crmPack

Roadmap and Discussion of Planned Features

Daniel Sabanés Bové

October 6, 2025

Start of SLA Phase

  • The crmPack package has entered the SLA phase on 11 September 2025 when we had a kick-off meeting with the crowd-funding companies 🚀
  • Here we give an overview of when the SLA services will be available
  • Plus we show the roadmap for the ongoing work on the package
  • We can discuss the new features on the roadmap and what can be important beyond the next year

Working model

  • We will keep our crmPack stand up meetings which allow to check in with you (the community) and collect general questions and feedback
  • Development will happen as before transparently on GitHub

Services

Service Available when
Free-of-charge technical software support via priority handling of written support requests, incl. product questions, feature requests and development issues Already
Access to the members area “My Account” of a dedicated website for an unlimited number of users End 2025
Annual 2 hours crmPack package training online (free-of-charge for unlimited number of participants) After CRAN release
Access to best practice descriptions and real-world examples: customer exclusive vignettes 2026

Services (cont’d)

Package validation/usability Available when
Get a copy of the formal validation documentation that is customized and licensed for exclusive use by your company 2026
Exchange your experiences within the user community (GitHub issues, regular user discussions, annual meeting) Already
Participate in the annual RPACT meeting (free-of-charge for two participants per company) today!
Get access to the development/beta versions of crmPack, influence development, etc. Already
Review specification documents and training material 2026

Roadmap

  1. Finish tidying up the existing package (main on GitHub) and release new version on CRAN - until end November
  2. Backfill cohorts simulation - until end Jan 2026
  3. Extended mixture prior distributions for (simple) BLRM - until end Feb 2026
  4. Joint BLRM by Neuenschwander et al. (2016) for 2+ drugs incl. mixture distributions - until end June 2026 (shoot for end Apr 2026)
  5. Validation documentation
  6. Other new features afterwards
  7. Exclusive vignettes

Overview of “tidy up” work

  • Currently 5 high priority CRAN milestone issues open (link)
    • These will definitely be addressed until the CRAN release
  • Plus 17 lower priority CRAN milestone issues (link)
    • I will go through these and decide which ones to address until the first CRAN release or later
  • Work will happen mainly now in October

Backfill cohorts simulation

Context:

  • Original simulations in the protocol did not account for additional patients (backfill cohorts).
  • Backfill cohorts allow enrolling up to 6 extra patients at a previously explored dose level, while dose escalation continues.
  • These additional patients are included in the BLRM to inform dose recommendations.

Simulation Setup:

  • Base model: Same as in the protocol (BLRM approach).
  • Dose-toxicity scenarios: 5 scenarios from protocol (e.g., realistic, low-toxicity, high-toxicity, too-toxic, non-logistic).
  • Iterations: 1000 per scenario for each assumption.

Backfill cohorts simulation (cont’d)

Assumptions for Backfill Cohorts:

  1. No backfill (baseline, as in existing simulations).
  2. 1 patient per backfill cohort (minimum extreme).
  3. 6 patients per backfill cohort (maximum extreme).
  4. Random backfill size (0–6 patients per dose per trial).

Outputs: For each assumption and scenario

  • Percentage of trials declaring MTD per dose.
  • Mean (min–max) number of patients per dose.

Extended mixture prior distributions

  • The basic mixture prior is already implemented in LogisticNormalFixedMixture and LogisticNormalMixture
    • Assumes \(\textrm{logit}[p(x)] = \alpha_0 + \alpha_1 \log(x/x*)\)
    • \((\alpha_0, \alpha_1)^\top \sim \sum_{k=1}^{K} w_k \textrm{Normal}_{2}(\mu_k, \Sigma_k)\)
    • Optionally \(\log(\alpha_1)\) instead of \(\alpha_1\) in the above prior
    • Weights are either fixed (for \(K \geq 2\)) or from a Beta distribution (for \(K=2\))
  • Extension to Dirichlet distribution for weights \(w_k\) instead of fixed weights for \(K > 2\)
    • Stephan and team have a working implementation already
    • This will be included in crmPack

Joint BLRM for 2 drugs

The idea is to implement the 5-parameter BLRM approach by Neuenschwander et al. (2014):

  • Let \(\textrm{odds}(p)=p/(1-p)\) be the odds transformation of the probability \(p\), such that \(\textrm{logit}(p) = \log(\textrm{odds}(p))\).

