Predicting ionizing radiation exposure using biochemically-inspired genomic machine learning

The transcriptomic revolution and radiation biology

PURPOSE

This article will briefly review the origins and evolution of functional genomics, first describing the experimental technology, and then some of the approaches applied to data analysis and visualization. It will emphasize application of functional genomics to radiation biology, using examples from the author's work to illustrate several key types of analysis. It concludes with a look at non-coding RNA, alternative reading of the genome, and single-cell transcriptomics, some of the innovative areas that may help to shape future research in radiation biology and oncology.

CONCLUSIONS

Transcriptomic approaches have provided insight into many areas of radiation biology and medicine, and innovations in technology and data analysis approaches promise continued contributions to radiation science in the future.

Pathway-extended gene expression signatures integrate novel biomarkers that improve predictions of patient responses to kinase inhibitors

Cancer chemotherapy responses have been related to multiple pharmacogenetic biomarkers, often for the same drug. This study utilizes machine learning to derive multi‐gene expression signatures that predict individual patient responses to specific tyrosine kinase inhibitors, including erlotinib, gefitinib, sorafenib, sunitinib, lapatinib and imatinib. Support vector machine (SVM) learning was used to train mathematical models that distinguished sensitivity from resistance to these drugs using a novel systems biology‐based approach. This began with expression of genes previously implicated in specific drug responses, then expanded to evaluate genes whose products were related through biochemical pathways and interactions. Optimal pathway‐extended SVMs predicted responses in patients at accuracies of 70% (imatinib), 71% (lapatinib), 83% (sunitinib), 83% (erlotinib), 88% (sorafenib) and 91% (gefitinib). These best performing pathway‐extended models demonstrated improved balance predicting both sensitive and resistant patient categories, with many of these genes having a known role in cancer aetiology. Ensemble machine learning‐based averaging of multiple pathway‐extended models derived for an individual drug increased accuracy to >70% for erlotinib, gefitinib, lapatinib and sorafenib. Through incorporation of novel cancer biomarkers, machine learning‐based pathway‐extended signatures display strong efficacy predicting both sensitive and resistant patient responses to chemotherapy.

Improved radiation expression profiling in blood by sequential application of sensitive and specific gene signatures

PURPOSE

Combinations of expressed genes can discriminate radiation-exposed from normal control blood samples by machine learning (ML) based signatures (with 8-20% misclassification rates). These signatures can quantify therapeutically relevant as well as accidental radiation exposures. The prodromal symptoms of acute radiation syndrome (ARS) overlap those present in influenza and dengue fever infections. Surprisingly, these human radiation signatures misclassified gene expression profiles of virally infected samples as false positive exposures. The present study investigates these and other confounders, and then mitigates their impact on signature accuracy.

METHODS

This study investigated recall by previous and novel radiation signatures independently derived from multiple Gene Expression Omnibus datasets on common and rare non-neoplastic blood disorders and blood-borne infections (thromboembolism, S. aureus bacteremia, malaria, sickle cell disease, polycythemia vera, and aplastic anemia). Normalized expression levels of signature genes are used as input to ML-based classifiers to predict radiation exposure in other hematological conditions.

RESULTS

Except for aplastic anemia, these blood-borne disorders modify the normal baseline expression values of genes present in radiation signatures, leading to false-positive misclassification of radiation exposures in 8-54% of individuals. Shared changes, predominantly in DNA damage response and apoptosis-related gene transcripts in radiation and confounding hematological conditions, compromise the utility of these signatures for radiation assessment. These confounding conditions (sickle cell disease, thrombosis, S. aureus bacteremia, malaria) induce neutrophil extracellular traps, initiated by chromatin decondensation, DNA damage response and fragmentation followed by programmed cell death or extrusion of DNA fragments. Riboviral infections (e.g. influenza or dengue fever) have been proposed to bind and deplete host RNA binding proteins, inducing R-loops in chromatin. R-loops that collide with incoming replication forks can result in incompletely repaired DNA damage, inducing apoptosis and releasing mature virus. To mitigate the effects of confounders, we evaluated predicted radiation-positive samples with novel gene expression signatures derived from radiation-responsive transcripts encoding secreted blood plasma proteins whose expression levels are unperturbed by these conditions.

