Exciting interdisciplinary research collaborations are flourishing across the computational cancer community! If you are developing a new collaboration using the computational resources found on the computational.cancer.gov portal, send us an email at computational-cancer-nih.gov. Descriptions of collaborations—and collaborators—will be posted here.
Learn about NCI AI events and funding opportunities at https://www.cancer.gov/research/resources/ai-cancer-research to help support your collaboration needs.
Read about current collaborations below:
Collaborators:
- Natallah Lanman, Purdue University
- Sara Jones and Ryan Weil, Frederick National Laboratory for Cancer Research
Objective:
Build a cross-species ML/AI model using human RNA-seq data to perform primary tumor type classification in canine RNA-seq data.
Summary:
TUmor cLassIfication Predictor (TULIP), an updated TC1 NCI-DOE resource, is a classification tool developed to predict primary tumor tissue types based on human RNA-seq data. The similarity between human and canine cancers and their respective gene expression profiles offers an opportunity to investigate whether a model trained on human RNA-seq data can predict canine tumor types.
To answer this question, scientists from Purdue University and the Frederick National Lab trained a human-based model with 18 primary tumor types using TULIP’s model architecture. One of these tumor types was glioblastoma, a primary tumor type of interest for collaborators at Purdue University. The inputs used in this model are genes shared between humans and canines. On a test dataset of 95 canine RNA-seq samples, this model achieved an overall accuracy of 80%.
Significance:
This collaborative work highlights the possibilities for successful adaptation of models trained on human RNA-seq data to predict the primary tumor tissue type of canine RNA-seq samples. Furthermore, this study represents an example of cross-species ML/AI model exploration that can inspire the use of other human-based models in the canine cancer research space.
Resource Used:
References:
Jones, S., Beyers, M., Shukla, M., Xia, F., Brettin, T., Stevens, R., Weil, M.R., & Ranganathan Ganakammal, S. (2022). TULIP: An RNA-seq-based Primary Tumor Type Prediction Tool Using Convolutional Neural Networks. Cancer Informatics, 21.
Collaborators:
- Sunita Chandrasekaran and Vineeth Gutta, University of Delaware
- Sara Jones and Ryan Weil, Frederick National Laboratory for Cancer Research
Objective:
Computational evaluation of regression-based deep learning neural network (DNN) and tree-based XGBoost algorithms to predict tumor response with a single drug.
Summary:
The NCI-DOE resources developed in the Cellular Level Pilot (Predictive Modeling for Pre-Clinical Screening) include three deep learning drug response prediction models:
- Single Drug Response Predictor (P1B3)
- Combination drug response predictor (Combo)
- Unified drug response predictor (Uno)
These models were trained using RNA-seq gene expression profiles, drug response data, and drug descriptors from various cancer cell line sources, including the following:
- NCI-60
- Cancer Therapeutics Response Portal (CTRP)
- Genomics of Drug Sensitivity in Cancer (GDSC)
- Genentech Cell Line Screening Initiative (gCSI)
- Cancer Cell Line Encyclopedia (CCLE)
Of these models, collaborators focused on the single drug response model, P1B3, and compared its performance on different types of cancer cell line sources with another type of model, the tree based XGBoost algorithm.
Collaborators plan to use CANDLE, another resource developed by the NCI-DOE Collaboration, for hyperparameter optimization. Collaborators also plan to enhance the usability of these models by developing a bring your own data (BYOD) application for cancer researchers to use the models to evaluate their own data.
Resources Used:
References:
Publications from this collaboration are forthcoming.