Pith. sign in

REVIEW 4 major objections 5 minor 27 references

The paper claims that a single latent-variable model can jointly reproduce tumor-volume trajectories, dropout patterns, and genetic conditioning in patient-derived xenograft data, while recovering treatment-effect parameters consistent with

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

An EB-VAE with a hazard-augmented decoder jointly models tumor growth and dropout, and genetic conditioning modestly improves held-out predictions in melanoma and breast cancer PDX data.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Useful extension of EB-VAE to joint longitudinal/dropout modeling, but the headline genetic gain is not cleanly identified because the with-genetics prior is trained under a different procedure than the baseline. the 4 major comments →

arxiv 2607.13984 v1 pith:ZWI6WBT3 submitted 2026-07-15 stat.ML cs.LG

Multimodal Empirical Bayes Variational Autoencoders for Joint Longitudinal and Time-to-Event Modeling

Anders Sj\"oberg , Nils Olsson , Marcus Baaz , Mats Jirstrand This is my paper
classification stat.ML cs.LG
keywords empirical Bayes variational autoencoderjoint longitudinal and time-to-event modelingneural ordinary differential equationstumor growth modelinginformative dropoutgenomic covariatesstability selectionpharmacometrics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that an empirical Bayes variational autoencoder (EB-VAE), augmented with a hazard model and a covariate-conditioned prior, can jointly model longitudinal tumor growth, time-to-dropout, and high-dimensional genetic covariates in one probabilistic framework. Specifically, it shows that a decoder mapping individual latent effects to tumor volume and cumulative hazard reproduces both tumor-volume distributions and dropout patterns in held-out patient-derived xenograft trajectories, and that a hybrid exponential-growth decoder recovers treatment-effect parameters broadly consistent with previously reported nonlinear mixed-effects estimates. The authors further report that conditioning the empirical Bayes prior on genetic indicators improves individual-level prior predictions in both cutaneous melanoma (13% lower RMSE) and breast cancer (11% lower RMSE). If these findings hold, they offer a modular alternative to hand-specified NLME models that can ingest genomic data and inform dropout-aware tumor dynamics without bespoke model building.

Core claim

The central claim is that the EB-VAE framework — latent individual effects k_i inferred by an amortized encoder, regularized by an empirical Bayes prior p_ψ(k|x), and decoded by a neural ODE or a hybrid exponential-growth model — can also carry the time-to-event process. The paper augments the decoder with a hazard head that adds cumulative hazard H_i(t) as a state variable, so the joint likelihood over (y_i, τ_i, δ_i) couples growth and dropout through the same latent variables. On PDX tumor data, the joint model's prior predictive simulations fell inside the observed percentiles in treatment-stratified VPCs and KM-VPCs for held-out individuals, while the hybrid decoder's post-hoc treatment

What carries the argument

The mechanism is the latent individual effect k_i shared by the longitudinal decoder and the hazard: the decoder is a differential equation in (z_i(t), H_i(t)) where H_i is cumulative dropout hazard, so the time-to-event likelihood is integrated into the ELBO. The empirical Bayes prior p_ψ(k|x) is conditioned on treatment (and, in a second stage, on genetic indicators through a KL-trained shift model), which is what lets the model borrow strength across individuals and treatments. The treatment-effect decomposition in Eq. 1 — prior means written as baseline growth minus monotherapy and combination-interaction terms — is what turns the latent population distribution into interpretable NLME-li

Load-bearing premise

The load-bearing premise is that the 13% (melanoma) and 11% (breast cancer) RMSE improvements come from the genetic information itself, but the genetics-conditioned prior was trained with an extra second-stage, KL-only optimization that the treatment-only baseline did not receive — so any part of the improvement could be an artifact of that extra optimization rather than of the DNA.

