key: cord-0325857-el2hfw7a
authors: Rockey, Nicole C.; Henderson, James B.; Chin, Kaitlyn; Raskin, Lutgarde; Wigginton, Krista R.
title: Predictive modeling of virus inactivation by UV
date: 2020-10-28
journal: bioRxiv
DOI: 10.1101/2020.10.27.355479
sha: 5f524d44277e21cb9b2d5d22fb163d82e6047c52
doc_id: 325857
cord_uid: el2hfw7a

Disinfection strategies are commonly applied to inactivate pathogenic viruses in water, food, air, and on surfaces to prevent the spread of infectious diseases. Determining how quickly viruses are inactivated to mitigate health risks is not always feasible due to biosafety restrictions or difficulties with virus culturability. Therefore, methods that would rapidly predict kinetics of virus inactivation by UV254 would be valuable, particularly for emerging and difficult-to-culture viruses. We conducted a rapid systematic literature review to collect high-quality inactivation rate constants for a wide range of viruses. Using these data and basic virus information (e.g., genome sequence attributes), we developed and evaluated four different model classes, including linear and non-linear approaches, to find the top performing prediction model. For both the (+) ssRNA and dsDNA virus types, multiple linear regressions were the top performing model classes. In both cases, the cross-validated root mean squared relative prediction errors were similar to those associated with experimental rate constants. We tested the models by predicting and measuring inactivation rate constants for two viruses that were not identified in our systematic review, including a (+) ssRNA mouse coronavirus and a dsDNA marine bacteriophage; the predicted rate constants were within 7% and 71% of the experimental rate constants, respectively. Finally, we applied our models to predict the UV254 rate constants of several viruses for which high-quality UV254 inactivation data are not available. Our models will be valuable for predicting inactivation kinetics of emerging or difficult-to-culture viruses.

Viruses can cause diverse and costly illnesses in humans and other animals (1). A variety 31 of approaches have therefore been developed to decontaminate food, water, air, and surfaces that 32 may contain infective viruses (2-7). UV254 treatment, in particular, is gaining popularity as an 33 alternative to more traditional chemical disinfection strategies (8-10). Viruses can have highly 34 variable UV254 susceptibilities (11, 12) . For example, two dsDNA viruses, adenovirus type 40 and 35 bacteriophage T6, are inactivated by UV254 at the widely varying rates of ~ 0.06 cm 2 mJ -1 (13-18) 36 and ~ 5.4 cm 2 mJ -1 (19), respectively. 37

Viruses have diverse genome types, including double-stranded RNA (dsRNA), single-38 stranded RNA (ssRNA), double-stranded DNA (dsDNA), and single-stranded DNA (ssDNA). 39 UV254 inactivates by primarily targeting viral genetic material, and the different biochemical 40 structures associated with these viral genome types result in distinct sensitivities to UV254 (20) . 41

Nucleic acid primary structure, or nucleotide base sequence, also affects UV254 genome reactivity 42 -pyrimidine bases, for instance, are about an order of magnitude more reactive with UV254 than 43 purine bases (21, 22). Different replication modes among viruses can also impact susceptibility to 44 UV254. For example, the reverse transcriptase enzymes involved in generation of retrovirus mRNA 45 may have different fidelities to photochemical modifications in nucleic acid compared to the RNA 46 dependent RNA polymerase enzymes used by other RNA viruses to synthesize mRNA (23). 47

Additional differences in viral infection cycles impact virus sensitivity to UV254 (24). dsDNA virus 48 genomes, for example, can undergo nucleic acid repair once inside host cells (24-26). This means 49 that a virus may be inactivated by UV254 treatment through base modification, only to be repaired 50 and thus rendered infectious again when such repair mechanisms are available. We note these 51 repair (34-37) impacts virus UV254 sensitivity. We included categorical predictors for genome 170 repair mode (i.e., host cell mediated, virus-gene controlled using one repair system, or virus-gene 171 controlled using multiple repair systems) and host cell type (i.e., prokaryotic host, eukaryotic host 172 with wild type repair, or eukaryotic host with reduced repair) in the dsDNA virus inactivation rate 173 constant models. Genome repair mode and host cell type were assigned based on available 174 information and are described in the SI Appendix. 175

