Permission to redistribute downstream PLS models trained on TITAN TCGA embeddings
We are conducting a non-commercial academic study to use the released TCGA_TITAN_features.pkl resource as a fixed 768-dimensional representation of TCGA diagnostic slides. We did not fine-tune TITAN or CONCH and did not use or modify the TITAN model weights.
Using these fixed representations, we fitted cancer-specific PLS regression and PLS–LDA models for molecular, genomic-instability and immune endpoints. We subsequently developed an R package to enable independent academic researchers to apply the fitted models to their own TITAN representations.
The proposed R package distribution would not contain:
- TITAN or CONCH model weights;
- the original TCGA_TITAN_features.pkl embeddings;
- TCGA whole-slide images;
- patient-level training embeddings; or
- patient-level molecular outcome data.
The fitted model objects contain only downstream statistical parameters required for prediction, including feature preprocessing values, PLS weights/loadings and regression coefficients, selected component numbers and, for binary models, LDA parameters. There are currently 323 cancer-specific fitted research models.
The project and package are intended exclusively for non-commercial academic research and are not intended for clinical use, diagnosis or commercial exploitation.
We understand that the TITAN terms state that models trained on outputs from TITAN are considered derivatives and that commercial use of such derivatives requires prior approval.
We would like to know if need explicit permission before publicly redistributing these fitted statistical model objects.
Specifically, we would be grateful for confirmation that we may:
- redistribute the fitted PLS and PLS–LDA parameter objects as part of the R package;
- make the package publicly available through GitHub and the Comprehensive R Archive Network (CRAN);
- allow CRAN and its mirrors/archive to continue distributing the package;
- make the package available exclusively for non-commercial academic/research use;
- distribute model-registry metadata, model checksums, training-range summaries and aggregate out-of-fold reference distributions necessary for interpretation of predictions; and
- allow users of the R package to apply these models to TITAN embeddings that they have generated or obtained in accordance with the TITAN terms.
We would also appreciate your guidance regarding the appropriate licence for the fitted model artifacts. We are happy to use CC BY-NC-ND 4.0, another non-commercial licence specified by your team, or any attribution and notice language that Mass General Brigham considers appropriate.
Because CRAN is a public, mirrored repository, it would also be helpful to clarify whether users of the fitted downstream models would need to register separately for TITAN when the TITAN model or original TITAN embeddings themselves are not redistributed by the R package.