When a patient's immune system goes into overdrive, dozens of circulating signaling proteins — cytokines — rise and fall together, and modern bead-based assays can measure 20 to 100 of them at once from a single blood draw. The problem is that most of these proteins go up in every serious inflammatory state: dengue, influenza, transplant injury, sepsis, and cell-therapy toxicity all produce a broadly similar "cytokine storm." A clinician who wants to know which disease is driving a patient, or a researcher who wants to know which cytokine is worth targeting, needs to separate the few responses specific to one etiology from the generic inflammatory background. Meanwhile, two structural constraints make this hard: cohorts are tiny (9 to 44 patients in the studies below), and the animal models used for mechanistic validation measure the same proteins on different assays, at different absolute concentrations, on a different timescale than human patients.
This proposal imports a mature idea from a seemingly unrelated field — the selectivity index, developed in computational neuroscience and deep-network interpretability to quantify whether a neuron responds specifically to one stimulus rather than broadly to everything — and applies it to multiplexed cytokine panels. The core claim: a unitless, scale-invariant Cytokine Selectivity Index (CSI), computed against a multi-disease reference distribution rather than against healthy controls alone, yields inflammatory signatures that (a) discriminate etiology where raw concentrations and fold-changes cannot, (b) are rank-conserved between mouse models and human cohorts, and (c) require far fewer statistical tests than univariate screening, which directly relieves the small-sample power problem.
Core Claim
Existing multiplex cytokine studies compare patients to healthy controls, so a marker like IP-10 — elevated nearly 50-fold in dengue, surging in influenza-driven liver injury in mice, and part of the predictive signature of lung-transplant graft dysfunction — appears "significant" in every disease and therefore discriminates none of them Cytokine Expression Profile of Dengue Patients at Different Phases of Illness, Systemic Inflammatory Burden Causes Liver Injury in H1N1-Infected Mice, Plasma protein biomarkers for primary graft dysfunction after lung transplantation. I claim that the right comparison class is not "healthy" but "other inflammatory states," and that the resulting normalized index is the quantity that survives the species jump, because the species jump destroys absolute scales (species-specific antibodies, different baseline concentrations, compressed rodent timelines) but should preserve relative specificity. This reframes cross-species validation from coarse qualitative parallelism ("marker X goes up in both mice and humans") into a testable quantitative invariant, exactly the gap flagged in the cross-species literature, where studies settle for "up in both" rather than scale-invariant signatures.
The proposal differs from the closest existing work in three ways. Meta-analysis informed machine learning: Supporting cytokine storm detection during CAR-T cell Therapy augments a 9-patient cohort with statistics from 17 published trials to fix prediction accuracy, but it stays within human data, relies on manual literature extraction, and never asks which cytokine responses are etiology-specific. GeneCytNet performs in-silico cytokine perturbation on latent gene-network representations, but only on synthetic cohorts designed to mimic rheumatoid arthritis, with no protein measurements and no species comparison. MuSe-GNN achieves genuine cross-species alignment across human, lemur, and mouse — but exclusively on transcriptomic graphs, never on the soluble secretome, and Identification of genomic determinants contributing to cytokine release shows mRNA abundance maps imperfectly onto circulating cytokine release. None of these compute a selectivity statistic; none connect it to power.
The Index
Following the class selectivity formulation of Selectivity considered harmful: evaluating the causal impact of class selectivity in DNNs, which compares a unit's mean activation for its preferred class against its mean activation across all other classes, define for each cytokine $i$:
$$CSI_i = \frac{\Delta C_{i,\text{target}} - \overline{\Delta C}_{i,\text{reference}}}{\Delta C_{i,\text{target}} + \overline{\Delta C}_{i,\text{reference}}}$$
Here $\Delta C_{i,\text{target}}$ is the log fold-change of cytokine $i$ in the target condition (say, dengue with warning signs at defervescence) relative to that cohort's own baseline, and $\overline{\Delta C}_{i,\text{reference}}$ is the average of the same quantity across a panel of other inflammatory conditions — pooled from the other cohorts, each internally normalized so no assay's absolute scale matters. The index runs from 0 (the cytokine moves just as much in every inflammatory state — pure pan-inflammatory noise) to 1 (it moves only in the target condition — etiology-specific). Because every quantity is a within-cohort log fold-change, the CSI is unitless by construction, which is what makes cross-species comparison meaningful at all.