The RMSrPE of the four optimized dsDNA model classes ranged from 0.31 to 1.6 (SI 176 Appendix, Table S3 All-virus model. Larger data sets generally add predictive power to models, though the 206 increased signal from additional data can be attenuated or negated by increased heterogeneity. We 207 therefore compared the performance of the separate (+) ssRNA and dsDNA virus models with a 208 model that incorporated data from all Baltimore classes. In addition to the genomic variables and 209 repair-related predictors (i.e., genome repair mode and host cell type) included for (+) ssRNA and 210 dsDNA viruses, a categorical predictor for nucleic acid type (i.e., double-stranded or single-211 stranded) was included. Boosted trees models were the top performing models using all viruses 212 (SI Appendix, Table S3 ); these performed significantly worse than the models trained using only 213 (+) ssRNA viruses (RMSrPE(+) ssRNA = 0.22 ± 0.23, RMSrPEall = 0.45 ± 0.33; SI Appendix, Table  214 S5) or only dsDNA viruses (RMSrPEdsDNA = 0.31 ± 0.28 vs RMSrPEall = 0.45 ± 0.35; SI Appendix, 215 diverse genome types and infection cycles into one model can negatively impact performance of 217 virus predictions, possibly owing to insufficient data from less studied classes. Based on these 218 results, we used the separate (+) ssRNA and dsDNA models for subsequent analyses. 219

Predicted rate constants align with new experimental rate constants. We applied the 220 optimized (+) ssRNA and dsDNA models to predict the rate constants of one (+) ssRNA virus and 221 one dsDNA virus for which experimental data were not available and then measured the rate 222 constants experimentally. Specifically, we predicted and measured the rate constants for MHV, a 223 (+) ssRNA mouse coronavirus, and HS2, a dsDNA marine bacteriophage. Based on its large 224 genome size (i.e., ~ 270% longer than the largest (+) ssRNA virus genome included in the training 225 and validation set) MHV provided an opportunity to assess the (+) ssRNA model's predictive 226 power using a virus with attributes outside those in the training and validation set (SI Appendix, 227 doses; the measured MS2 rate constants were in line with those in the literature (0.12 to 0.14 cm 2 231 mJ -1 ; SI Appendix, Fig. S4 and Table S1 ). 232

The predicted inactivation rate constant for MHV (kpred = 2.05 ± 0.88 cm 2 mJ -1 ; mean ± 233 95% margin of error) was not significantly different than the experimental rate constant (kexp = 234 1.92 ± 0.17 cm -2 mJ -1 ), with a percent error of only 7% (Figure 4a high quality UV254 inactivation rate constants in the literature. We therefore applied the (+) ssRNA 260 and dsDNA predictive models to estimate the inactivation rates constants for several viruses, 261 including human norovirus, dengue virus, SARS-CoV-2, and several herpesviruses (Table 1) . 262

These predictions resulted in a range of inactivation rate constants, from 0.28 for human norovirus 263 to 3.0 cm 2 mJ -1 for human cytomegalovirus. 264 Table S2 ) but dissimilar UV254 inactivation rate constants (5.1 cm -2 302 mJ -1 for T2 and 1.7 cm -2 mJ -1 for T4; SI Appendix, Table S1 ). T4 phage's UV254 resistance is due 303 to an additional virus-controlled repair gene in the T4 genome not present in the T2 genome (50, 304 51). Interestingly, the relative contribution of genomic variables in the dsDNA model was 305 significantly less than the genome repair predictors, which suggests that genome repair is a more 306 important factor in dsDNA UV254 inactivation than genomic variables. 307