What would settle it

Take the already-trained treatment-only model, run the identical second-stage KL-only training with the genetic indicator vectors replaced by random binary vectors (or by permutations of the real ones across individuals), and measure RMSE on held-out individuals. If RMSE drops by roughly the same 13%/11%, the genetic signal is an artifact of the extra optimization; if it stays flat, the genetic effect is real and attributable to the covariates.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • The framework subsumes the classic two-step pharmacometric workflow (longitudinal NLME plus separate survival model) into one joint latent-variable model, so dropout and tumor dynamics inform each other automatically.
  • Hybrid decoders preserve interpretability: treatment-effect parameters can be extracted post hoc from the learned prior means, giving a bridge between neural models and classical population modeling.
  • If the treatment-holdout results generalize, models trained on a diverse set of monotherapies and combinations can predict tumor trajectories under unseen regimens without retraining.
  • Genetic conditioning lowers out-of-sample RMSE, and the stability-selection scheme turns the model into a hypothesis generator that ranks candidate biomarkers (e.g., BRAF V600E, NRAS, NF1, MDM2) for further study.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's genetic result is confounded: the genetics-conditioned prior is trained in a second stage with a KL-only objective, while the treatment-only baseline is the stage-1 prior. A fair experiment would retrain the treatment-only prior with the same two-stage protocol and genetic inputs permuted; if the RMSE gain persists, it is genetic, if not, it is optimization artifact.
  • The framework's hazard augmentation is endpoint-agnostic, so the same decoder state could carry progression, survival, or other clinical events beyond dropout — a natural extension the paper mentions only implicitly.
  • Because most of the genetic benefit appears at early time points, the model suggests genetics mostly sharpens initial trajectory predictions; an untested corollary is that later-time predictions would benefit more from combining genetics with continuous molecular readouts rather than binary alteration indicators.
  • The authors themselves label the genetic experiment a proof-of-concept restricted to two tumor types, so the 13%/11% improvements should be read as scoped to melanoma and breast cancer PDX data, not as a general claim about all solid tumors.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper extends the empirical Bayes variational autoencoder (EB-VAE) framework to jointly model longitudinal tumor-volume trajectories and time-to-event dropout, and further incorporates genomic covariates through a two-stage prior adaptation. The authors evaluate two decoder formulations (fully neural NODE and hybrid exponential-growth plus NODE hazard) on patient-derived xenograft (PDX) data from cutaneous melanoma and breast cancer. They report that the hybrid decoder recovers treatment-effect parameters broadly consistent with previous NLME estimates, that the joint model reproduces tumor-volume and dropout distributions in held-out individuals, and that genetic conditioning improves individual-level prior predictions. The paper also uses stability selection to rank genetic indicators, identifying plausible candidates such as BRAF V600E, NRAS, NF1, and MDM2.

Significance. If the claims hold, the EB-VAE framework would be a useful contribution to pharmacometric machine learning: it provides a single probabilistic framework for longitudinal dynamics, informative dropout, treatment effects, and high-dimensional genomic covariates, while retaining a mechanistic interpretation via the hybrid decoder. The paper makes good use of standard pharmacometric diagnostics (VPCs, KM-VPCs, pcVPCs) and includes treatment-holdout and multi-treatment training experiments that go beyond simple reconstruction. The hybrid parameter recovery, although approximate, is a valuable bridge between amortized variational inference and classical NLME models. However, the headline genetics claim is not currently supported by a controlled comparison, and there are internal inconsistencies in the reported parameter values. The core methodological machinery is promising and the experimental design is largely appropriate, but the genetics result requires re-analysis before the paper's central claims can be accepted.