  • Let \(x_i\) be the dose of drug \(i=1,2\), and \(p(x_1, x_2)\) be the probability of DLT with doses \(x_1\) and \(x_2\).

  • The reference doses for the two compounds are again denoted by stars.

  • Then the model assumes a linear interaction function:

    \[ \textrm{odds}(p(x_1, x_2)) = \textrm{odds}(p_0(x_1, x_2)) \cdot \exp(\eta x_1/x_1^{*} x_2/x_2^{*}), \]

  • where \(\eta\) is the interaction coefficient (positive values correspond to synergistic toxicity, zero corresponds to additive effect without interaction, and negative values correspond to antagonistic toxicity).

Joint BLRM for 2 drugs (cont’d)

  • Under no interaction with \(\eta=0\), this reduces the probability \(\textrm{odds}(p(x_1, x_2))\) to \[ p_0(x_1, x_2) = p(x_1) + p(x_2) - p(x_1)p(x_2) = 1 - (1 - p(x_1))(1 - p(x_2)). \]

  • Now for the single-agent DLT probabilities \(p(x_1)\) and \(p(x_2)\) we assume the logistic log-normal models: \(\textrm{logit}[p(x_i)] = \alpha_i + \beta_i \log(x_i/x_i^{*}),\) with prior e.g. \((\alpha_i, \log(\beta_i))^\top \sim \textrm{Normal}(\mu_i, \Sigma_i)\) for \(i=1,2\).

Joint BLRM for 2 drugs (cont’d)

  • Different priors can be used in principle for the 5 parameters \(\theta = (\alpha_1, \beta_1, \alpha_2, \beta_2, \eta)^\top\).
  • Typical is a meta-analytic combined (MAC) prior which separates
    • mono drug 1 (\(\theta_1\))
    • combination of drugs 1 and 2 (\(\theta_2\))
    • mono drug 3 (\(\theta_3\))
  • And then uses the hierarchical prior \(\theta_j \sim \textrm{Normal}(\mu, \Sigma)\), \(j=1,2,3\).
  • Hyperpriors on \(\mu = (\mu_{\alpha_1}, \mu_{\beta_1}, \mu_{\alpha_2}, \mu_{\beta_2}, \mu_{\eta})^\top\) e.g.:
    • \(\mu_{\alpha_i} \sim \textrm{Normal}(\textrm{logit}(0.25), 2.5^2)\), \(i=1,2\)
    • \(\mu_{\beta_1} \sim \textrm{Normal}(0, 0.7^2)\)
    • \(\mu_{\beta_2}, \mu_{\eta} \sim \textrm{Normal}(0, 1)\)
  • Structured assumptions for \(\Sigma\) with 6 parameters (details omitted here)

Joint BLRM for 2 drugs (cont’d)

  • This will be quite some work to implement well, therefore the timeline until latest end June 2026
  • However, we have early code already in the dose_combinations branch (last changes happened 9 years ago!)

Discussion

  • What could be good ways for the user to specify the backfill cohort simulations?
  • How important is it to go beyond 2 drugs in the joint BLRM? (e.g. 3 or 4 drugs)
  • …

References

Neuenschwander, Beat, Alessandro Matano, Zhongwen Tang, Satrajit Roychoudhury, Simon Wandel, and SA Bailey. 2014. “Bayesian Industry Approach to Phase i Combination Trials in Oncology.” Statistical Methods in Drug Combination Studies, 95–135.