Using Machine Learning to Predict Cancer Treatment Effectiveness
Scientists develop algorithms that can not only identify tumors that will respond well to chemotherapy, but which are also likely to resist treatment.
Rapid cell proliferation or spreading is a staple of cancer, especially when it is combined with therapeutic sensitivity that can result in DNA replication stress (RS). This issue with tumors lies in their ability to become resistant to medications known as RS-inducing (RSi) agents, also known as chemotherapy.
In a study published in Cancer Discovery,1 investigators sought to develop machine learning (ML) algorithms that can predict tumors that are likely to resist treatments and those that may respond well to it.
In order to be able to replicate accurate responses to RSi drugs, the investigators focused on a set of 718 genes assessed by current clinical cancer gene panels, including one or more of FoundationOne CDx, Tempus xT, and Project GENIE. Specifically, they used the genomic alteration status of these genes in a tumor sample, including the presence or absence of mutation and copy number aberration, as input. The models were trained using drug response data for genomically characterized tumor cell lines gathered from the Cancer Therapeutics Response Portal and the Genomics of Drug Sensitivity in Cancer databases.
Being that these databases featured the measured responses to many RSi drugs targeting DNA replication or DNA damage response, the authors decided to prioritize six RSi agents:
- cisplatin
- gemcitabine
- camptothecin
- etoposide
- olaparib
- CD437
To be able to compile genotypes for all of the cell lines, investigators extracted non-synonymous coding mutations and copy number alterations for the clinical panel genes from the Cancer Cell Line Encyclopedia. Any gene mutations were indicated as either being present or absent, and these mutations were filtered for missense, nonsense, and non-stop mutations; splice site and region variations; frameshift insertions and deletions, and in-frame insertions and deletions. If there were any gene copy number deletions or amplifications, they were to be separately marked using binary symbols.
One multi-drug visible neural network (VNN) and six single-drug visible neural networks were trained, where the VNN architecture followed the connections of the 718 genes and 131 assemblies in the NeST. They conducted this training by diminishing the mean squared error between the predicted and observed drug responses using standard back-propagation methods.
Using nested cross validation, the precision of drug response prediction was measured for each model, in which 64% of cell lines were chosen at random for model training, 16% for model validation and hyperparameter tuning, and 20% as held-out cell lines that were yet to be seen during training or validation. This assessment produced predictive odds ratios in the range of 2.2-3.2 across the six RSi drug models.
The model was able to find 41 protein assemblies that returned RSi drug responses, meaning that genetic alterations affected drug efficacy. Following the training, the model was tested for cervical cancer, in which approximately 35% of cervical and lung tumors remained after treatment. The model was successfully recognizing tumors that would respond well to therapy; in turn, this resulted in improved patient outcomes. The model also found tumors likely would resist treatment.
"Unraveling an AI model's decision-making process is crucial, sometimes as important as the prediction itself," Trey Ideker, PhD, one of the study’s authors and a UC San Diego of Medicine professor in the department of medicine, said in an interview.2 "Our model's transparency is one of its strengths, first because it builds trust in the model, and second because each of these molecular assemblies we’ve identified becomes a potential new target for chemotherapy. We’re optimistic that our model will have broad applications in not only enhancing current cancer treatment, but also in pioneering new ones."
Reference
1. Zhao X, Singhal A, Park S, Kong JH, Bachelder R, Ideker T. Cancer mutations converge on a collection of protein assemblies to predict resistance to replication stress. Cancer Discov 2024. https://doi.org/10.1158/2159-8290.CD-23-0641
2. Martin, Miles. AI Harnesses Tumor Genetics to Predict Treatment Response. UC San Diego Today. January 18, 2024. Accessed February 15, 2024. https://today.ucsd.edu/story/ai-harnesses-tumor-genetics-to-predict-treatment-response#:~:text=In%20a%20groundbreaking%20study%20published,when%20cancer%20will%20resist%20chemotherapy.