major comments (4)
  1. [§2.5, §3.3, Table 4] The claim that genetic conditioning improves prior predictions rests on a confounded comparison. The genetics-conditioned prior is trained post hoc with a KL-only objective after freezing the encoder, decoder, and stage-1 prior, while the treatment-only baseline is trained jointly through the full ELBO. The two models differ in both covariate conditioning and optimization procedure, so the RMSE gain (Table 4: CM 161.4→140.0, BC 131.5→117.1) cannot be attributed to genetic information. A matched control is needed: e.g., train a treatment-only prior with the same two-stage KL-only objective, or train the genetics-conditioned prior end-to-end, or permute genetic labels. Without this, the abstract's statement that genetic conditioning improved predictions is unsupported.
  2. [§2.1, Appendix C] Genetic information is linked at the PDX-model level, and the model includes a learned PDX-model embedding. If cross-validation splits individuals (mice/trajectories) rather than PDX models, trajectories from the same PDX model can appear in both training and test sets, allowing the embedding or the shared genetic vector to memorize model-specific characteristics. The manuscript does not state that all trajectories of a given PDX model are confined to one fold. If this is the case, the genetics improvement could be at least partly an artifact of leakage. The authors should clarify the split and preferably report results with PDX-model-level cross-validation.
  3. [§3.1, Table 2, Table 3] There is an internal inconsistency in the hybrid parameter recovery. The text reports monotherapy parameters for LEE011, encorafenib, and binimetinib as 0.0372, 0.0340, and 0.0212, compared with reference values 0.0444, 0.0393, and 0.0090. However, Table 3 lists the same estimates as net growth rates and gives the encorafenib reference as 0.0462, while Table 2 reports different drug-effect parameters (a_LEE011=0.0458, a_encorafenib=0.0490, a_binimetinib=0.0618) with reference values 0.0156, 0.0138, 0.051. These tables and the text use 'parameter' inconsistently and cannot all be correct. This needs to be reconciled before the 'broadly consistent' recovery claim can be assessed.
  4. [§3.3, Table 4, Figure 5] The genetics comparison reports mean RMSE improvements of 13.3% and 11.0% over five runs each, but no uncertainty intervals or paired significance tests are provided. Given the high variance typical of small PDX cohorts and the confounded training procedure, a permutation test or bootstrap confidence interval for the RMSE difference is necessary. The histograms in Figure 5 are descriptive only and do not establish that the improvement is systematic across subjects beyond the aggregate means.
minor comments (5)
  1. [§3.2] The text refers to 'Figure 7' for the KM-VPC/pcVPC comparison, but the main-text figure is numbered Figure 3; Figure 7 appears in the supplementary material. Please correct the cross-reference.
  2. [§3.4] Typo: 'benfit' should be 'benefit'.
  3. [§3.1 and Table 3] The reference values for encorafenib differ between the text (0.0393) and Table 3 (0.0462). Also clarify whether 'monotherapy parameter' means the net growth rate or the drug-effect coefficient a_j; the notation in Eq. (1) suggests the latter, but the results are reported in a mixed way.
  4. [Appendix C] Please state explicitly whether the learned PDX-model embedding is trained per PDX model and whether it is used in the genetics experiments; its role is not described in the main text and its interaction with cross-validation should be clarified.
  5. [General] The manuscript lacks a data and code availability statement. Given the complexity of the training procedure and the importance of the two-stage genetic adaptation, providing code would substantially aid reproducibility.

Circularity Check

0 steps flagged

No circular derivation: hybrid parameter recovery is a post-hoc mapping of fitted prior means to an external benchmark, and the genetics comparison is an empirical held-out evaluation.

full rationale

The paper's derivation chain does not reduce any predicted quantity to a fitted input by construction. The hybrid treatment-effect parameters are explicitly post hoc: 'these treatment-effect parameters are not explicit decoder parameters during training. They are instead obtained post hoc by solving the linear system implied by Eq. 1 using the treatment-specific prior means.' The comparison with [19] is an external published NLME benchmark, not a training target, so 'broadly consistent' is a validation statement rather than a circular derivation. The joint model's dropout component enters through a standard hazard log-likelihood term in the ELBO (Section 2.3), and the genetics-conditioned prior is trained in a second stage against the fixed encoder posterior via KL (§2.5); the resulting held-out RMSE improvement is an empirical result, not an algebraic identity. The self-citations [18,19] supply the base framework and a literature reference, but neither is invoked as a uniqueness theorem or as a constraint that forces the reported values. The main weakness — that the with-genetics and treatment-only priors differ in both conditioning and training procedure — is a validity/confounding concern, which the paper partly acknowledges ('The two-stage genetic adaptation used here was introduced to stabilize learning'), not a circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 2 invented entities