Including genome repair as a model predictor presented some limitations. First, the mode 308 and extent of genome repair is not known for many viruses and has not been well-studied across 309 virus families. A single predictor encompassing the contribution of genome repair was therefore 310 not possible. We instead applied multiple categorical predictors. With this approach, only viruses 311 that shared a particular genome repair mode or host cell type with at least one other virus in the 312 dsDNA data set could be used in cross-validation. Ultimately, the data set used for dsDNA model 313 development and validation lacked numerous forms of dsDNA viruses with distinct repair modes 314 and host cell types, resulting in uncertainty in model performance for certain dsDNA viruses not 315 represented in the training and validation set. To improve future dsDNA virus models, it is critical 316 to have a better understanding of genome repair mechanisms and how they affect UV254 317 inactivation. 318

Our top performing UV254 virus prediction models provide improvements over earlier 319 prediction approaches (28, 29). On average, the (+) ssRNA and dsDNA virus models predicted 320 rate constants to within ~0.2x and ~0.3x of experimental constants, respectively. A previous 321 approach using genome length to determine genome size-normalized sensitivity values for a 322 number of virus families expected uncertainties in predicted values of ~2x (28). A more recent 323 approach developed predictive models for ssRNA and dsDNA UV254 inactivation using genome 324 dimer formation potential, a value that incorporated pyrimidine doublets, genome length, and 325 purines with adjacent pyrimidine doublets (29). Their reported error as a coefficient of 326 determination (i.e., R 2 ) was 0.67 for ssRNA viruses compared to 0.74 (adjusted R 2 ) for our model, 327

and an R 2 value of 0.62 for dsDNA viruses compared to 0.99 (adjusted R 2 ) for our model. Several 328 factors can be attributed to the improved performance of our models, including extensive curation 329 of data based on quality and the incorporation of genome repair into dsDNA modeling. 330

In light of the coronavirus disease 2019 (COVID-19) pandemic and the need for effective 331 decontamination strategies, our predictive models provided an opportunity to predict rate constants 332 for a critical group of viruses with very little published inactivation data. Limited data on UV254 333 inactivation for coronaviruses in aqueous suspension are available and the published information 334 did not pass the inclusion criteria of our rapid systematic review (10, 52-54). This paucity of 335 information on the susceptibility of coronaviruses to UV254 is of critical importance for developing 336 effective decontamination strategies. Our predicted rate constants for SARS-CoV-1, SARS-CoV-337 2, and MERS, and our measured rate constant for the mouse coronavirus MHV, suggest that 338 coronaviruses are much more susceptible to UV254 inactivation than other (+) ssRNA viruses. A 339 recent estimate of SARS-CoV-2 UV254 susceptibility using the previously developed Lytle and 340

Sagripanti approach (28) is ~ 1.7x greater than our estimate indicates (55). Discrepancies in new 341 experimental coronavirus data still persist, likely stemming from a lack of checks on UV254 342 attenuation of suspensions. 343

More robust models are possible with larger data sets that consist of more diverse viruses. 344

Unfortunately, a large portion of UV254 inactivation data found during the rapid systematic review 345 did not pass our inclusion criteria. The most common reason for excluding data from our 346 systematic review was a failure to report solution UV254 attenuation. An earlier study of SARS-347 CoV-1 inactivation by UV254 (54), for example, did not account for UV254 attenuation in the 348 experimental DMEM suspension. The reported inactivation rate constant of 0.003 cm 2 mJ -1 was 349 nearly three orders of magnitude lower than our predicted rate constant for SARS-CoV-1 and our 350 measured value for MHV, likely in part due to solution attenuation. We estimate that their rate 351 constant would be closer to 0.35 cm 2 mJ -1 after accounting for solution attenuation. This value 352 more closely aligns with our coronavirus values. Similarly, several studies reported UV254 353 inactivation of viruses in blood products without describing how attenuation was considered in 354 their reported doses (10, 56-58). Although these doses are likely representative for these fluids, 355 they cannot be extrapolated to other matrices. More stringent reporting of UV254 experimental 356 conditions (59), including matrix solution transmission at 254 nm, will facilitate future modeling 357 efforts. We note when UV254 inactivation rate constants are known for a solution with 100% 358 transmittance (e.g., purified virus in buffer solution), the rate constant can be adjusted to account 359 for a solution with significant attenuation (e.g., blood products) based on the Beer-Lambert law 360 (60). 361

