Graduate Research Fellows | Session 5 from the 2026 NIJ Forensic Science Research and Development Webinar Series

Tracey Johnson: Welcome to the NIJ Forensic Science Research and Development Webinar Series. We are excited for you to join us. My name is Tracy Johnson, and I'm a supervisory physical scientist in the Office of Investigative and Forensic Sciences at NIJ, and I'll be monitoring today's session. The National Institute of Justice, or NIJ, is the research, evaluation, and technology agency of the U.S. Department of Justice, and is dedicated to improving knowledge and understanding of crime and justice issues through science.

Today's webinar will present research from our graduate research fellows, highlighting transformative research across several forensic disciplines. Presentations will highlight technology innovations such as bioinformatics, high-res mass spec, and molecular biology, operational efficiency such as the development of high-throughput user-friendly methods, enhanced accuracy through the implementation of multiple validation approaches and quality controls, accessibility through open access resources and compatible workflows for crime labs, and impact for public safety by expanding investigative capabilities in the forensic sciences.

Next slide.

Before we begin, we'll review some important information. To adjust your audio settings, click on the down arrow button next to the grayed-out mic icon to adjust as needed. Questions can be submitted at any time and will be answered after each presentation during a Q&A session. To submit your questions, click the Q&A icon at the top of your screen.

Please note, we may not be able to address all questions. Closed captioning is available. To access, click the More button with the three dots and select Language and Speech, and then select Show Live Captions. A recording of the webinar will be posted to NIJ's website in the coming days, and certificates are not being issued for this webinar series.

Today's presenters are in the following order: Betzaida Maldonado, PhD candidate from the University of Colorado Anschutz. Jack Trimble, PhD candidate from the University of Central Florida, and Akshita Verma, PhD candidate from Florida International University.  

Next slide.

And with that, I will turn the floor over to Betzaida for her presentation.  

Betzaida Maldonado: Good afternoon, and thank you so much for this opportunity. I'm excited to present my work, which is entitled Forensic DNA Mixture Interpretation of SNP Sequencing Data. Next. First of all, I just want to provide a motivation for why I'm interested in sequencing. Over the last decade, advances in sequencing technologies have allowed us to obtain more genetic information than traditional forensic DNA testing.  

For example, improving our ability to identify human remains when they're highly degraded and enabling the use of forensic genetic genealogy, an approach that is able to identify relatives of a person of interest by searching a sample against profiles in commercial DNA databases, such as GenMatch. This can be used in cases where all resources have been exhausted, such was the case with the highly televised identification of the Golden State Killer. Next. It's important to highlight that I'm specifically talking about sequencing data that detects SNPs or single nucleotide polymorphisms, which are single base variations in our genome, where oftentimes only two alleles exist. Importantly, when talking about sequencing data, we're moving from traditionally testing 20 to 26 DNA sites, or for some of you who might be familiar with STR locations, to testing hundreds to thousands of sites, therefore obtaining a lot more genetic information.  

Next.

One key challenge in forensic genetics is DNA mixtures, which are samples that contain the genetic contribution of multiple individuals. This is, of course, not always known at the time of sample collection, but becomes a problem during data interpretation. Through mixture interpretation tools, it is possible to understand the underlying genetic contribution of each individual, and perhaps even assign major or minor contributors using a mixture ratio or a proportion, which is simply a value between zero and one and indicates the amount of DNA that each person is contributing.

Next. Even though there are commonly used mixture interpretation tools for STR data, or the data that has been traditionally generated in crime laboratories, a current major gap in forensic genetics is that there's no widely accepted methods to interpret SNP mixture data. This limits the utility of sequencing technologies and hinders the interpretation of samples that are mixtures, which are commonly encountered in crime scenes.

Next. For that reason, the objective of our project is to develop a bioinformatics framework for the interpretation of SNP forensic mixtures. We hope that this work further enables the implementation of SNP technologies in crime laboratories and also expands the number of samples that would be eligible for forensic genetic genealogy.  

Next. I want to provide an overview as to what the problem looks like with sequencing data. When a single individual is sequenced, the output will be a number of read counts detected given the genotypes of that individual. These read counts are a quantitative measure of the DNA that is detected and depends on the amount of DNA that is present. But it also depends on the sequencing depth of the technology that is being used. Now, the problem becomes when we have more than one individual, which going forward, I'll specifically be talking about two-person mixtures, in which the main questions then become: what are the genotypes of these individuals, and what is the relative contribution in that sample? In other words, we're trying to infer the genotypes of each individual and their relative mixture proportion. Next.

All of our work is using simulated data because it allows for flexibility and allows us to develop and test a model over a wide range of mixture proportions and conditions. We use a likelihood-based model to infer the mixture proportion, which we note as lambda, and then apply a probabilistic approach to infer the genotypes for contributor one, which is noted as C1, and for contributor 2, which is noted as C2. Essentially, our model identifies the parameters that best fit the data, and then we evaluate the accuracy of our estimates by comparing them to ground truth. Next.

