Adsorption Isotherms on Disordered Receptor Tails: Predicting JAK2 Inhibitor Efficacy Across the In Vitro–In Vivo Gap

Targeted drugs often work beautifully in a plastic dish and then fail in a patient, and the field still lacks models that explain why in mechanistic terms. This proposal argues that one specific, well-documented failure — the gap between the potent in vitro activity of the JAK2 inhibitor Fedratinib against glioma cells and its unverified effectiveness in the brain — can be attacked by importing a mathematical tool from environmental chemistry. Class 1 cytokine receptors relay growth and immune signals into cells by recruiting the enzyme JAK2 onto long, floppy intracellular tails. These tails are "intrinsically disordered regions" — chain-like segments with no fixed 3D shape, decorated with many short docking motifs that physically overlap, so different signaling proteins must compete for the same stretch of chain. That is structurally the same situation as molecules competing for binding sites on a heterogeneous adsorbent surface, which chemists already model with adsorption isotherms — equilibrium curves that map how much solute binds a surface as its concentration rises.

The core claim: if we treat the disordered receptor tail as a one-dimensional, energetically heterogeneous adsorbent, fit competitive isotherm parameters (site capacities and affinities) to measured JAK2 recruitment, and then link JAK2's tail occupancy to downstream signaling output, we can convert an in vitro dose–response curve into a mechanistic prediction of in vivo efficacy that fails loudly when physiology (blood–brain barrier transit, clearance) will break it. No study in the file models receptor–kinase recruitment this way; the closest work stops at sequence-level analysis of the same receptors.

Core Claim and Relation to Prior Work

Three literatures motivate this and none of them meet. Seiffert et al. established that all 29 human class 1 cytokine receptor intracellular domains are entirely disordered and saturated with overlapping short linear motifs, creating "scaffolding hot spots" where signaling molecules compete, and that splicing can delete almost the whole tail (the PRLR SF1b isoform loses 90% of the intracellular domain, removing JAK docking sites) — but they did sequence analysis and biophysical characterization of isolated fragments, with no binding-kinetics model and no pharmacology. Yu et al. identified Fedratinib as a lead compound reversing a high-risk glioma expression signature via the Connectivity Map, and confirmed it kills glioma cell lines with IC50 values between 4 and 9 µg/mL — but they explicitly flag that blood–brain barrier permeability in humans is unverified, and they never examine the receptor-level PTM biology upstream of JAK2. The adsorption literature supplies the mathematics of competitive binding on heterogeneous, multi-site surfaces: co-exposure of Cd(II) and Zn(II) on biochar cut Zn(II) uptake by 56–63% and follows Freundlich power-law fits with $R^2 > 0.85$, showing that competition on heterogeneous surfaces produces large, quantifiable, non-additive penalties. None of it touches biology. The proposal builds the missing bridge: the disordered tail is the adsorbent, JAK2 and competing effectors are the solutes, and post-translational modifications (phosphorylation, lysine acetylation and ubiquitination at overlapping sites) act as time-varying modifiers of site energies.

The isotherm formalism earns its place with one equation. If the tail carries $n$ motif-sites with affinities $K_i$, the JAK2 occupancy $q$ (average number of bound JAK2 molecules per tail) at free JAK2 concentration $C$ is the competitive multi-site Langmuir sum

$$q(C) = \sum_{i=1}^{n} \frac{q_{m,i} K_i C}{1 + K_i C + \sum_j K_{ij} C_j}$$

where $q_{m,i}$ is the capacity of site $i$ and the $C_j$ are concentrations of competing effectors (STATs, SOCS proteins, other kinases) that displace JAK2 from overlapping sites. Heterogeneous-affinity distributions recover Freundlich power-law behavior at low concentration and saturation at high concentration — exactly the Sips interpolation between regimes. This is the one piece of math the note needs; everything else is experimental design.

