer types, not patients. For each cancer type and organ system we correlate the sex-specific FAERS signal strength from Arm A against the sex-specific microenvironment quantities from Arm B, then test stability by leaving each cancer type out in turn and permuting sex labels to destroy any true structure. The Stropharia multi-omic study is the methodological template here: it showed that Pearson correlation between continuous transcript levels and continuous metabolite abundances can isolate a handful of governing genes from 6,726 candidates (Integrative Analysis of Transcriptome and Metabolome Sheds Light on Flavonoid Biosynthesis in the Fruiting Body of Stropharia rugosoannulata); we run the same style of analysis with irAE signal strength in the metabolite's role and infiltration scores in the transcript's role.
Proposed Experiments
Experiment 1 reproduces the pooled world. On FAERS, pooled disproportionality and a multivariable logistic regression with sex as a binary covariate, exactly as in the OneFlorida+ study design, give the reference numbers: female sex OR 1.15 for any 1-year irAE, and the male cardiac OR 1.17 from the dual-ICI analysis. On TCGA, pooled (sex-aggregated) partial correlations between myeloid axis scores and infiltration give the reference biology numbers. We expect the pooled correlations to sit near zero.
Experiment 2 tests signal recovery under stratification. Within each sex stratum we recompute everything and ask two questions: how many organ-system–cancer-type cells gain a consensus-clearing signal that the pooled analysis missed, and how large is the gap between stratified and pooled correlations, in Fisher-$z$ units, after false-discovery-rate correction across the multiple tests. A masking calibration on the Gancao Xiexin pattern — an effect present in one sex only — tells us quantitatively how much dilution pooling causes at each sample size, so we know what effect sizes the design can and cannot see. This matters because stratification halves effective sample sizes, a known power cost of sex-disaggregated analysis (Genome-Wide Sex and Gender Differences in Cancer).
Experiment 3 is the head-to-head against covariate regression. Both approaches are asked the same predictive question: rank organ systems by expected male-versus-female irAE bias. The covariate model produces one sex coefficient per cohort; ours produces per-organ association maps tied to microenvironment quantities. We compare rank correlation with the observed sex ratios held out from fitting. If the covariate model ranks as well, the stratified machinery adds nothing for prediction and we will say so; its remaining value would be explanatory only.
Experiment 4 is the deconvolution ablation. We rerun Arm B correlating raw bulk expression of CCL2, MYBL2, and PPAR$\gamma$ against raw checkpoint transcripts, with no purity correction, and measure how many sex-differential correlations flip sign or vanish. The purity paradox and molecular mimicry documented by Aran and Butte predict spurious associations in exactly this configuration (Digitally deconvolving the tumor microenvironment); the file reports no benchmark performance number found on Lacuna for TIMER, so we establish our own reference rather than cite one.
Experiment 5 probes the ecological bridge from the biology side. In the GEO anti-PD-1-treated melanoma cohort (GSE78220) and the FUSCC triple-negative breast cancer cohort of 360 patients (AI-powered omics-based drug pair discovery for pyroptosis therapy targeting triple-negative breast cancer), we test the within-cohort, sex-stratified correlation between myeloid axis activity and infiltration. These cohorts lack systematic irAE annotation, so this experiment can confirm the biology half (sex-differential axis wiring) but not the toxicity half; we state that limit openly.
Baselines
The naive reference point is the do-nothing pooled analysis: FAERS disproportionality computed over all reports regardless of sex, using the exact consensus thresholds of the donepezil study, and pooled TCGA correlations without sex strata. It is the strongest competitor for the claim that stratification matters, because it is what every cited pharmacovigilance study except the sex-focused ones actually does. It runs on the identical FAERS extract and TCGA snapshots, with the same consensus thresholds and the same false-discovery correction, so any difference is attributable to stratification alone. Beating it means recovering sex-specific organ signals it misses, judged by the Fisher-$z$ gap in Experiment 2, and by demonstrating on the masked-signal calibration that it provably cannot detect one-sex-only effects at the cohort sizes in question.
The strongest existing methodological competitor is multivariable logistic regression with sex as an adjusted covariate, the design used in the OneFlorida+ study of 6,526 ICI-treated patients, which reports female sex OR 1.15 for 1-year irAE incidence alongside combination-therapy OR 1.56 and breast-cancer OR 2.48 (A real-world cohort study of immune-related adverse events in patients receiving immune checkpoint inhibitors). It is the incumbent because it already found the headline sex effect, it handles confounders by construction, and it requires no new data linkage. We run it on the same FAERS-derived organ-stratified tables with age, regimen, and cancer type as covariates, and with the same compute (all methods here are statistical, so fairness is about identical data partitions rather than tuning budgets; Pearson and regression have no hyperparameters beyond thresholds we fix in advance from the cited studies). Beating it is defined by Experiment 3: better rank prediction of organ-level sex bias, or — the more likely and still publishable outcome — equivalent risk prediction plus a continuous microenvironment linkage the regression cannot express, which would reposition the contribution as explanatory rather than predictive.
The key ablation isolates purity-corrected deconvolution: raw bulk expression correlations versus TIMER-style corrected infiltration scores, everything else held fixed (Experiment 4). No reported head-to-head accuracy result was found on Lacuna for TIMER versus uncorrected bulk correlation, so the ablation itself establishes the number. A second ablation replaces the continuous Pearson layer with binarized contingency tables at the consensus thresholds, isolating what continuous measurement contributes over the categorical standard.