CONCLUSIONS

This approach identifies and eliminates misclassified samples with underlying hematological or infectious conditions, leaving only samples with true radiation exposures. Diagnostic accuracy is significantly improved by selecting genes that maximize both sensitivity and specificity in the appropriate tissue using combinations of the best signatures for each of these classes of signatures.

Gene expression for biodosimetry and effect prediction purposes: promises, pitfalls and future directions - key session ConRad 2021

PURPOSE

In a nuclear or radiological event, an early diagnostic or prognostic tool is needed to distinguish unexposed from low- and highly exposed individuals with the latter requiring early and intensive medical care. Radiation-induced gene expression (GE) changes observed within hours and days after irradiation have shown potential to serve as biomarkers for either dose reconstruction (retrospective dosimetry) or the prediction of consecutively occurring acute or chronic health effects. The advantage of GE markers lies in their capability for early (1-3 days after irradiation), high-throughput, and point-of-care (POC) diagnosis required for the prediction of the acute radiation syndrome (ARS).

CONCLUSIONS

As a key session of the ConRad conference in 2021, experts from different institutions were invited to provide state-of-the-art information on a range of topics including: (1) Biodosimetry: What are the current efforts to enhance the applicability of this method to perform retrospective biodosimetry? (2) Effect prediction: Can we apply radiation-induced GE changes for prediction of acute health effects as an approach, complementary to and integrating retrospective dose estimation? (3) High-throughput and point-of-care diagnostics: What are the current developments to make the GE approach applicable as a high-throughput as well as a POC diagnostic platform? (4) Low level radiation: What is the lowest dose range where GE can be used for biodosimetry purposes? (5) Methodological considerations: Different aspects of radiation-induced GE related to more detailed analysis of exons, transcripts and next-generation sequencing (NGS) were reported.

Transcriptomics for radiation biodosimetry: progress and challenges

PURPOSE

Transcriptomic-based approaches are being developed to meet the needs for large-scale radiation dose and injury assessment and provide population triage following a radiological or nuclear event. This review provides background and definition of the need for new biodosimetry approaches, and summarizes the major advances in this field. It discusses some of the major model systems used in gene signature development, and highlights some of the remaining challenges, including individual variation in gene expression, potential confounding factors, and accounting for the complexity of realistic exposure scenarios.

CONCLUSIONS

Transcriptomic approaches show great promise for both dose reconstruction and for prediction of individual radiological injury. However, further work will be needed to ensure that gene expression signatures will be robust and appropriate for their intended use in radiological or nuclear emergencies.

Estimating partial-body ionizing radiation exposure by automated cytogenetic biodosimetry

PURPOSE

Inhomogeneous exposures to ionizing radiation can be detected and quantified with the dicentric chromosome assay (DCA) of metaphase cells. Complete automation of interpretation of the DCA for whole-body irradiation has significantly improved throughput without compromising accuracy, however, low levels of residual false positive dicentric chromosomes (DCs) have confounded its application for partial-body exposure determination.

MATERIALS AND METHODS

We describe a method of estimating and correcting for false positive DCs in digitally processed images of metaphase cells. Nearly all DCs detected in unirradiated calibration samples are introduced by digital image processing. DC frequencies of irradiated calibration samples and those exposed to unknown radiation levels are corrected subtracting this false positive fraction from each. In partial-body exposures, the fraction of cells exposed, and radiation dose can be quantified after applying this modification of the contaminated Poisson method.

RESULTS

Dose estimates of three partially irradiated samples diverged 0.2-2.5 Gy from physical doses and irradiated cell fractions deviated by 2.3%-15.8% from the known levels. Synthetic partial-body samples comprised of unirradiated and 3 Gy samples from 4 laboratories were correctly discriminated as inhomogeneous by multiple criteria. Root mean squared errors of these dose estimates ranged from 0.52 to 1.14 Gy² and from 8.1 to 33.3%² for the fraction of cells irradiated.

CONCLUSIONS

Automated DCA can differentiate whole- from partial-body radiation exposures and provides timely quantification of estimated whole-body equivalent dose.