The central claims rest on the latent-variable joint model, the exponential-growth mechanistic assumption for the hybrid decoder, the linear treatment-effect decomposition, and the validity of the two-stage genetic prior training. The treatment-effect parameters are not free in the usual sense; they are post hoc linear transforms of fitted prior means. The genetic program count and indicator set are hand-chosen. No new physical entities are postulated.

free parameters (5)
  • Latent dimensionality of ki = 1 or 2 (per dataset)
    Hand-chosen capacity; shorter follow-up uses 1 latent dimension, longer follow-up uses 2 (§C). Affects model expressiveness.
  • Number of genetic programs = 6
    Chosen by hand for genetic conditioning module (§C); not justified by data.
  • Genetic indicator set = 463 binary indicators
    Constructed ad hoc from COSMIC, literature, ChatGPT candidates, and random sampling (§2.5).
  • Treatment-specific prior means μ_a = estimated from data per treatment arm
    Population-level latent means; treatment-effect parameters a_j, a_jl are post hoc transforms (Eq. 1).
  • Naive dropout threshold = 1500 mm3
    Arbitrary baseline threshold, used only as comparison, not part of the model.
axioms (6)
  • domain assumption Longitudinal and event processes are conditionally independent given latent individual effect k_i
    Stated in §2.3; this is the core factorization that makes joint modeling tractable. If dropout depends on the exact tumor volume trajectory not captured by k, the likelihood is misspecified.
  • standard math Hazard and survival functions follow standard survival model S(t)=exp(-H(t)) with time-dependent hazard λ, and dropout time simulated by exponential threshold
    Used to define the event likelihood in §2.3; standard within survival analysis.
  • domain assumption Hybrid decoder uses exponential tumor growth dz/dt = k z
    Assumed in §2.4 for the Hybrid-EB-VAE to enable comparison with NLME exponential growth model. May not hold for all treatments.
  • domain assumption Treatment effects enter the prior mean linearly with pairwise interactions only (Eq. 1)
    Assumed in §2.4; no higher-order interactions because arms have at most two drugs. If true effects are non-linear, the post hoc a_j, a_jl will be biased.
  • ad hoc to paper The encoder posterior qφ(k|y,x) is a good enough approximation to the true posterior that a genetics-conditioned prior trained to match it will improve predictions
    The two-stage genetic training (§2.5) optimizes the prior against the fixed encoder's posterior on training data. This assumes the encoder is accurate and that the shift model can generalize.
  • domain assumption PDX data are representative and dropout is informative
    Underlies the joint model evaluation; if dropout is mostly administrative or protocol-driven, the hazard component adds little.
invented entities (2)
  • genetic programs no independent evidence
    purpose: Project 463 binary genetic indicators into six learned factors that gate treatment-dependent prior shifts
    Learned latent representation; no external biological validation, purely internal to the model.
  • PDX-model embedding no independent evidence
    purpose: Per-cell-line learned embedding combined with genetic representation to condition the prior
    Learned per-line embedding; no external validation.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Empirical Bayes Variational Autoencoders for Joint Longitudinal and Time-to-Event Modeling." pith.science (2026). https://pith.science/paper/ZWI6WBT3