The developed models allow us to predict the effectiveness of current UV254 treatment 362 strategies on viral pathogens that are difficult or impossible to culture. For example, human 363 norovirus, which causes gastrointestinal disease, is a major target of UV254 disinfection processes 364 in water treatment and food processing. Our (+) ssRNA virus model predicts an inactivation rate 365 constant of 0.28 cm 2 mJ -1 for human norovirus GII.4, which is similar to our recently reported rate 366 constant of k = 0.27 cm 2 mJ -1 for human norovirus GII.4 Sydney using RT-qPCR data coupled 367 with a full-genome extrapolation approach (61). This finding indicates that current water treatment 368 guidelines for adequate UV254 virus inactivation, which are defined to treat adenovirus 41 (62), are 369 more than sufficient to inactivate human norovirus to acceptable levels. In fact, none of the viruses 370 for which we predicted rate constants had UV254 resistance greater than viruses in the Adenoviridae 371 family. 372

The limited and unbalanced data set that we obtained from the systematic review and used 373 in modeling efforts created challenges in our modeling work. Of primary concern, we could not 374 take a commonly used approach to evaluating models, in which a portion of data is held back 375 during model development to assess performance. Holding back the typical 10 -20% of data 376 would correspond to holding back only two to four viruses from the (+) ssRNA or dsDNA classes 377 for testing. This could result in high variance estimates of prediction performance that would also 378 be highly dependent on the viruses withheld during training. We consequently used leave-one-379 virus-out cross-validation to more efficiently estimate prediction performance on out of sample 380 data. Another limitation of our models is that they were developed and validated for only (+) 381 emerging and noteworthy human viruses belong to other classes. In particular, the (-) ssRNA virus 383 class includes several important human pathogens, such as lassa virus, nipah virus, influenza virus, 384 and ebolavirus. Since only two (-) ssRNA viruses were included in our data set, we were unable 385 to assess whether inactivation rate constants for viruses in this group could be accurately predicted 386 with our (+) ssRNA model. More high quality UV254 experimental inactivation data for a broader 387 set of viruses would facilitate the holdout approach for validating models and the development of 388 models for other virus Baltimore classification groups. 389

This research demonstrates the value of predictive models for estimating virus fate in 390 various settings. Using readily available viral genome data, we developed models to predict UV254 391 inactivation of (+) ssRNA and dsDNA viruses. The benefits of predictive models are underlined 392 by the ongoing COVID-19 pandemic: access to the biosafety level 3 laboratories required to work 393 with SARS-CoV-2 has been limited and, as a result, few experimental inactivation studies have 394 been performed. Our approach can rapidly determine virus susceptibility to UV254 using available 395 genomes, but without relying on culture systems that are often unavailable or difficult to access. 396

Other potential applications of our models including identifying outlier UV254 data that are 397 published and predicting potential worst-case scenarios for viruses and their susceptibility to 398 UV254. Ultimately, we expect that this predictive modeling approach can be applied to estimate 399 inactivation of microorganisms with other disinfectants and in different settings, such as on 400 surfaces or in air. nm; viruses were irradiated in a liquid suspension; infective viruses were enumerated with 408 quantitative culture-based approaches (e.g., plaque assay); attenuation through the sample solution 409 was taken into account, or negligible UV254 attenuation was reported (transmittance > 95%) or 410 could be assumed based on the reported viral stock purification techniques and matrix solution 411 composition; stirring was reported when attenuation was significant (transmittance < 95%); first-412 order kinetics were reported or could be confirmed with reported data points for at least two UV254 413 fluences; the first-order inactivation rate constant or log-removal dose (e.g., D99) was provided or 414 could be determined with data presented in a plot or 