The forensic application we began with is a scenario where in a two-person mixture, we have a reference sample for one of the individuals, say, for example, the victim. Therefore, we have two unknown parameters in this particular mixture problem, one of them being the genotype of the unknown contributor or that second individual, and then also the mixture proportion. This case is specifically relevant in forensic cases where an individual can be reasonably assumed, which can be the case for sexual assault evidence. Next.

Our simulation study tested our model in a wide range of mixture proportions from zero to 1. More specifically, we tested 90 different mixture proportions within that range. Each proportion was evaluated 30 different mixtures, so a total of 2,700 mixtures were evaluated in our simulation study. Again, we're just testing our model in the scenario I mentioned before, so we can assume the genotype of 1 contributor, and then we're trying to infer the genotype of the second contributor and the relative contribution of both individuals.

Next. So for each simulated mixture, we're asking two key questions. The first one being, did the model correctly estimate the mixture proportion? And did the model correctly infer the unknown contributor genotype? And as a reminder, lambda represents the proportion of DNA of the known contributor or contributor one. So for example, lambda of 0.60 means that 60% of the DNA within that sample is from contributor one, and the 40% is from contributor 2, or the unknown individual. We evaluated the performance across varying number of SNPs and sequencing depth. Next.  

So we begin by testing 100 SNPs at 100x sequencing depth, chosen just for simplicity as a baseline. I want to briefly orient you on these graphs, and in every graph how, percent going forward on the x-axis, we will have the true mixture proportion. That is what we know is ground truth from the simulated mixtures.

Now, focusing on the left-hand side graph, we have a mixture proportion accuracy, which is measured as root mean square error on the y-axis. And on the right-hand side, we have unknown genotype accuracy, measured as percent incorrect on the y-axis. For both of those figures, a value close to 0 means high accuracy, which is observable for most mixture proportions.

However, we note that the inference accuracy, both in mixture proportion and unknown genotypes, decreases as the unknown contributor becomes the minor. This is not necessarily surprising, given that at those mixture ratio values, we have less information or list read counts for that individual. Next.

Now, what would happen if we increased the number of SNPs? Say we have a different sequencing kit with more SNPs. Here on the left-hand side, we have the same graph I showed in the last slide, but the right graph is with the higher number of SNPs. So going from 100 SNPs to 10,000 SNP, which is possible with some sequencing kits, we show that accuracy remained the same.

However, the estimate variability decreased as is observed by the bars, the error bars on the right-hand side graph. Next. So far, we've evaluated the accuracy of two parameters, finding that inference accuracy decreases as the unknown contributor becomes the minor contributor, and that increasing the number of SNPs does not necessarily improve our accuracy, but it does reduce the variability. So the remaining question, or one obvious next question is, would increasing the sequencing depth help? Next.

Here, we have increased the sequencing depth, which increases the number of read counts for each individual in the mixture, to explore how changing this variable impacts accuracy. Starting with mixture proportion accuracy for both of these graphs, on the x-axis, again, we have the true mixture ratio from the simulation, from the simulated data. And on the y-axis, we have the parameter estimate. And we see an improvement in mixture ratio inference, going from being able to accurately infer this parameter up to a lambda of 0.7 to being able to accurately infer this parameter up to a lambda of 0.87, as can be seen on the right-hand side of a graph. Next.  

Now, importantly, it’s being able to determine how increasing the sequencing depth impacts genotype inference accuracy. We see the same trend on the left-hand side as the graph that I had shown before. And on the right-hand side, we can see that we're now able to accurately infer the genotype of the unknown contributor, as long as that mixture parameter is less than approximately. 0.87. Next. So in summary, we have developed a computational framework to simulate SNP mixtures and developed a model to infer mixture proportions and the genotypes of an unknown individual. Our simulation studies highlight the importance of sequencing depth and mixture interpretation using our model.

Of course, there's limitations to our work, and that is that all mixtures are computationally generated, that we assume two contributors, and that we haven't tested the model with different populations. Next. For that reason, some of our next steps involve applying the model to publicly available genomic data, extending our model to two and three unknown contributors, and incorporate parameters that mimic real case scenarios to eventually determine metrics for mixture deconvolution, and hopefully use those deconvoluted or inferred genotypes for forensic genetic genealogy.

Next.

I want to give a big shout out to my lab and my mentor, who has been so helpful in guiding me through this project, my university and my PhD program for their encouragement, and the Department of Justice, given that this work is supported through the NIJ Graduate Fellowship Award. Next. Happy to take any questions.  

Tracey Johnson: We're now moving on to the Q&A portion of the program. So please submit your questions in the Q&A box, but I'll start with the following. Betzaida, yesterday we heard from Dr. August Woerner about his efforts to lower barriers for crime labs to access SNP technology. By developing bioinformatics for low pass or low sequencing depth, you described improvements to unknown genotype inference accuracy with higher sequencing depth.

