Review Papers Claim by Claim

Topline Critique

The preprint introduces a biologically motivated dynamic PD-L1 state epsilon into an ODE immunotherapy model (Avelumab plus NHS-muIL12) and shows that therapy-dependent PD-L1 regulation can explain adaptive resistance and some nonmonotone tumor dynamics observed in EMT-6 mouse data; the approach is plausible and useful but has identifiability and generalizability limits (murine single model, digitized data, omission of explicit T cell exhaustion) that the authors acknowledge and quantify with RSS/AIC comparisons Dynamic PD-L1 Regulation Paper

Detailed Evidence Based Review

1 Key claims and what the paper actually shows

• Claim Dynamic PD-L1 regulation (state ϵ) determines adaptive immune resistance and shapes outcomes for NHS-muIL12 and anti-PD-L1 (Avelumab) combination therapy. Evidence: the authors add an ODE for ϵ whose production depends on tumor volume and NHS-muIL12 and whose suppression depends on Avelumab; the dynamic-ϵ model reproduces six therapy arms including nonmonotone trajectories that the original constant-ϵ model failed to capture Dynamic ϵ ODE and Fits
• Claim Avelumab suppresses effective tumor PD-L1 (lowers fitted ϵ) while NHS-muIL12 increases it (IFN-gamma mediated upregulation proxy), explaining monotherapy failure with NHS-muIL12 and synergy in combination Avelumab suppresses PD-L1; IL12 upregulates

2 Model formulation and parameterization (what is rigorous vs what is assumed)

The model is an extension of the 4-ODE tumor–T cell–Avelumab–NHS-muIL12 system from prior work by Nikolopoulou et al. The added equation for ϵ is:

Comments:

• Biological plausibility: linking PD-L1 production to tumor volume and immunostimulant A2 as a proxy for IFN-gamma signaling is reasonable and supported by experimental literature that IFN-gamma upregulates PD-L1 (paper cites Fallon et al. IFN-gamma KO mice) IFN-gamma drives PD-L1
• Pharmacokinetics: dA1 and dA2 updated to murine half-lives (dA1≈0.3726 day^-1, dA2≈0.0730 day^-1) — good practice to use species-appropriate PK values Murine PK updates
• Fitting strategy: they fit the base model with epsilon constant per therapy and then fit the dynamic epsilon model globally across all six datasets using lsqnonlin; tumor volumes were digitized from Xu et al. — digitization is acceptable but reduces raw-data fidelity and can introduce error

3 Strengths

• Mechanistic linkage of clinically relevant biology (IFN-gamma driven PD-L1 upregulation and PD-L1 blockade) into a quantitative model that can generate testable hypotheses about dosing and timing Mechanistic modeling strength
• Transparent comparison between constant-epsilon fits and dynamic model with RSS/AIC — authors do not hide that constant-epsilon gave better global AIC but dynamic model captured treatment-specific effects
• Inclusion of plausible alternative mechanisms and limitations (T-cell exhaustion, clonal evolution) in Discussion shows epistemic humility Acknowledged alternatives and limits

4 Weaknesses, uncertainties, and potential biases

• Data fidelity Tumor volumes were digitized from published figures (PlotDigitizer) rather than using raw experimental datasets; digitization is subject to resolution and sampling noise and the authors excluded at least one outlier combination arm — that reduces reproducibility and may bias fits Data digitization and outlier exclusion
• Identifiability The dynamic ϵ ODE adds five parameters yet the observable data are bulk tumor volume time series; identifiability of alpha_A2, alpha_A1, k_basal, K_V, d_epsilon from tumor volume alone is weak — authors partially mitigate by fixing/borrowing literature values but identifiability remains a concern (and may explain better global AIC for constant-epsilon) Parameter identifiability concern
• Missing biology T-cell exhaustion is discussed but not modeled explicitly; other compartments (myeloid cells, regulatory T cells, exosomal PD-L1) can influence PD-L1 dynamics — therefore ϵ may be an effective parameter conflating multiple mechanisms T cell exhaustion omission
• Generalisability Single preclinical EMT-6 murine model; translation to human tumors (heterogeneous PD-L1 regulation, tumor microenvironment complexity) remains untested. The authors use murine PK values but extrapolation beyond EMT-6/BALB-c is speculative.