Why not just…?
Why not just adjust for sex as a covariate in one regression? For average risk, this works — it is how the OR 1.15 was found — and we treat it as our strongest baseline rather than argue it away. Where it falls short is structure, not averages: a covariate assumes the same gene-to-cell-type slope in both sexes with only an intercept shift, while genome-wide evidence shows roughly 15% of autosomal genes with sex-biased copy alterations and sex-dimorphic regulatory networks (Genome-Wide Sex and Gender Differences in Cancer), and the rat toxicology study shows pooling can erase a one-sex effect entirely (Acute and subacute oral toxicity assessment of Gancao Xiexin decoction in Sprague-Dawley rats). Experiment 3 settles this directly: if the covariate model predicts organ-level sex bias as well as the stratified maps, the covariate approach wins and we report that.
Why not just run standard disproportionality with chi-square tests, as everyone does? We do run it — Arm A is exactly the consensus machinery of the donepezil and eculizumab papers, and for pure signal detection it needs no replacement. What it cannot do is the linking: contingency tables binarize reports, collapsing the continuous gradients of chemokine expression and infiltration that carry the microenvironment information, while Pearson correlation preserves them — the Stropharia study recovered 8 structural genes and 21 transcription factors from continuous data that binarized thresholds would have flattened (Integrative Analysis of Transcriptome and Metabolome Sheds Light on Flavonoid Biosynthesis in the Fruiting Body of Stropharia rugosoannulata). The contingency ablation makes this concrete: if binarized versions of our pipeline produce the same sex-differential map, the continuous layer is decorative and can be cut.
Why not just train a black-box model on all features? If the goal were predicting which individual patient develops toxicity, a structured model would plausibly win — in the TNBC drug-pair study, a biologically structured network reached a C-index of 0.90 and beat random forests and logistic regression (AI-powered omics-based drug pair discovery for pyroptosis therapy targeting triple-negative breast cancer). But that study had 360 patients with matched omics and outcomes; our FAERS side has no per-patient tumor data at all, so there is no patient-level feature vector to train on. The goal here is association mapping with coefficients a clinician can read, not prediction, and Pearson plus Fisher-$z$ is the right tool for that goal. This is a genuine scope limit, not a defense: we do not claim predictive gains.
Why not just correlate raw bulk expression and skip deconvolution? Because the purity paradox is documented, not hypothetical: uncorrected bulk analysis attributes infiltrating immune transcripts to tumor cells, and tumor cells themselves aberrantly express checkpoint molecules like CTLA-4, faking infiltration that is not there (Digitally deconvolving the tumor microenvironment). Experiment 4 quantifies the damage; if uncorrected correlations match the corrected ones, deconvolution is unnecessary here and we will report that too.
Decisive Checks
Three results would make the case; their failure would kill it. Permuting sex labels in Arm B and Arm C must destroy the sex-differential correlation structure — if structure survives permutation, it is an artifact of the pipeline. At least several organ-system–cancer-type pairs must show a significant Fisher-$z$ sex difference in correlation surviving false-discovery correction, with the female-excess organ systems (endocrine, skin) loading on T-cell-inflamed fingerprints and the male-excess cardiac signals loading on myeloid ones; a random assignment of fingerprints to organ systems falsifies the hypothesis. And the ecological bridge must triangulate: cancer types whose tumors show strong sex-differential myeloid axis wiring in TCGA should be the ones whose FAERS sex-biased organ signals are strongest, stable under leave-one-cancer-type-out. We flag one unavoidable weakness up front: the Arm C correlation is across cancer types because no dataset in the file links per-patient irAE outcomes to per-patient tumor transcriptomes, and the OneFlorida+ and Aetna cohorts explicitly lack tumor molecular data (Real-world data analyses unveiled the immune-related adverse effects of immune checkpoint inhibitors across cancer types). The proposal is therefore a triangulated association study whose strongest deliverable, if the checks pass, is a testable mechanistic prediction — that the MYBL2–CCL2 and PPAR$\gamma$ axes operate at different strengths in male and female tumors of the same type — which the wet-lab community can take directly into the syngeneic and organoid models already validated for those axes.
References
- A real-world cohort study of immune-related adverse events in patients receiving immune checkpoint inhibitors
- Cardiac adverse events associated with dual immune checkpoint inhibitors: a pharmacovigilance analysis from the FDA adverse event reporting system
- The MYBL2–CCL2 axis promotes tumor progression and resistance to anti-PD-1 therapy in ovarian cancer by inducing immunosuppressive macrophages
- IDDF2020-ABS-0201 Targeting hepatoma-intrinsic pparγ signaling overcomes immune checkpoint therapy resistance by inflaming the tumor microenvironment
- Acute and subacute oral toxicity assessment of Gancao Xiexin decoction in Sprague-Dawley rats
- Genome-Wide Sex and Gender Differences in Cancer
- Comprehensive safety assessment of donepezil: pharmacovigilance analysis based on the FDA adverse event reporting system
- Adverse drug events (ADEs) risk signal mining related to eculizumab based on the FARES database
- Digitally deconvolving the tumor microenvironment
- Integrative Analysis of Transcriptome and Metabolome Sheds Light on Flavonoid Biosynthesis in the Fruiting Body of Stropharia rugosoannulata
- AI-powered omics-based drug pair discovery for pyroptosis therapy targeting triple-negative breast cancer
- Real-world data analyses unveiled the immune-related adverse effects of immune checkpoint inhibitors across cancer types