A proposed molecular mechanism for pathogenesis of severe RNA-viral pulmonary infections

Background: Certain riboviruses can cause severe pulmonary complications leading to death in some infected patients. We propose that DNA damage induced-apoptosis accelerates viral release, triggered by depletion of host RNA binding proteins (RBPs) from nuclear RNA bound to replicating viral sequences. Methods: Information theory-based analysis of interactions between RBPs and individual sequences in the Severe Acute Respiratory Syndrome CoronaVirus 2 (SARS-CoV-2), Influenza A (H3N1), HIV-1, and Dengue genomes identifies strong RBP binding sites in these viral genomes. Replication and expression of viral sequences is expected to increasingly sequester RBPs - SRSF1 and RNPS1. Ordinarily, RBPs bound to nascent host transcripts prevents their annealing to complementary DNA. Their depletion induces destabilizing R-loops. Chromosomal breakage occurs when an excess of unresolved R-loops collide with incoming replication forks, overwhelming the DNA repair machinery. We estimated stoichiometry of inhibition of RBPs in host nuclear RNA by counting competing binding sites in replicating viral genomes and host RNA. Results: Host RBP binding sites are frequent and conserved among different strains of RNA viral genomes. Similar binding motifs of SRSF1 and RNPS1 explain why DNA damage resulting from SRSF1 depletion is complemented by expression of RNPS1. Clustering of strong RBP binding sites coincides with the distribution of RNA-DNA hybridization sites across the genome. SARS-CoV-2 replication is estimated to require 32.5-41.8 hours to effectively compete for binding of an equal proportion of SRSF1 binding sites in host encoded nuclear RNAs. Significant changes in expression of transcripts encoding DNA repair and apoptotic proteins were found in an analysis of influenza A and Dengue-infected cells in some individuals. Conclusions: R-loop-induced apoptosis indirectly resulting from viral replication could release significant quantities of membrane-associated virions into neighboring alveoli. These could infect adjacent pneumocytes and other tissues, rapidly compromising lung function, causing multiorgan system failure and other described symptoms.

A proposed molecular mechanism for pathogenesis of severe RNA-viral pulmonary infections

Background: Certain riboviruses can cause severe pulmonary complications leading to death in some infected patients. We propose that DNA damage induced-apoptosis accelerates viral release, triggered by depletion of host RNA binding proteins (RBPs) from nuclear RNA bound to replicating viral sequences. Methods: Information theory-based analysis of interactions between RBPs and individual sequences in the Severe Acute Respiratory Syndrome CoronaVirus 2 (SARS-CoV-2), Influenza A (H3N2), HIV-1, and Dengue genomes identifies strong RBP binding sites in these viral genomes. Replication and expression of viral sequences is expected to increasingly sequester RBPs - SRSF1 and RNPS1. Ordinarily, RBPs bound to nascent host transcripts prevents their annealing to complementary DNA. Their depletion induces destabilizing R-loops. Chromosomal breakage occurs when an excess of unresolved R-loops collide with incoming replication forks, overwhelming the DNA repair machinery. We estimated stoichiometry of inhibition of RBPs in host nuclear RNA by counting competing binding sites in replicating viral genomes and host RNA. Results: Host RBP binding sites are frequent and conserved among different strains of RNA viral genomes. Similar binding motifs of SRSF1 and RNPS1 explain why DNA damage resulting from SRSF1 depletion is complemented by expression of RNPS1. Clustering of strong RBP binding sites coincides with the distribution of RNA-DNA hybridization sites across the genome. SARS-CoV-2 replication is estimated to require 32.5-41.8 hours to effectively compete for binding of an equal proportion of SRSF1 binding sites in host encoded nuclear RNAs. Significant changes in expression of transcripts encoding DNA repair and apoptotic proteins were found in an analysis of influenza A and Dengue-infected cells in some individuals. Conclusions: R-loop-induced apoptosis indirectly resulting from viral replication could release significant quantities of membrane-associated virions into neighboring alveoli. These could infect adjacent pneumocytes and other tissues, rapidly compromising lung function, causing multiorgan system failure and other described symptoms.

