Development of Deep Learning Models for Biomarker Identification and Classification

Period of Performance: 08/17/2017 - 02/16/2018


Phase 1 SBIR

Recipient Firm

Amplion, Inc.
BEND, OR 97702
Principal Investigator


Nearly half of all biomedical publications ever written have been published in the last 15 years? the era since the completion of the human genome project. The rate of increase is exponential, shows no signs of slowing down, and can be seen in all of the information sources relevant to precision medicine development. This constitutes a nearly insurmountable burden for the drug-development and diagnostics professionals who develop precision medicines. Next-generation automated evaluation of this data that will enable rapid, supportable, and innovative product-development decisions would lead to the development and approval of many more precision medicines, resulting in improved public health, decreased precision medicine time-to-market, and increased efficiency and profitability for drug-development and diagnostics companies. Developing precision medicines is exceedingly difficult due to an underlying chicken-and-egg problem: patient populations for whom a drug will be effective cannot be identified without a test, but test development is notworth justifying without a drug that is demonstrated to be effective in that population. No system for the identification of actionable precision medicine opportunities exists. The ideal approach would 1) examine multiple systems of record that impact marketplace decisions; 2) be aware of the identify of individual molecular biomarkers mentioned therein, as well as bring in technical details such as measurability and measurement method; and 3) be accessible to drug-development companies looking to define a patient population and diagnostics companies looking to develop a test to provide that definition. Therefore, the overall goal of this multi-phase SBIR project is to capitalize on our preliminary success in building a biomarker database (BiomarkerBase?) that supports this critical interface of precision medicine development decisions. Amplion's highly qualified data science R&D team will collaborate with Dr. Parag Mallick of Stanford University to pursue three Aims: 1) train a Deep Learning Model for biomarker identification in clinical trials, 2) extend Model application and prove performance for classification of biomarker usage intent, and 3) prove that Model-identified and classified biomarkers match and expand expert opinions. Showing that we can identify biomarkers and their usage classifications in clinical trials and publications will establish the predictive potential of our unique algorithms within the limited scope of this Phase I feasibility project and will set the stage for a larger Phase II demonstration. Phase I will provide data that can be incorporated directly into Amplion's BiomarkerBase? product and will allow us to assess how well our Model meets user requirements for this data. Phase II work will allow us to expand/extend this Model to cover additional sources. Phase III work (with Industry partners) will allow us to integrate this commercial service directly into customer work flows. Our next-generation capabilities will provide a critical linkage at the challenging interface between the diagnostics and drug-development efforts and will accelerate the development of novel precision medicines.