where k " v is the inverse variance weighted mean for the virus, n is the number of experimental rate 432 constants for the virus, ki is the inactivation rate constant for experiment i, and wi is the weight for 433 experiment i, defined as: 434

where SE,i is the standard error of the inactivation rate constant for experiment i. The standard 436 error of the inverse variance weighted mean, SEv, was evaluated for each virus as: 437

We estimated the inter-experimental error for viruses with more than one experimental rate 439 predictors nucleic acid type, genome repair mode, and host cell type were not included in the PCA. 464

We then developed linear regression models containing either the first, first and second, or first, 465 second, and third principal components, as well as the other predictors. Only the first through third 466 principal components were assessed for inclusion in the linear regression models, because they 467 cumulatively explained 97% of the variation in genomic variables. Genomic variables were 468 standardized to unit variance prior to PCA to eliminate dissimilarities in the magnitude of variable 469 values. Linear regression can include one or more predictors that can affect model accuracy. We 470 therefore used best subset selection to evaluate a wide range of potential multiple linear regression 471 models. 472

Elastic net regularization. As an alternative to best subset selection, we considered linear 473 regression with parameter regularization using L1 ("Lasso") and L2 ("Ridge") penalties, a 474 technique known as the elastic net. We used the 'glmnet' package in R to create models with elastic 475 net regularization. The alpha and lambda hyperparameters, which control the relative contribution 476 and overall scale of the L1 and L2 penalties, respectively, were tuned using a grid search to find 477 the optimal hyperparameters for the data set as determined by leave-one-virus-out cross-validation. 478

Specifically, 11 different values ranging from 0 to 1 with a step of 0.1 were assessed for the 479 hyperparameter alpha, and 100 different lambda values were evaluated for each alpha. 480

Random forests. To accommodate the use of the modified inverse variance weights, the 481 random forests model was developed in R using the 'xgboost' package with a single round of 482 boosting, and other hyperparameters were set to match defaults from the 'randomForest' package 483 as well as possible (66). 484

Boosted trees. Boosted trees modeling was conducted using the 'xgboost' package in R. 485

The number of boosting rounds was selected to minimize the cross-validated error. The 486 hyperparameters for learning rate, tree depth, and minimum terminal node weight were 0.3, 6, and 487 1, respectively. 488

bacteriophage HS2. To consider how well the models may predict inactivation of a virus not 490 already included in the collected data set, we determined the UV254 inactivation rate constant of 491 MHV, a virus in the Coronaviridae family and Betacoronavirus genus, and of HS2, a marine 492 UV254 inactivation of viruses. All UV254 inactivation experiments were conducted with a 495 custom-made collimated beam reactor containing 0.16 mW cm -2 lamps (model G15T8, Philips). 496 UV254 irradiance was determined using chemical actinometry (67, 68) and MS2 (ATCC 15597-497 B1) was included in all experimental solutions as a biodosimeter to further confirm UV254 doses. 498

Infective MS2 was assessed using the double agar overlay approach with host Escherichia coli 499 (ATCC 15597) (69). For each UV254 exposure, 2 mL of the experimental solution was added to a 500 10 mL glass beaker and continuously stirred. Sample solution depth (0.8 cm) and transmittance (~ 501 47% to 53% for MHV experiments, ~ 79% to 80% for HS2 experiments) were used to determine 502 the average UV254 irradiance of the sample according to the Beer-Lambert law (60). Infective 503 viruses were assayed immediately following experiments. Dark controls were conducted with each 504 experiment and consisted of the virus suspended in experimental solution but stored in the dark on 505 ice for the duration of experiments. Three independent replicates were conducted for each 506 inactivation experiment. 507

