Public Abstract

DE-SC0025298: Combined Experimental and AI-Based Deep Learning Approach to Low Dose and Dose-Rate Breast Cancer Radiation Risk Prediction

Award Status: Active

Institution: Board of Regents, obo, Nevada System of Higher Education (NSHE) - University of Nevada, Las Vegas, Las Vegas, NV
UEI: DLUTVJJ15U66
DUNS: 098377336

Most Recent Award Date: 08/16/2024
Number of Support Periods: 1
PM: Kulkarni, Resham

Current Budget Period: 09/01/2024 - 08/31/2025
Current Project Period: 09/01/2024 - 08/31/2027
PI: Cucinotta, Francis

Supplement Budget Period: N/A

Public Abstract

Combined Experimental and AI-Based Deep Learning Approach to Low Dose and Dose-Rate Breast Cancer Radiation Risk Prediction Francis A. Cucinotta, Professor, UNLV (Principal Investigator) Janice M. Pluth, Associate Professor, UNLV (Co-Investigator) Mingon Kang, Assistant Professor, UNLV (Co-Investigator) Cancer risk following chronic low dose/low dose-rate (LDLDR) irradiation is the predominant concern for the public, radiation workers in the medical field, nuclear energy, and pilots and flight crews, and in the use of diagnostic radiation such as mammography and computed tomography (CT)-scans. Low doses are often considered tissue doses below 100 mGy. Radiation cancer risk has been described by a linear non-threshold dose response (LNT) model based on epidemiology analysis of cohorts exposed to medium to high doses of acute or fractionated radiation. Radiation breast cancer risk is found to be one of the highest risks in epidemiology findings. However, epidemiology findings are severely limited at low doses (<100 mGy) and low dose-rates (<5 mGy/h), including chronic irradiation and cannot rule out alternate hypothesis, which include a dose threshold or a supra-linear dose response. A dose threshold would indicate no risk or a risk that is so small to not be detectable. A supra-linear dose response would indicate a risk larger than a linear dose response model predicts. A variety of molecular and cellular marker studies including assorted “omics’’ data have been reported in humans with breast cancer, and in mouse or cellular models exposed to medium to high radiation doses. Single cell RNA (scRNA-seq) is a more recent and powerful technology allowing for studies of variation in many genes across a cell population. In addition, there are limited genomics information on subjects with breast cancer amongst the Japanese atomic-bomb survivors and Chernobyl nuclear accident. However, these studies are severely limited at low doses and have used a variety of methodologies. The goal of our study is to develop biophysics and artificial intelligence (AI) based deep learning models that combine information from both high and low dose irradiation in humans and experimental models to inform radiation breast predictions for low dose radiation exposures. Our primary hypothesize is that low dose and dose-rate radiation will induce transcriptional changes in a human mammary 3D acini model system, and that stochastic effects will occur due to fluctuations in signaling process that impact transcriptional changes. Furthermore, that data and biophysics approaches can be used in a deep learning model that combines these data with existing data from mouse and humans exposed to radiation at higher doses and human mammary cancer data bases to discern LDLDR changes associated with breast cancer risk. This research will lead to a significantly innovative, biologically interpretable deep learning risk score model to provide breast cancer risk assessment for chronic exposures to low dose (<100 mGy) radiation. The Specific Aims in support of these goals described in our research plan are: Aim 1: To perform scRNA-seq on cells from LDLDR exposed human mammary 3D acini cultures to provide insights into the cellular heterogeneity and identification of rare changes induced by LDLDR irradiation and shed light on stochastic gene expression, regulatory mechanisms, and functional diversity within the exposed populations of cells. Aim 2: Develop a stochastic biochemical approach to relate DNA damage and oxidative stress signaling events after LDLDR irradiation that precedes transcriptional changes. Our previous model of DNA damage repair, ATM, and TGFβ signaling pathways will be extended and recast using a stochastic formalism and studied in relation to the transcription changes in the experiments of Aim 1. Aim 3: Develop an AI deep learning approach using Pathway Graph Convolution Networks (PathGCN) that characterizes cell type-specific transcriptional and pathway-based mechanisms using gene expression (e.g., RNA-seq and scRNA-seq) data from human and mouse studies to predict breast cancer risks from low to high doses of radiation. Our proposal is from a multi-disciplinary research team with expertise in computational modeling in radiation biophysics and AI DL approaches, and experimental radiobiology. The proposed research uses an innovative experimental model in Aim 1, while Aim 2 focusses on modeling radiation effects leading to the observed transcriptional changes. Aim 3 will investigate methods to combine the results from Aims 1 and 2 with experimental data from other radiation studies and breast cancer patient data. This research will establish a highly innovative LDLDR radiation model for breast cancer risk predictions.