@misc{pith2026260713984,
  author       = {Pith},
  title        = {Pith review of: Multimodal Empirical Bayes Variational Autoencoders for Joint Longitudinal and Time-to-Event Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZWI6WBT3}},
  note         = {Machine review of arXiv:2607.13984}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Longitudinal tumor measurements, dropout information, and genetic covariates provide complementary information about treatment response, but integrating these data sources within a single population modeling framework remains challenging. We extend the empirical Bayes variational autoencoder (EB-VAE) framework to joint longitudinal and time-to-event modeling and evaluate it on tumor growth data. The framework represents inter-individual variability using latent individual effects regularized by a covariate-conditioned empirical Bayes prior, while a decoder maps these latent effects to tumor-volume trajectories. To account for informative dropout, the decoder was augmented with a hazard model, yielding joint predictions of tumor growth and time to dropout. We further compared fully neural and hybrid semi-mechanistic decoder formulations and incorporated genomic covariates through a genetics-conditioned prior adaptation. The hybrid decoder recovered treatment-effect parameters broadly consistent with previously reported nonlinear mixed-effects estimates, while achieving prior predictive performance comparable to the neural decoder. The joint model reproduced both tumor-volume distributions and dropout patterns in held-out individuals, and genetic conditioning improved individual-level prior predictions in both cutaneous melanoma and breast cancer experiments. Stability selection identified several biologically plausible genetic indicators, including alterations in BRAF, NRAS, NF1, and MDM2. These results demonstrate that EB-VAE provides a flexible probabilistic framework for combining neural dynamics, mechanistic structure, time-to-event modeling, and high-dimensional covariates in pharmacometric applications.

Figures

Figures reproduced from arXiv: 2607.13984 by Anders Sj\"oberg, Marcus Baaz, Mats Jirstrand, Nils Olsson.

Figure 1
Figure 1. Figure 1: Overview of the empirical Bayes variational autoencoder framework for joint longitudinal and time-to-event modeling. Longitudinal tumor observations, treatment information, and optional covariates are used by the encoder to infer individual-specific latent effects ki . A covariate-conditioned empirical Bayes prior pψ(ki | xi) captures systematic population-level variability, including differences between t… view at source ↗
Figure 2
Figure 2. Figure 2: Treatment-stratified VPCs for the Hybrid-EB-VAE, illustrating model fit to the observed tumor-volume data. Observed percentiles are shown in black/gray and predicted percentiles in blue/orange, with shaded regions indicating 95% confidence intervals for the predicted percentiles. Simulated trajectories were censored according to the learned dropout process, and VPCs were truncated once fewer than 10 indivi… view at source ↗
Figure 3
Figure 3. Figure 3: Kaplan–Meier visual predictive checks (KM-VPCs; panels a,c) and prediction-corrected visual predictive checks (pcVPCs; panels b,d) for a representative cross-validation split of the cutaneous melanoma dataset, with test individuals pooled across treatment groups. Observed percentiles (not used for training) are shown in black and gray, while predicted percentiles are shown in blue and orange. Shaded region… view at source ↗
Figure 4
Figure 4. Figure 4: Treatment-holdout VPCs for the cutaneous melanoma (first row) and breast cancer (second row) datasets. Each panel evaluates prior predictive performance for a treatment group excluded during training. For each cancer type, panels were selected to show the untreated group, the poorest-fitting held-out treatment group, and the best-fitting held-out treatment group among the evaluated treatment-holdout experi… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of prior predictive performance with and without genetic conditioning. (a) Cumulative mean individual RMSE improvement for the 100-day CM experiment, expressed as percent improvement from conditioning the prior on genetic information. (b) Cumulative mean individual RMSE with and without genetic conditioning for the 100-day CM experiment. (c) Distribution of individual RMSE improvements for the e… view at source ↗
Figure 6
Figure 6. Figure 6: Mean individual prior RMSE for selected cutaneous melanoma treatment groups under two training regimes. “Same treatment only” denotes models trained only on individuals from the evaluated treatment group. “All treatments” denotes models trained on the same treatment-specific training individuals together with individuals from the remaining treatment groups. Both models were evaluated on the same held-out i… view at source ↗
Figure 7
Figure 7. Figure 7: (left) KM-VPCs (right) pcVPCs for one representative cross-validation split of the breast cancer dataset with test individuals pooled across treatment groups. Observed percentiles (not used for training) are shown in black/gray and predicted percentiles in blue/orange, with shaded regions indicating 95% confidence intervals for the predicted percentiles. (a) KM-VPC using the learned survival-model dropout … view at source ↗
Figure 8
Figure 8. Figure 8: Prediction-corrected visual predictive checks for the remaining cross-validation folds of the cutaneous melanoma dataset. Panels show held-out test individuals pooled across treatment groups for folds 1–4. Observed percentiles are shown in black/gray and predicted percentiles in blue/orange, with shaded regions indicating 95% confidence intervals for the predicted percentiles. These diagnostics complement … view at source ↗
Figure 9
Figure 9. Figure 9: Prediction-corrected visual predictive checks for the remaining cross-validation folds of the breast cancer dataset. Panels show held-out test individuals pooled across treatment groups for folds 1–4. Observed percentiles are shown in black/gray and predicted percentiles in blue/orange, with shaded regions indicating 95% confidence intervals for the predicted percentiles. 26 [PITH_FULL_IMAGE:figures/full_… view at source ↗
Figure 10
Figure 10. Figure 10: Additional treatment-holdout visual predictive checks for the CM dataset. Each panel evaluates prior predictive performance for a treatment group that was excluded during training. These panels show the remaining held-out treatment groups not included in the main treatment-holdout figure. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Additional treatment-holdout visual predictive checks for the BC dataset. Each panel evaluates prior predictive performance for a treatment group that was excluded during training. These panels show the remaining held-out treatment groups not included in the main treatment-holdout figure. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 1 linked inside Pith