For MHV experiments, solutions contained MHV and MS2 diluted in 1X PBS to a final 508 concentration of ~ 10 5 pfu/mL and ~ 10 10 pfu/mL, respectively. Samples were exposed to UV254 509 for 0 s, 5 s, 15 s, 25 s, and 35 s, which corresponded to UV254 doses of approximately 0 mJ cm -2 , 510 0.62 mJ cm -2 , 1.2 mJ cm -2 , 1.9 mJ cm -2 , 3.1 mJ cm -2 , and 4.3 mJ cm -2 . MS2 infectivity was assayed 511 after larger UV254 doses due to its slower inactivation kinetics, namely 37 mJ cm -2 and 74 mJ cm -512 2 . For HS2 experiments, solutions contained HS2 and MS2 diluted in 1X PBS to a final 513 concentration of ~ 10 8 pfu/mL and ~ 10 9 pfu/mL, respectively. Samples were irradiated for 0 s, 514 by the following equation: 518 5 6 6 & 7 = 89: • <=>?@ (4) 519

where C0 and C are infectious virus concentrations before and after UV254 exposure, respectively, 520 in pfu/mL, and DUV254 is the average UV254 dose, in mJ cm -2 . 521 Experimental inactivation rate constants (i.e., kexp) were determined with linear regression 522 analyses conducted in Prism version 8.4.2 (GraphPad) to obtain experimental inactivation rate 523 constants (i.e., kexp). UV254 inactivation curves for some viruses exhibited tailing at high doses. In 524 these situations, only the linear portions of the inactivation curves were included in the linear 525 regression analyses. 526 MHV and HS2 inactivation rate constant prediction. The UV254 inactivation rate constants 527 of MHV and HS2 were predicted using the best-performing inactivation models for (+) ssRNA 528 viruses and dsDNA viruses, respectively. The MHV genome sequence was provided by Dr. 529

Leibowitz (SI Appendix, Supplementary Text File S1), and the HS2 genome sequence is available 530 in NCBI (accession no. KF302036). 531

rates of several emerging and difficult-to-culture viruses, including SARS-CoV-2, were predicted 533 using the best-performing inactivation model. Sequence data for these viruses were obtained from 534 NCBI and all viruses with sequence information are included in SI Appendix, Table S2 . 535

Foundation project #1545756. 541

Viral Inactivation in Foods: A Review of Traditional and Novel 549

Food-Processing Technologies

Healthcare Infection Control Practices Adviosry Committee, 551 "Guideline for Disinfection and Sterilization in Healthcare Facilities

Disinfection of Microbial Aerosols" in Modeling the Transmission 553 and Prevention of Infectious Disease

Water quality engineering

Wastewater engineering : treatment and resource recovery

Efficacy of ultraviolet (UV-C) light in reducing foodborne 559 pathogens and model viruses in skim milk

The effect of ultraviolet-561

C technology on viral infection incidence in a pediatric long-term care facility

UV-C treatment on the safety of skim milk: Effect on microbial 564 inactivation and cytotoxicity evaluation

Inactivation of Ebola virus and Middle East respiratory syndrome 566 coronavirus in platelet concentrates and plasma by ultraviolet C light and methylene blue 567

Human T-cell leukaemia virus type I is highly sensitive to UV-C light

Removal of enteric 573 viruses and selected microbial indicators by UV irradiation of secondary effluent

Effect of Host Cells on Low-and Medium-Pressure UV 579

Feline Calicivirus and Adenovirus Type 40 by UV Radiation

Comparative inactivation of enteric adenoviruses, poliovirus and 584 coliphages by ultraviolet irradiation

Inactivation of Pathogens with Innovative UV Technologies

Water Works Association Research Foundation

Genetic Recombinations Leading to Production of Active 588

Bacteriophage from Ultraviolet Inactivated Bacteriophage Particles

Photochemistry of the Nucleic Acids

Suppression of hydrate and dimer formation in ultraviolet-597 irradiated poly (A + U) relative to poly U

