Translating biomarkers from basic research to clinical utility involves transfer of information across a series of contexts (from cells to disease animal models to humans) in which data are progressively harder to obtain. It is known that biomarkers identified in cell lines often do not translate to clinical settings and that is one of the main roadblocks in Translational Medicine. Presently, the state-of-the-art machine learning models require a number of training samples. The inventors show that conventional machine learning models, such as Random Forest, Linear Regression Model, Nearest Neighbors, cannot achieve accurate predictions and therefore there is a need for more accurate models.
Researchers at UC San Diego wanted to improve biomarker transfer across contexts (from cells to disease animal models to humans), so they formulated a neural network model, Translation of Cellular Response Prediction (TCRP), using a recently developed technique of few-shot learning. Few-shot learning refers to the practice of feeding a learning model with a very small amount of training data. Few-shot learning aims to identify widely applicable input features by optimizing their transferability, rather than their overall prediction accuracy as in conventional learning approaches. In an initial “pre- training” phase, the model is exposed to a series of pre-defined contexts, with each context represented by numerous training samples. In a second “few-shot learning” phase, the model is presented with a new context not seen previously, and further learning is performed on a small number of new samples. Neural networks trained by this two-phase design have been shown to learn surprisingly rapidly in the new context relative to models trained by conventional means. It can leverage the information of very few samples of human biological experiments, for example (which are precious) and make accurate predictions of drug response.
One of the potential commercial applications would be to provide pharmaceutical/biotech companies with our computation framework, which would be very useful for them to understand the functional mechanisms of drug responses.
Conventional machine learning models require a remarkable amount of training data. However, in the precision medicine domain, it is typical that each biological experiment only consists of 5-10 samples. Our model can leverage the information of these very few samples of human biological experiments and make accurate predictions of drug response
Our model is at experimental data stage.
The invention is patent pending and is available for licensing.
Neural network model, cellular response, drug response, molecular markers, in-vitro platforms, human cell lines, biomarker