5 Quantitative model comparison (data table reproduction)

Interpretation: for several individual arms (notably combination low-dose Av+NHS) constant-epsilon fits give lower RSS/AIC, but dynamic model uniquely reproduces nonmonotone tumor trajectories and gives mechanistic ϵ time courses; global AIC parity suggests the extra parameters trade explanatory power for overfitting risk.

6 Cross-check with independent biology resources

PD-L1 biology summarized in UniProt corroborates key mechanisms used by the model: PD-L1 is a single-pass type I membrane protein, regulated by IFN-gamma, subject to ubiquitination and recycling via CMTM6/CMTM4, and can translocate to nucleus under hypoxia — all mechanisms that change effective surface PD-L1 and could be lumped into ϵ; UniProt notes PD-L1 stabilization via deubiquitinases and regulation by cytokine signaling which supports the model choice to let ϵ be treatment-responsive rather than fixed PD-L1 UniProt Evidence

7 Practical implications and testable predictions

• Prediction 1: NHS-muIL12 dosing schedules that produce large transient T-cell activation bursts will raise tumor ϵ (PD-L1) and may reduce monotherapy efficacy — testable by measuring tumor and circulating PD-L1 and IFN-gamma time courses after NHS-muIL12 dosing in mice or patients Therapy-driven ϵ predictions
• Prediction 2: Combining Avelumab with NHS-muIL12 at doses/timers that maintain A1 levels during the IL12-induced ϵ rise will avert adaptive resistance — can be validated with staggered dosing experiments and serial PD-L1 measurement.

8 What would disprove the central claim

• High-quality longitudinal PD-L1 measurements (cell surface and exosomal) in the same tumor model showing no consistent therapy-dependent increase in PD-L1 following NHS-muIL12 would falsify the main mechanism.
• If inclusion of an explicit T-cell exhaustion compartment (without dynamic PD-L1) reproduces all six arms and nonmonotone dynamics with fewer parameters that would argue the ϵ mechanism is unnecessary.

9 Recommendations for improving the study

1. Obtain raw tumor-volume and PD-L1 time-course data (surface and soluble/exosomal) rather than digitized traces to re-fit and assess identifiability.
2. Perform structural and practical identifiability analyses (profile likelihoods, parameter unidentifiability maps) for the ϵ parameters and present confidence intervals for fitted parameters.
3. Extend the model to include an explicit exhausted T-cell compartment or dynamic T-cell dysfunction term and compare model evidence (AIC/BIC) rigorously, not only RSS.
4. Validate model predictions prospectively in the EMT-6 model (serial PD-L1 and IFN-gamma measurements) and in a second syngeneic tumor model with different baseline PD-L1 regulation.

10 Minimal reproducible experiment to settle mechanism

Perform an in vivo EMT-6 experiment with these arms: isotype, NHS-muIL12 alone, Avelumab alone, combination. Collect tumor biopsies at baseline, 24h, 72h, 7d for tumor cell surface PD-L1 (IHC and flow cytometry), exosomal PD-L1 in plasma, and IFN-gamma levels. If PD-L1 increases after NHS-muIL12 and is suppressed by Avelumab as modelled, that supports ϵ mechanism; if PD-L1 does not change but T cell exhaustion markers increase, then exhaustion is more likely.

11 Actionable bottom line

The paper delivers a useful mechanistic hypothesis: PD-L1 is dynamic and therapy responsive and this dynamicity can generate adaptive resistance that explains some complex tumor volume responses. The model is a solid starting point for experiment-driven optimization of dose and timing for combinations of immunostimulants and checkpoint blockade, but the claim requires prospective longitudinal PD-L1 measurement and identifiability work before clinical translation Paper conclusion caution

Author reviews