Virus Type 1 (HIV-1) by Ultraviolet and X Irradiation

Biological effects of ultraviolet radiation

Harm, Gene-controlled reactivation of ultraviolet-inactivated bacteriophage

Cellular reactivation of ultraviolet-irradiated human adenovirus 2 in normal 604 and xeroderma pigmentosum fibroblasts

Expression of animal virus genomes

Predicted inactivation of viruses of relevance to biodefense 607 by solar radiation

A Genomic Model for Predicting the 609

Ultraviolet Susceptibility of Viruses

Applied Predictive Modeling

On the relationship between host-cell reactivation and UV-reactivation in UV-613

Studies on Repair of Adenovirus 2 by Human Fibroblasts Using Normal

Xeroderma Pigmentosum, and Xeroderma Pigmentosum Heterozygous Strains

Defective repair of UV-damaged DNA in human tumor and SV40-621 transformed human cells but not in adenovirus-transformed human cells

DNA repair in species with extreme lifespan differences

Host Cell Reactivation of Epstein-Barr Virus in Normal and 626

Host-cell Reactivation in Mammalian Cells

Physical and Chemical Changes Induced in Nucleic Acids by Ultraviolet 630

The disappearance of thymine dimers from DNA: an error-632 correcting mechanism

Origin of ultraviolet damage in DNA

Thymine Dimerization in DNA Is an Ultrafast Photoreaction

Contribution of thymine dimers to the ultraviolet light inactivation of 641 mutants of bacteriophage T4

Sequence-dependent thymine 643 dimer formation and photoreversal rates in double-stranded DNA

Pyrimidine hydrates and dimers in ultraviolet-646 irradiated tobacco mosaic virus ribonucleic acid

UV-induced DNA damage and repair: a review. Photochem

Ultraviolet Irradiation of Herpes Simplex Virus (Type 650 1): Delayed Transcription and Comparative Sensitivities of Virus Functions

Specificity of DNA Lesion Bypass by the Yeast DNA Polymerase η

Prediction complements explanation in 655 understanding the developing brain

The genetic control of ultraviolet sensitivity levels in bacteriophages T2 657 and T4

Mutants of phage T4 with increased sensitivity to ultraviolet

Evaluation of inactivation methods for severe acute 663 respiratory syndrome coronavirus in noncellular blood products

Inactivation of the 666 coronavirus that induces severe acute respiratory syndrome, SARS-CoV

Estimated Inactivation of Coronaviruses by Solar Radiation 669

With Special Reference to COVID-19

Inactivation of dengue, chikungunya, and Ross River viruses in 671 platelet concentrates after treatment with ultraviolet C light

Evaluation of the effectiveness of the SurePure Turbulator ultraviolet-674

C irradiation equipment on inactivation of different enveloped and non-enveloped viruses 675 inoculated in commercially collected liquid animal plasma

A novel approach to pathogen reduction in platelet concentrates using 677 short-wave ultraviolet light

Standardization of Methods for Fluence

Determination in Bench-Scale UV Experiments

Absorption Effects in Volume Irradiation of Microorganisms

The long term 2 enhanced surface water treatment rule. EPA-HQ

A scoping review of rapid review methods

Expediting systematic reviews: methods and 689 implications of rapid reviews

Core Team, R: A Language and Environment for Statistical Computing

66. xgboost developers, Random Forests in XGBoost

The lodide/lodate Actinometer in UV Disinfection: 693 Determination of the Fluence Rate Distribution in UV Reactors

Potassium Iodide as a Chemical Actinometer for 254 nm Radiation: Use of 696 lodate as an Electron Scavenger

United States Environmental Protection Agency

Somatic Coliphage in Water by Two-step Enrichment Procedure

The authors declare no competing financial interests. 544