  1. [1]

    Joint modelling of longitudinal and repeated time-to-event data using nonlinear mixed-effects models and the stochastic approximation expectation–maximization algorithm,

    C. Mbogning, K. Bleakley, and M. Lavielle, “Joint modelling of longitudinal and repeated time-to-event data using nonlinear mixed-effects models and the stochastic approximation expectation–maximization algorithm,”Journal of Statistical Computation and Simulation, vol. 85, no. 8, pp. 1512–1528, 2015

    2015

  2. [2]

    Translational pharmacoki- netic/pharmacodynamic modeling of tumor growth inhibition supports dose-range se- lection of the anti–pd-1 antibody pembrolizumab,

    A. Lindauer, C. R. Valiathan, K. Mehta,et al., “Translational pharmacoki- netic/pharmacodynamic modeling of tumor growth inhibition supports dose-range se- lection of the anti–pd-1 antibody pembrolizumab,”CPT: Pharmacometrics & Systems Pharmacology, vol. 6, no. 1, pp. 11–20, 2017

    2017

  3. [3]

    Model-based prediction of phase iii overall survival in colorectal cancer on the basis of phase ii tumor dynamics,

    L. Claret, P. Girard, P. M. Hoff,et al., “Model-based prediction of phase iii overall survival in colorectal cancer on the basis of phase ii tumor dynamics,”Journal of Clinical Oncology, vol. 27, no. 25, pp. 4103–4108, 2009

    2009

  4. [4]

    Multi-omics approaches to disease,

    Y. Hasin, M. Seldin, and A. Lusis, “Multi-omics approaches to disease,”Genome biology, vol. 18, no. 1, p. 83, 2017. 16

    2017

  5. [5]

    High-throughput screening using patient-derived tumor xenografts to predict clinical trial drug response,

    H. Gao, J. M. Korn, S. Ferretti, J. E. Monahan, Y. Wang, M. Singh, C. Zhang, C. Schnell, G. Yang, Y. Zhang,et al., “High-throughput screening using patient-derived tumor xenografts to predict clinical trial drug response,”Nature medicine, vol. 21, no. 11, pp. 1318– 1325, 2015

    2015

  6. [6]