A second, compartment-level variant applies the same formula within a single organism: in the H1N1 mouse model, setting the reference to lung and serum changes and the target to liver changes yields a "liver-tropic selectivity" per cytokine — exactly the liver-versus-lung spillover question that Systemic Inflammatory Burden Causes Liver Injury in H1N1-Infected Mice raises but does not quantify, despite measuring the full panel in all three compartments.
Research Program
The program has three components. First, build a pan-inflammatory reference atlas by pooling the shared analytes (IP-10, MCP-1, MIP-1$\beta$, RANTES, IL-6, IL-8, G-CSF, IFN-$\gamma$, TNF-$\alpha$, eotaxin, IL-1Ra) across the four multiplex cohorts in the literature: the 29-plex dengue longitudinal cohort ($n = 44$), the 27-plex lung-transplant cohort ($n = 40$, 12 with persistent severe graft dysfunction), the 14-cytokine CAR-T cohort ($n = 9$ patients, 102–109 draws), and the H1N1 mouse Bio-Plex panel. Each cohort contributes z-scored within-cohort changes, so the atlas is a distribution of "how much does cytokine $i$ move in inflammation generally," against which any target condition's CSI is computed.
Second, run the cross-species test. The mouse side uses the H1N1 model's compartment data to rank cytokines by liver-tropic selectivity. The human side uses the dengue cohort's reported associations between hepatic injury (AST/ALT) and individual cytokines (IL-1Ra, IL-10, IP-10). Both species exhibit hepatic injury driven by systemic inflammation rather than local infection — in the mouse, viral RNA is undetectable in liver — making this the one shared phenotype with protein-level data on both sides in the available literature. The decisive statistic is the rank correlation (Spearman $\rho$) between mouse liver-tropic CSI ranks and human dengue hepatic-injury-association ranks over shared analytes, compared head-to-head against the same correlation computed on raw fold-changes. If fold-changes correlate as well as CSI, the selectivity normalization adds nothing and the idea should shrink; if CSI correlates substantially better, the invariance claim is supported.
Third, map the power consequences. Because the CSI collapses a 27-analyte panel into one summary statistic per condition (or a sparse set of high-CSI analytes), the multiple-testing burden that destroys small-cohort studies shrinks by an order of magnitude. I will quantify this by bootstrap subsampling the two larger cohorts — the 40-patient transplant cohort with 12 events and the 44-patient dengue cohort whose severe subgroup has only 4 patients — and plotting discovery yield at a 5% false-discovery rate and minimum detectable effect at 80% power, for univariate screening versus CSI-based testing. As a worked translational example, the program will compute what CSI-scale effect a 72-patient septic-shock trial could have detected, connecting to the anisodamine trial where microcirculation improved significantly ($p < 0.05$) yet 28-day mortality of 14% versus 16% was hopelessly non-significant ($p = 0.86$) Effect of anisodamine hydrobromide on improving early microcirculation disturbance in septic shock patients — and ask whether a CSI-based surrogate endpoint would have required fewer patients than a mortality endpoint.
Baselines
The naive reference point is standard univariate screening: a Mann-Whitney or t-test per cytokine with Benjamini-Hochberg correction, which is effectively what the dengue and transplant studies did. It is the strongest competitor on raw discovery count in a single cohort because it makes no assumptions, and it will be run on exactly the same cohorts and significance threshold as the CSI analysis. Beating it means two things: recovering the same or more validated discoveries at the same false-discovery rate, and producing rankings that replicate across cohorts and species, which univariate testing cannot even attempt since its output is a list of "elevated in disease" analytes that is nearly identical across diseases.