Research Program

Phase one generates what does not yet exist: isotherm parameters for JAK2 recruitment on cytokine receptor tails. No reported result found on Lacuna for JAK2–receptor binding isotherm parameters, so we will measure them directly, using recombinant intracellular domains of two receptors from the Seiffert et al. panel (the long and SF1b-spliced prolactin receptor, plus one JAK2-dominant receptor such as the erythropoietin or thrombopoietin receptor tail) and quantifying bound JAK2 across a concentration series by fluorescence polarization or pulldown, in single-solute and two-competitor conditions. PTM crosstalk enters as an experimental treatment: tails pre-phosphorylated or acetylated at specific lysines, with modification sites nominated computationally by MethylSight 2.0, which predicts four competing lysine modifications simultaneously (its multitask design raised AUPRC from 0.267 for prior hand-crafted predictors to 0.672, and its lab validation confirmed 68 of 100 predicted sites). Each PTM state yields a separate isotherm; the readout is how acetylation or ubiquitination shifts the effective site capacities and affinities for JAK2 — the molecular counterpart of the Cd/Zn competition penalty, where a competitor at shared sites cut uptake by more than half.

Phase two connects occupancy to drug response. The model's core quantity is not "JAK2 inhibited" but "JAK2 tail occupancy over time under a Fedratinib concentration profile": Fedratinib blocks kinase activity, so we link occupancy to signaling output (phosphorylated STAT3) via a transfer function fit on the SW1783 and SW1088 glioma lines Yu et al. used, and predict viability curves. The decisive comparison is against standard Hill-sigmoid IC50 fits: if the tail behaves as a heterogeneous multi-site adsorbent, dose–response curves under PTM perturbation or isoform switching (long-form vs. SF1b tail) should deviate from single-sigmoid shapes in the direction the isotherm model predicts, and the isotherm parameters fitted on one condition should transfer to held-out conditions (different competitor concentrations, different PTM states).

Phase three attacks the translation gap head-on. The in vitro IC50 of 4–9 µg/mL assumes constant drug exposure in a non-clearing medium. We wrap the occupancy model in a simple pharmacokinetic shell: plasma concentration time-courses from published Fedratinib dosing, an estimated brain-extracellular concentration capped by a blood–brain barrier transmission parameter, and clearance. The model then outputs a predicted time-integrated STAT3 suppression in brain tissue — and, crucially, an explicit regime where it predicts failure (occupancy never crosses the signaling threshold) even though the dish IC50 looked excellent. Validation uses an orthotopic glioma xenograft arm with measured brain drug levels, evaluated PDX-style against tumor growth inhibition.

Baselines

The naive reference point is direct translation: assume the in vitro IC50 (4–9 µg/mL in SW1783 and SW1088, reported by Yu et al.) predicts in vivo potency at matched plasma concentrations. This is what the field implicitly does now, and it is the null hypothesis our PK-wrapped occupancy model must beat — specifically, it must predict which dosing regimens and barrier conditions break efficacy, which the IC50 alone cannot.

The strongest existing competitor is the static-omics machine-learning pipeline of Yu et al.: a LASSO expression signature (their 3-gene CASP1/CASP3/CASP9 signature reached a 1-year survival ROC AUC of 0.883 on TCGA and held at roughly 0.70 on the independent CGGA and Rembrandt cohorts) feeding CMap/L1000FWD drug matching. Its published strength is retrospective prognostic stratification and compound nomination; it has no dose axis, no receptor-level mechanism, and no in vivo prediction. A fair comparison on the same cell lines and the same xenograft data: does adding the occupancy layer change the dose and regimen recommendation, and does its predicted growth inhibition correlate better with observed growth inhibition than a CMap-style binary "candidate / not candidate" call? We will run both on identical compute, with the same tuning budget, and treat any improvement in predicted-vs-observed in vivo response correlation as the success criterion.

The broader drug-response-modeling baseline, following the xenograft assessment study, is a Random Forest or XGBoost regressor trained on cell-line expression and response (CCLE/GDSC) and tested on in vivo responses. That study reports average drug-response $R^2 \approx 0.15$ from such models, with complete failure to predict gemcitabine and paclitaxel responses in PDX cohorts and only moderate success (correlation ≈ 0.5) for the EGFR inhibitor erlotinib. We reproduce this baseline configuration on our JAK2 data as the quantitative bar: the occupancy model must predict in vivo response meaningfully better than a static-expression regressor, evaluated by correlation between predicted and observed tumor growth inhibition with the same cross-validation scheme.