Combined Experimental and AI-Based Deep Learning Approach

to Low Dose and Dose-Rate Breast Cancer Radiation Risk Prediction

Francis A. Cucinotta, Professor, UNLV (Principal Investigator)

Janice M. Pluth, Associate Professor, UNLV (Co-Investigator)

Mingon Kang, Assistant Professor, UNLV (Co-Investigator)

Cancer risk following chronic low dose/low dose-rate (LDLDR) irradiation is the predominant concern for the public, radiation workers in the medical field, nuclear energy, and pilots and flight crews, and in the use of diagnostic radiation such as mammography and computed tomography (CT)-scans. Low doses are often considered tissue doses below 100 mGy. Radiation cancer risk has been described by a linear non-threshold dose response (LNT) model based on epidemiology analysis of cohorts exposed to medium to high doses of acute or fractionated radiation. Radiation breast cancer risk is found to be one of the highest risks in epidemiology findings. However, epidemiology findings are severely limited at low doses (<100 mGy) and low dose-rates (<5 mGy/h), including chronic irradiation and cannot rule out alternate hypothesis, which include a dose threshold or a supra-linear dose response. A dose threshold would indicate no risk or a risk that is so small to not be detectable. A supra-linear dose response would indicate a risk larger than a linear dose response model predicts.

A variety of molecular and cellular marker studies including assorted “omics’’ data have been reported in humans with breast cancer, and in mouse or cellular models exposed to medium to high radiation doses. Single cell RNA (scRNA-seq) is a more recent and powerful technology allowing for studies of variation in many genes across a cell population. In addition, there are limited genomics information on subjects with breast cancer amongst the Japanese atomic-bomb survivors and Chernobyl nuclear accident. However, these studies are severely limited at low doses and have used a variety of methodologies.

The goal of our study is to develop biophysics and artificial intelligence (AI) based deep learning models that combine information from both high and low dose irradiation in humans and experimental models to inform radiation breast predictions for low dose radiation exposures. Our primary hypothesize is that low dose and dose-rate radiation will induce transcriptional changes in a human mammary 3D acini model system, and that stochastic effects will occur due to fluctuations in signaling process that impact transcriptional changes. Furthermore, that data and biophysics approaches can be used in a deep learning model that combines these data with existing data from mouse and humans exposed to radiation at higher doses and human mammary cancer data bases to discern LDLDR changes associated with breast cancer risk. This research will lead to a significantly innovative, biologically interpretable deep learning risk score model to provide breast cancer risk assessment for chronic exposures to low dose (<100 mGy) radiation.

The Specific Aims in support of these goals described in our research plan are:

Aim 1: To perform scRNA-seq on cells from LDLDR exposed human mammary 3D acini cultures to provide insights into the cellular heterogeneity and identification of rare changes induced by LDLDR irradiation and shed light on stochastic gene expression, regulatory mechanisms, and functional diversity within the exposed populations of cells.

Aim 2: Develop a stochastic biochemical approach to relate DNA damage and oxidative stress signaling events after LDLDR irradiation that precedes transcriptional changes. Our previous model of DNA damage repair, ATM, and TGFβ signaling pathways will be extended and recast using a stochastic formalism and studied in relation to the transcription changes in the experiments of Aim 1.

Aim 3: Develop an AI deep learning approach using Pathway Graph Convolution Networks (PathGCN) that characterizes cell type-specific transcriptional and pathway-based mechanisms using gene expression (e.g., RNA-seq and scRNA-seq) data from human and mouse studies to predict breast cancer risks from low to high doses of radiation.

Our proposal is from a multi-disciplinary research team with expertise in computational modeling in radiation biophysics and AI DL approaches, and experimental radiobiology. The proposed research uses an innovative experimental model in Aim 1, while Aim 2 focusses on modeling radiation effects leading to the observed transcriptional changes. Aim 3 will investigate methods to combine the results from Aims 1 and 2 with experimental data from other radiation studies and breast cancer patient data. This research will establish a highly innovative LDLDR radiation model for breast cancer risk predictions.