A High Throughput Approach to Reconstruct Partial-Body and Neutron Radiation Exposures on an Individual Basis

Biodosimetry-based individualized reconstruction of complex irradiation scenarios (partial-body shielding and/or neutron + photon mixtures) can improve treatment decisions after mass-casualty radiation-related incidents. We used a high-throughput micronucleus assay with automated scanning and imaging software on ex-vivo irradiated human lymphocytes to: a) reconstruct partial-body and/or neutron exposure, and b) estimate separately the photon and neutron doses in a mixed exposure. The mechanistic background is that, compared with total-body photon irradiations, neutrons produce more heavily-damaged lymphocytes with multiple micronuclei/binucleated cell, whereas partial-body exposures produce fewer such lymphocytes. To utilize these differences for biodosimetry, we developed metrics that describe micronuclei distributions in binucleated cells and serve as predictors in machine learning or parametric analyses of the following scenarios: (A) Homogeneous gamma-irradiation, mimicking total-body exposures, vs. mixtures of irradiated blood with unirradiated blood, mimicking partial-body exposures. (B) X rays vs. various neutron + photon mixtures. The results showed high accuracies of scenario and dose reconstructions. Specifically, receiver operating characteristic curve areas (AUC) for sample classification by exposure type reached 0.931 and 0.916 in scenarios A and B, respectively. R² for actual vs. reconstructed doses in these scenarios reached 0.87 and 0.77, respectively. These encouraging findings demonstrate a proof-of-principle for the proposed approach of high-throughput reconstruction of clinically-relevant complex radiation exposure scenarios.

Multigene signatures of responses to chemotherapy derived by biochemically-inspired machine learning

Pharmacogenomic responses to chemotherapy drugs can be modeled by supervised machine learning of expression and copy number of relevant gene combinations. Such biochemical evidence can form the basis of derived gene signatures using cell line data, which can subsequently be examined in patients that have been treated with the same drugs. These gene signatures typically contain elements of multiple biochemical pathways which together comprise multiple origins of drug resistance or sensitivity. The signatures can capture variation in these responses to the same drug among different patients.

Predicting responses to platin chemotherapy agents with biochemically-inspired machine learning

The selection of effective genes that accurately predict chemotherapy responses might improve cancer outcomes. We compare optimized gene signatures for cisplatin, carboplatin, and oxaliplatin responses in the same cell lines and validate each signature using data from patients with cancer. Supervised support vector machine learning is used to derive gene sets whose expression is related to the cell line GI₅₀ values by backwards feature selection with cross-validation. Specific genes and functional pathways distinguishing sensitive from resistant cell lines are identified by contrasting signatures obtained at extreme and median GI₅₀ thresholds. Ensembles of gene signatures at different thresholds are combined to reduce the dependence on specific GI₅₀ values for predicting drug responses. The most accurate gene signatures for each platin are: cisplatin: BARD1, BCL2, BCL2L1, CDKN2C, FAAP24, FEN1, MAP3K1, MAPK13, MAPK3, NFKB1, NFKB2, SLC22A5, SLC31A2, TLR4, and TWIST1; carboplatin: AKT1, EIF3K, ERCC1, GNGT1, GSR, MTHFR, NEDD4L, NLRP1, NRAS, RAF1, SGK1, TIGD1, TP53, VEGFB, and VEGFC; and oxaliplatin: BRAF, FCGR2A, IGF1, MSH2, NAGK, NFE2L2, NQO1, PANK3, SLC47A1, SLCO1B1, and UGT1A1. Data from The Cancer Genome Atlas (TCGA) patients with bladder, ovarian, and colorectal cancer were used to test the cisplatin, carboplatin, and oxaliplatin signatures, resulting in 71.0%, 60.2%, and 54.5% accuracies in predicting disease recurrence and 59%, 61%, and 72% accuracies in predicting remission, respectively. One cisplatin signature predicted 100% of recurrence in non-smoking patients with bladder cancer (57% disease-free; N = 19), and 79% recurrence in smokers (62% disease-free; N = 35). This approach should be adaptable to other studies of chemotherapy responses, regardless of the drug or cancer types.