The key-ingredient ablation removes competition and heterogeneity: a single-site Langmuir model with one affinity per tail, no overlapping competitors, no PTM state dependence — effectively a Hill equation with mechanistic dress. If the full competitive model does not beat this ablation at predicting held-out competition and PTM-perturbation conditions, the heterogeneity machinery is not earning its keep and the idea should be abandoned in favor of the simpler model.

Why not just…?

Why not just train XGBoost on gene expression? The file's evidence is blunt: static-expression models reach only $R^2 \approx 0.15$ for drug response on average and fail almost entirely when transferred to xenografts, while the isoform-biomarker study shows that gene-level averaging actively masks predictive signal — aggregate gene expression was not predictive where specific isoforms were, and fewer than 50% of expression biomarkers validated across independent cohorts. A black-box regressor also gives no dose axis and no mechanism, so it cannot tell us which physiological constraint breaks Fedratinib in the brain. This is a real threat on raw prediction accuracy within a single dataset; the experiments settle it by comparing out-of-distribution performance (new PTM states, new competitor levels, in vivo conditions), which is exactly where the file shows these models collapse.

Why not just fit standard Hill/IC50 curves? Because the file documents that IC50 is the weakest link in the chain: the xenograft-assessment study found area-under-the-curve response metrics significantly outperform IC50-based models ($p < 0.001$), and the CCLE/GDSC re-analysis found only 5 of 15 overlapping drugs reproducible between consortia. More fundamentally, a sigmoid fits net viability and treats target engagement as single-site saturation; the Cd/Zn work shows that when sites are shared and heterogeneous, uptake becomes non-additive with penalties up to 63%. If the receptor tail is that kind of surface, the Hill fit will silently absorb the wrong mechanism and mispredict the moment competitors or PTM states change — which is precisely the regime that differs between a dish and a tumor. Our phase-two held-out-condition test is designed to expose this: if single-site and competitive models fit equally well everywhere, the reviewer is right and we will say so.

Why not just use sequence-based PTM predictors like MethylSight 2.0 directly? MethylSight 2.0 is genuinely strong at what it does (AUPRC 0.672 versus 0.267 for prior predictors, with 68% lab confirmation), and we adopt it as a tool for nominating lysine modification sites. But it predicts biochemical capacity for a modification, not whether the modification happens in a given cell state, and it is entirely cell-type and dynamics agnostic — its own open-problem list says so. It cannot produce an occupancy, a dose, or an in vivo statement. We use it upstream; it cannot substitute for the program.

Why not just push a DeepDRA-style multi-omics autoencoder across the gap? DeepDRA reaches AUPRC 0.99 within mixed training/testing sets but drops to 0.72 from GDSC to CCLE, and the xenograft study shows empirical cell-line models fail outright in vivo for most drugs. The autoencoder is physiologically blind by construction — no barrier, no clearance, no target engagement. If, however, a multi-omics model with a PK feature appended matched our mechanistic model's in vivo correlation, that would be a decisive result against us, and we will run that comparison rather than assume it away.

Proposed Experiments and Decisive Checks

Experiment one fits isotherms on the two-tail panel across concentration series, single and two-competitor conditions, and three PTM states per tail. Decisive check: competitive multi-site fits beat single-site Langmuir fits on held-out competitor concentrations by a pre-registered margin in cross-validated error; the analog of the Cd/Zn penalty (a competitor cutting JAK2 occupancy by tens of percent at shared sites) is reproducible rather than noise. If no competition penalty appears, the central mechanism is unsupported and the project stops there.

Experiment two tests prediction of Fedratinib response in SW1783 and SW1088 cells under tail-isoform and PTM perturbation (knockdown/re-expression of long vs. SF1b tails, lysine-mutant tails). Decisive check: parameters fitted on unperturbed cells predict perturbed-condition viability curves better than a re-fit Hill baseline, and better than the best static-expression regressor. Experiment three is the translation test: the PK-wrapped occupancy model, given only the in vitro parameters and a barrier-permeability estimate, predicts orthotopic xenograft tumor growth inhibition and the plasma/brain dosing threshold at which efficacy collapses, judged against the measured in vivo outcomes. Success is a materially higher predicted–observed correlation than both the naive IC50 translation and the static-expression baseline, plus a correct qualitative call on the failure regime. The honest risk is identifiability — tail-level competition may be too weak or too transient to produce distinguishable isotherm shapes — and experiment one is deliberately cheap and early so that a negative answer costs little.