MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Using Raman Spectroscopy and Machine Learning to Predict and Monitor Cellular Radiation Responses

X Deng*, K Milligan, R Ali-Adeeb, P Shreeves, S Van Nest, J Andrews, A Brolo, J Lum, A Jirasek, University of British Columbia, Kelowna, BC, CA, University of Victoria, Victoria, BC, CA, Deeley Research Centre, BC Cancer, Victoria, BC, CA, Weill Cornell Medicine, New York, NY, USA

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose:
To use machine-learning algorithms applied to Raman spectroscopic data acquired on irradiated cells for the prediction and monitoring of cellular radiation responses

Methods:
Normal breast cell line MCF10A and four breast cancer cells (MCF7, BT474, MDA-MB231, and SKBR3) were exposed to varying doses of single fractions of radiation (0, 10, 30, 50Gy). Raman spectral acquisitions were performed on the irradiated cell cultures before and 18, 24 and 66 hours post-irradiation. For each cell line, 240 random spectra were collected per cell line. Additionally, a Raman library of biochemical base chemicals was collected for subsequent input into a constrained non-negative matrix factorization (CNMF) algorithm. Non-negative matrix factorization, logistic regression, lasso logistic regression, and random forest methods were applied to acquired Raman data for model building and response prediction.

Results:
Using CNMF with the Raman library of chemicals, we obtained corresponding scores for each biochemical contribution to the Raman data. Several models such as logistic regression, lasso logistic regression, and random forest were fitted to the biochemical scores for biochemical response signature selection, radiation response prediction and monitoring, and immunohistochemistry predictions. Currently, random forest models provide the best result in achieving these three goals. Using the CNMF biochemical score fitted random forest model, we can classify cancer cell lines and the normal cell line with an accuracy> 90% (specificity and sensitivity >90%). From an empirical threshold of SF2 set based on literature review, we can also classify the cell lines into the relatively radiosensitive group versus relatively radioresistant group.


Conclusion:
The application of machine learning methods facilitated the Raman data analysis for predicting and monitoring cellular radiation responses significantly. Combining CNMF and random forest resulted in a robust model to identify important biochemical signatures for radiation response and to distinguish cancer cells and normal cells in terms of the radiation response.

Funding Support, Disclosures, and Conflict of Interest: CIHR (258926, AGB, AJ, JJL), CFI (32859, AGB, AJ), The BC Cancer Foundation (JJL), The authors declare that there is no conflict of interest.

Keywords

Radiosensitivity, Spectrometry, Modeling

Taxonomy

TH- Radiobiology(RBio)/Biology(Bio): Rbio - Outcome models combining dose, imaging, radiomics/radiogenomics and clinical factors: machine learning

Contact Email