The strongest existing method is the meta-analysis-informed ensemble M2-CRS from Meta-analysis informed machine learning, which reports accuracy of 94.1% and AUC of 0.936 for cytokine-storm detection on the CAR-T cohort, against a plain Random Forest baseline of 89.3% accuracy and 0.881 AUC trained on the local data alone. It is the strongest competitor because it is the only published method that already attacks the small-sample cytokine problem with external information rather than pure local data. I will run both the raw-concentration Random Forest and a CSI-feature classifier on the identical CAR-T cohort, splits, and tuning budget, and additionally ask the question M2-CRS cannot answer: does the model transfer when features are recomputed against a different reference atlas, and does its feature set mean anything in the mouse data. Beating M2-CRS means matching or exceeding its AUC while removing its manual literature-extraction requirement and gaining cross-species validity; matching its accuracy while transferring where it cannot would already count as a win on this proposal's actual question, and I will state results that way rather than claiming a single-cohort leaderboard.
The key ablation isolates the multi-disease reference: compute the same index but with the reference distribution replaced by healthy controls only, which reduces the CSI to a glorified fold-change. This ablation will be run on every task. If healthy-referenced fold-change performs as well as full CSI on discrimination, cross-species rank conservation, and power, then the reference atlas — the proposal's central ingredient — is doing no work, and the contribution collapses to "normalize your data," which no one should publish.
One method deserves mention as a non-runnable baseline: MuSe-GNN's gene-centric graph alignment MuSe-GNN, which outperformed GIANT and Gene2vec by 20.1% to 97.5% on cross-species functional similarity. It cannot be applied to secretome data — it requires co-expression graphs — and that incompatibility is itself the gap this proposal fills; I will not pretend to compare against it numerically.
Why not just…?
Why not just run univariate tests with FDR correction? Because the method's own outputs demonstrate its ceiling: IP-10 is flagged in dengue (nearly 50-fold elevation), in the H1N1 mouse liver/serum/lung panel, in the transplant graft-dysfunction signature, and in cytokine-release syndrome — a marker significant everywhere has zero etiological information content. The multiple-testing arithmetic is equally unforgiving: 27–29 correlated tests on 40–44 patients leaves almost no power after correction, and the dengue severe subgroup has 4 patients. The asthma-genetics study makes the structural point crisply: NPSR1 showed no significant association in isolation but was significant in a joint interaction model ($H = 4.22$, $P = 0.04$) Correlation of 5-HTT, BDNF and NPSR1 gene polymorphisms — isolated per-analyte tests miss structure that joint statistics recover. Experiment 3 measures this directly as discovery yield per patient.
Why not just train a Random Forest or XGBoost on the raw panel? The CAR-T paper already ran this experiment: the Random Forest reached 89.3% accuracy on 9 patients while leaning on IFN-$\gamma$ and IL-10, features that turned out to be idiosyncratic artifacts of the local cohort, and its reliance shifted to robust analytes only when external trial statistics were injected. Synthetic oversampling does not fix this — it interpolates within the same unrepresentative cloud, a limitation that paper analyzes explicitly. The deeper problem is that off-the-shelf supervised models cannot cross species at all: capture antibodies are species-specific and mouse and human concentration distributions occupy different scales, so a model trained on human picograms-per-milliliter has no mouse input to eat. If the CSI-feature classifier fails to beat the Random Forest on the within-human CAR-T task, that result will be reported plainly and the claim narrowed to the cross-species and power contributions, which the Random Forest cannot contest.