    Nonlinear mixed effects models for repeated measures data,

    M. J. Lindstrom and D. M. Bates, “Nonlinear mixed effects models for repeated measures data,”Biometrics, vol. 46, no. 3, pp. 673–687, 1990

    1990

  7. [7]

    A review of mixed-effects models of tumor growth and effects of anticancer drug treatment used in population analysis,

    B. Ribba, N. H. Holford, P. Magni,et al., “A review of mixed-effects models of tumor growth and effects of anticancer drug treatment used in population analysis,”CPT: Pharmacometrics & Systems Pharmacology, vol. 3, no. 5, p. e113, 2014

    2014

  8. [8]

    A novel method based on unbiased correlations tests for covariate selection in nonlinear mixed effects models: the COSSAC approach,

    G. Ayral, J.-F. Si Abdallah, C. Magnard, and J. Chauvin, “A novel method based on unbiased correlations tests for covariate selection in nonlinear mixed effects models: the COSSAC approach,”CPT: Pharmacometrics & Systems Pharmacology, vol. 10, no. 4, pp. 318–329, 2021

    2021

  9. [9]

    Covariate modeling in pharmacometrics: general points for consideration,

    K. Sanghavi, J. Ribbing, J. A. Rogers, M. A. Ahmed, M. O. Karlsson, N. Holford, E. Chasseloup, M. Ahamadi, K. G. Kowalski, S. Cole,et al., “Covariate modeling in pharmacometrics: general points for consideration,”CPT: Pharmacometrics & Systems Pharmacology, vol. 13, no. 5, pp. 710–728, 2024

    2024

  10. [10]

    Neural-ODE for pharmacokinetics modeling and its advantage to alternative machine learning models in predicting new dosing regimens,

    J. Lu, K. Deng, X. Zhang, G. Liu, and Y. Guan, “Neural-ODE for pharmacokinetics modeling and its advantage to alternative machine learning models in predicting new dosing regimens,”iScience, vol. 24, no. 7, 2021

    2021

  11. [11]

    Low-dimensional neural ordinary differential equations accounting for inter-individual variability implemented in Monolix and NONMEM,

    D. S. Br¨ am, B. Steiert, M. Pfister, B. Steffens, and G. Koch, “Low-dimensional neural ordinary differential equations accounting for inter-individual variability implemented in Monolix and NONMEM,”CPT: Pharmacometrics & Systems Pharmacology, vol. 14, no. 1, pp. 5–16, 2025

    2025

  12. [12]

    Variational autoencoder for inference of nonlinear mixed effect models based on ordinary differential equations,

    Z. Li, M. Prague, R. Thi´ ebaut, and Q. Clairon, “Variational autoencoder for inference of nonlinear mixed effect models based on ordinary differential equations,”arXiv preprint arXiv:2601.17400, 2026

    arXiv 2026

  13. [13]

    Phar- macometrics and machine learning partner to advance clinical data analysis,

    G. Koch, M. Pfister, I. Daunhawer, M. Wilbaux, S. Wellmann, and J. E. Vogt, “Phar- macometrics and machine learning partner to advance clinical data analysis,”Clinical Pharmacology & Therapeutics, vol. 107, no. 4, pp. 926–933, 2020

    2020

  14. [14]

    The evolution of nonlinear mixed effects modeling in pharmacometrics: toward AI-based variational autoencoders,

    J. Rohleff, G. Koch, and J. Schropp, “The evolution of nonlinear mixed effects modeling in pharmacometrics: toward AI-based variational autoencoders,”Journal of Pharmacokinetics and Pharmacodynamics, vol. 53, p. 37, 2026

    2026

  15. [15]

    Neural ordinary differential equations,

    R. T. Chen, Y. Rubanova, J. Bettencourt, and D. K. Duvenaud, “Neural ordinary differential equations,”Advances in Neural Information Processing Systems, vol. 31, 2018

    2018

  16. [16]

    Auto-encoding variational Bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational Bayes,”arXiv preprint arXiv:1312.6114, 2013