Predicting responses to platin chemotherapy agents with biochemically-inspired machine learning

The selection of effective genes that accurately predict chemotherapy responses might improve cancer outcomes. We compare optimized gene signatures for cisplatin, carboplatin, and oxaliplatin responses in the same cell lines and validate each signature using data from patients with cancer. Supervised support vector machine learning is used to derive gene sets whose expression is related to the cell line GI₅₀ values by backwards feature selection with cross-validation. Specific genes and functional pathways distinguishing sensitive from resistant cell lines are identified by contrasting signatures obtained at extreme and median GI₅₀ thresholds. Ensembles of gene signatures at different thresholds are combined to reduce the dependence on specific GI₅₀ values for predicting drug responses. The most accurate gene signatures for each platin are: cisplatin: BARD1, BCL2, BCL2L1, CDKN2C, FAAP24, FEN1, MAP3K1, MAPK13, MAPK3, NFKB1, NFKB2, SLC22A5, SLC31A2, TLR4, and TWIST1; carboplatin: AKT1, EIF3K, ERCC1, GNGT1, GSR, MTHFR, NEDD4L, NLRP1, NRAS, RAF1, SGK1, TIGD1, TP53, VEGFB, and VEGFC; and oxaliplatin: BRAF, FCGR2A, IGF1, MSH2, NAGK, NFE2L2, NQO1, PANK3, SLC47A1, SLCO1B1, and UGT1A1. Data from The Cancer Genome Atlas (TCGA) patients with bladder, ovarian, and colorectal cancer were used to test the cisplatin, carboplatin, and oxaliplatin signatures, resulting in 71.0%, 60.2%, and 54.5% accuracies in predicting disease recurrence and 59%, 61%, and 72% accuracies in predicting remission, respectively. One cisplatin signature predicted 100% of recurrence in non-smoking patients with bladder cancer (57% disease-free; N = 19), and 79% recurrence in smokers (62% disease-free; N = 35). This approach should be adaptable to other studies of chemotherapy responses, regardless of the drug or cancer types.

Predicting responses to platin chemotherapy agents with biochemically-inspired machine learning

The selection of effective genes that accurately predict chemotherapy responses might improve cancer outcomes. We compare optimized gene signatures for cisplatin, carboplatin, and oxaliplatin responses in the same cell lines and validate each signature using data from patients with cancer. Supervised support vector machine learning is used to derive gene sets whose expression is related to the cell line GI₅₀ values by backwards feature selection with cross-validation. Specific genes and functional pathways distinguishing sensitive from resistant cell lines are identified by contrasting signatures obtained at extreme and median GI₅₀ thresholds. Ensembles of gene signatures at different thresholds are combined to reduce the dependence on specific GI₅₀ values for predicting drug responses. The most accurate gene signatures for each platin are: cisplatin: BARD1, BCL2, BCL2L1, CDKN2C, FAAP24, FEN1, MAP3K1, MAPK13, MAPK3, NFKB1, NFKB2, SLC22A5, SLC31A2, TLR4, and TWIST1; carboplatin: AKT1, EIF3K, ERCC1, GNGT1, GSR, MTHFR, NEDD4L, NLRP1, NRAS, RAF1, SGK1, TIGD1, TP53, VEGFB, and VEGFC; and oxaliplatin: BRAF, FCGR2A, IGF1, MSH2, NAGK, NFE2L2, NQO1, PANK3, SLC47A1, SLCO1B1, and UGT1A1. Data from The Cancer Genome Atlas (TCGA) patients with bladder, ovarian, and colorectal cancer were used to test the cisplatin, carboplatin, and oxaliplatin signatures, resulting in 71.0%, 60.2%, and 54.5% accuracies in predicting disease recurrence and 59%, 61%, and 72% accuracies in predicting remission, respectively. One cisplatin signature predicted 100% of recurrence in non-smoking patients with bladder cancer (57% disease-free; N = 19), and 79% recurrence in smokers (62% disease-free; N = 35). This approach should be adaptable to other studies of chemotherapy responses, regardless of the drug or cancer types.

Machine learning has identified genetic signatures that predict how patients will respond to three of the most widely used cancer drugs. Chemotherapy regimens are usually based on how groups of people with similar cancers respond to them, but genetic differences can render the drugs more or less effective in individual patients. Machine learning provides a way of sifting through large amounts of data to identify patterns—in this case, in gene signatures associated with cancer recurrence and remission. The authors investigated cellular responses to cisplatin, carboplatin, and oxaliplatin and identified signatures in 11–15 genes which were the most predictive for each drug. The compositions of these signatures are also tailored to how well these therapies prevent growth of cancer cells. Accuracy varied, but one cisplatin signature was able to predict all instances of disease recurrence in non-smokers with bladder cancer.