Why not just validate in mice and be done? The translational record in this literature is blunt: Decitabine produced a clean survival benefit in the murine sepsis model ($p = 0.0081$) Inhibiting DNA Methylation Improves Survival in Severe Sepsis, while a human septic-shock intervention that genuinely improved physiology showed nothing on survival ($p = 0.86$) — and orthologs diverge functionally, as with the murine tumor suppressor whose human counterpart behaves differently Ruonan Xu et al.. Mouse-only validation systematically overstates effects and says nothing about required human sample sizes. This is a real threat in the other direction too: if mouse CSI ranks and human ranks turn out not to correlate, my cross-species invariance claim is falsified by Experiment 2, and I have designed that experiment so it can fail cleanly.
Proposed Experiments and Decisive Checks
Experiment 1 computes the CSI atlas across the four human cohorts and tests its discrimination: the prediction, grounded in the dengue paper's phase structure, is that MIP-1$\beta$ (early febrile, warning signs) and MCP-1 (defervescence) earn high phase-specific CSI while IP-10 earns near-zero CSI against a multi-disease reference — despite IP-10 being the most dramatically elevated analyte in that cohort. A parallel test asks whether the celebrated 7-protein transplant signature survives selectivity filtering or is mostly pan-inflammatory. An honest data limitation: no reported result found on Lacuna for absolute per-analyte concentrations in either the dengue or transplant cohort (both papers report signatures and fold-changes, not full concentration tables), so this experiment requires collaborative re-analysis of the underlying panels; the reported directional and fold-change findings are sufficient to pre-register the predictions above.
Experiment 2 is the cross-species rank-conservation test described in the research program, with the pre-registered success criterion of Spearman $\rho > 0.5$ between mouse liver-tropic CSI and human hepatic-injury association ranks over shared analytes, and the fold-change ablation run identically. Experiment 3 is the power mapping: bootstrap discovery-yield and minimum-detectable-effect curves for univariate versus CSI testing on the transplant and dengue cohorts, plus the worked 72-patient septic-shock sample-size calculation, with the FLUVACS trial ($n = 1{,}952$, where low infection counts still compressed power despite the large enrollment HAI and NAI titer correlates) as a calibration anchor for what adequate event counts actually buy. Experiment 4 is the prediction benchmark against the Random Forest (89.3% accuracy, 0.881 AUC) and M2-CRS (94.1% accuracy, 0.936 AUC) on the identical CAR-T cohort and protocol.
The closing check ties the four directions together: if the CSI is genuinely scale-invariant, then a biomarker prioritization derived from the mouse CLP sepsis model should predict which cytokines move in human sepsis — the discovery-to-validation step that the methylation study executed sequentially, with no alignment metric, and that this proposal would execute with one.
References
- Cytokine Expression Profile of Dengue Patients at Different Phases of Illness
- Systemic Inflammatory Burden Causes Liver Injury in H1N1-Infected Mice
- Plasma protein biomarkers for primary graft dysfunction after lung transplantation: a single-center cohort analysis
- Meta-analysis informed machine learning: Supporting cytokine storm detection during CAR-T cell Therapy
- GeneCytNet: an interpretable deep learning framework for rheumatoid arthritis classification and in silico cytokine perturbation modeling
- MuSe-GNN: Learning Unified Gene Representation From Multimodal Biological Graph Data
- Identification of genomic determinants contributing to cytokine release in immunotherapies and human diseases
- Selectivity considered harmful: evaluating the causal impact of class selectivity in DNNs
- Effect of anisodamine hydrobromide on improving early microcirculation disturbance in septic shock patients
- Correlation of 5-HTT, BDNF and NPSR1 gene polymorphisms with anxiety and depression in asthmatic patients
- Inhibiting DNA Methylation Improves Survival in Severe Sepsis by Regulating NF-κB Pathway
- BC094916 suppressed SP 2/0 xenograft tumor by down-regulating Creb1 and Bcl2 transcription
- HAI and NAI titer correlates of inactivated and live attenuated influenza vaccine efficacy