    Pith/arXiv arXiv 2013

  17. [17]

    An introduction to variational autoencoders,

    D. P. Kingma and M. Welling, “An introduction to variational autoencoders,”Foundations and Trends in Machine Learning, vol. 12, no. 4, pp. 307–392, 2019

    2019

  18. [18]

    The empirical Bayes variational autoencoder – a neural ODE approach for population modeling in pharmacology,

    M. Baaz, A. Sj¨ oberg, and M. Jirstrand, “The empirical Bayes variational autoencoder – a neural ODE approach for population modeling in pharmacology,”CPT: Pharmacometrics & Systems Pharmacology, vol. 15, no. 7, p. e70280, 2026. 17

    2026

  19. [19]

    Optimized scaling of translational factors in oncology: from xenografts to RECIST,

    M. Baaz, T. Cardilin, F. Lignet, and M. Jirstrand, “Optimized scaling of translational factors in oncology: from xenografts to RECIST,”Cancer Chemotherapy and Pharmacology, vol. 90, no. 3, pp. 239–250, 2022

    2022

  20. [20]

    Convergence of a stochastic approximation version of the em algorithm,

    B. Delyon, M. Lavielle, and E. Moulines, “Convergence of a stochastic approximation version of the em algorithm,”The Annals of Statistics, vol. 27, no. 1, pp. 94–128, 1999

    1999

  21. [21]

    C. M. Bishop,Pattern Recognition and Machine Learning. New York: Springer, 2006

    2006

  22. [22]

    The COSMIC cancer gene census: describing genetic dysfunction across all human cancers,

    Z. Sondka, S. Bamford, C. G. Cole, S. A. Ward, I. Dunham, and S. A. Forbes, “The COSMIC cancer gene census: describing genetic dysfunction across all human cancers,” Nature Reviews Cancer, vol. 18, no. 11, pp. 696–705, 2018

    2018

  23. [23]

    Stability selection,

    N. Meinshausen and P. B¨ uhlmann, “Stability selection,”Journal of the Royal Statistical Society Series B: Statistical Methodology, vol. 72, no. 4, pp. 417–473, 2010

    2010

  24. [24]

    Prediction-corrected visual predictive checks for diagnosing nonlinear mixed-effects models,

    M. Bergstrand, A. C. Hooker, J. E. Wallin, and M. O. Karlsson, “Prediction-corrected visual predictive checks for diagnosing nonlinear mixed-effects models,”The AAPS Journal, vol. 13, pp. 143–151, 2011

    2011

  25. [25]

    BRAF V600E in cancer: Exploring structural complexities, mutation profiles, and pathway dysregulation,

    J. Bharti, P. Gogu, S. K. Pandey, A. Verma, J. P. Yadav, A. K. Singh, P. Kumar, A. R. Dwivedi, and P. Pathak, “BRAF V600E in cancer: Exploring structural complexities, mutation profiles, and pathway dysregulation,”Experimental Cell Research, vol. 446, no. 1, p. 114440, 2025

    2025

  26. [26]

    MDM2/MDM4 amplification and CDKN2A deletion in metastatic melanoma and glioblastoma multiforme may have implications for targeted therapeutics and immunotherapy,

    T. E. Arnoff and W. S. El-Deiry, “MDM2/MDM4 amplification and CDKN2A deletion in metastatic melanoma and glioblastoma multiforme may have implications for targeted therapeutics and immunotherapy,”American Journal of Cancer Research, vol. 12, no. 5, p. 2102, 2022

    2022

  27. [27]

    Signal pathways of melanoma and targeted therapy,

    W. Guo, H. Wang, and C. Li, “Signal pathways of melanoma and targeted therapy,”Signal transduction and targeted therapy, vol. 6, no. 1, p. 424, 2021. Appendix A Detailed Evaluation Procedures Models were evaluated using both prior and posterior predictive simulations, depending on the objective of the analysis. Prior predictive simulations were used to as...

    2021

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.