How do you train machine learning models to translate between data types? In precision oncology, an impressive ML model uses few-shot learning to predict patient responses to cancer drugs using data from complex, clinically-derived samples. The independent reproduction of the model’s capabilities are a promising step towards further tailoring cancer therapy for individual patients.
Predicting Cancer Drug Response
One of the major early goals of personalized medicine is to use genomic data to select the best cancer treatment possible for a given patient. With a wealth of pharmacogenomic data and sophisticated computational and bioinformatic tools, scientists have been attempting to make that goal a reality, using genetic signatures to train machine learning models to predict a patient’s response to a given drug.
Unfortunately, there is often insufficient clinical genomic data to train these models, and thus, preclinical data are used for training before the model is validated on a smaller dataset of real patient outcomes. These models trained on in vitro data often struggle to translate from predicting the correct response in a cell line to predicting the outcomes in actual patients. However, new ML methods may help bridge this gap.
The Translation of Cellular Response Prediction Model
A recent Nature Machine Intelligence paper by So et al. attempted to reproduce one of the most impressive ML models for cancer drug response prediction and determine its accuracy using clinical trial data. The original model, published in 2021 in Nature Cancer by Ma et al., used a newer technique called few-shot learning that allows models to be tuned to new contexts using a relatively small number of additional samples.
Few-shot learning is particularly well-suited to solving the challenge of translating preclinical predictions, as clinical datasets are much smaller than the cell line screening datasets generated during preclinical studies. Ma et al.’s model, termed ‘translation of cellular response prediction’ or TCRP, was able to quickly adapt from cell line-based models to complex clinical models including patient-derived tumor cells and patient-derived xenografts.
So et al. were able to independently reproduce the impressive results of TCRP using the code published by the original authors. So et al. noted that some missing elements and instructions led their model to produce slightly different values at times, but overall TCRP robustly outperformed existing statistical and ML approaches and reproduced the performance seen in the original paper. So et al. also tested TCRP in novel clinical contexts and observed consistent performance.
The Future of Personalized Medicine
The promise of personalized medicine, particularly personalized oncology, is that tools like these will be an essential part of every oncologist’s toolbox. By matching a patient’s molecular profile to the best possible therapy, doctors and patients can save precious time, reduce unpleasant side effects of ill-compatible drugs, and ideally dramatically improve patient outcomes.
These kinds of studies, both the development of new models to address current challenges in precision oncology like Ma et al.’s work and the independent reproduction of those models by groups like So et al. are critical for bringing this vision to fruition.
Outsourcing Bioinformatics Analysis: How Bridge Informatics Can Help
Groundbreaking studies like these are made possible by technological advances making biological data generation, storage and analysis faster and more accessible than ever before. From pipeline development and software engineering to deploying existing bioinformatics tools, Bridge Informatics can help you on every step of your research journey.
As experts across data types from leading sequencing platforms, we can help you tackle the challenging computational tasks of storing, analyzing and interpreting genomic and transcriptomic data. Bridge Informatics’ bioinformaticians are trained bench biologists, so they understand the biological questions driving your computational analysis. Click here to schedule a free introductory call with a member of our team.
Jane Cook, Biochemist & Content Writer, Bridge Informatics
Jane Cook, the leading Content Writer for Bridge Informatics, has written over 100 articles on the latest topics and trends for the bioinformatics community. Jane’s broad and deep interdisciplinary molecular biology experience spans developing biochemistry assays to genomics. Prior to joining Bridge, Jane held research assistant roles in biochemistry research labs across a variety of therapeutic areas. While obtaining her B.A. in Biochemistry from Trinity College in Dublin, Ireland, Jane also studied journalism at New York University’s Arthur L. Carter Journalism Institute. As a native Texan, she embraces any challenge that comes her way. Jane hails from Dallas but returns to Ireland any and every chance she gets. If you’re interested in reaching out, please email [email protected] or [email protected].