Do you have plans to enhance your bioinformatic tool to accommodate low-pass strategies?  

Betzaida Maldonado: Definitely. I found the talk very informative and it's something that, it's very important to evaluate in our model. I do think that there's going to have to be some changes to the algorithm that we're using and also the equations that we're using. But I do think that low pass whole genome sequencing needs to be looked at just because it is an affordable way to encourage crime laboratories to implement these technologies.  

Tracey Johnson: Thank you, Betzaida. I'm so glad you were able to join yesterday and hear Dr. August's, or Dr. Woerner's presentation. So next question. You describe your data as computationally generated and that your next steps include evaluating your tools with actual genomic data. Will this include an assessment of how biogeographic ancestry contributes to genotype inference accuracy?  

Betzaida Maldonado: Yeah, definitely. I, so the currently the data is all computer generated. So the way that we can just use publicly available data is by downloading it and then using those files and kind of loading them into our computer and being able to then use what we have developed as a bioinformatics framework to then essentially, in silico, mix those genomic regions or genomic data. As a population and geneticist, as somebody that is very interested in population genetics, I'm definitely interested in understanding how the model would perform in different populations.

And with that, I understand that allele frequencies may be different. So far, what we've seen in this particular model that we've developed is that the inference is really driven by the data, by the amount of read counts, which is why sequencing depth is really important. And when we try to incorporate some allele frequency differences, we didn't really see much of a change. But it's definitely something that I want to look at and be able to evaluate with the 1,000 Genomes Project.  

Tracey Johnson: Great. I'm so glad to hear that you're going to be considering that as an option. So next question, what did you use for the simulation? Was it pre-developed code or algorithm, or did you code it yourself? Also, what did you use to code? So I guess what programming language, I'm inferring.  

Betzaida Maldonado: Yeah, yeah. So I coded it myself, and part of the reason for that was, honestly, a learning opportunity. I was at the crime lab scientist for a few years, and so really stepping into bioinformatics was kind of like a new skill that I wanted to develop. And so I developed the code myself in Python.  

Tracey Johnson: Okay, thank you for that very much. One more question and then we will be moving on to our next presenter. Can you talk a little bit about what particular mixture ratios you will be studying? If you could go into a little bit more detail about what your plans are with that.  

Betzaida Maldonado: Yeah, I think that one of the things I was really excited about with this project and some of the data that we've seen is that we would be able to interpret an unknown contributor if they're, say, like contributing 50% of the mixture or more in a two-person mixture. I'm really interested in really being able to hone in or determine what the limit of detection would be or the limit of interpretation would be as they become the minor contributor. It's oftentimes the person of interest that is that forensic scientists are trying to identify could be that minor contributor.

So trying to see if we could modify our model so that we can kind of go at lower levels. And of course there would be some challenges, given that some data can, low level data can have additional challenge like stochastic effects. But I think that going to mixture proportions lower than 30% would be really important to look at and to see what our limit of interpretation would be.  

Tracey Johnson: All right, thank you very much, Betzaida, for your presentation. And now we will hear from Jack Trimble. Thank you, Jack.  

Jack Trimble: Thank you.

Hey, well, good afternoon. Thank you all for joining. I want to thank the NIJ R&D team for giving me this invite. And today I get to share a little bit about my PhD work, which is about evaluating and developing a chemical characterization technique of pollen for forensic investigations. Next. And so I think it's important to first kind of explain forensic palynology, as not a lot of people might be familiar with this sort of niche trace evidence, but as the name implies, forensic palynology or pollen analysis is the use of palynomorphs such as pollen and spores to aid in criminal investigations. Next. So the general idea of using pollen as evidence is that we can use it to link individuals or objects to a crime. There are a lot of features about pollen that make it useful to investigators.

One, it is produced in large quantities and is extremely resistant to decay, so it can persist in our environment for millions of years. Also, different locations do have different vegetative assemblages or flora, and so the presence of certain pollen can allow investigators to perform geospatial tracking of an object or person. And then lastly, plants do tend to bloom at specific times annually. And so the presence, again, of a specific species of pollen could provide some temporal estimations or like the creation of a timeline.  

Next slide.

And so traditionally, how pollen is characterized in a forensic setting is that these grains are analyzed using microscopy to look at different physical or morphological features such as size, shape, and sculpturing that tend to vary between taxa. And so I think this image kind of demonstrates that pretty well about how varied these pollen can be. And so palynologists want to be highly trained in palynology to be able to look at these grains under a microscope and then compare with reference libraries in order to make that identification. Next.

And so the goal for forensic palynologists would be to identify the species of origin that that pollen came from. And once that is established, you can then compare with comparative samples from a crime scene, for example, or you could look at something like a vegetation map pictured here, which would allow for more questioning or narrow their line of inquiry once they establish what plant that pollen has come from.

However, there are a couple limitations. Actually, it's probably several, but there's two that I want to highlight for this presentation alone. Next. The one being that these physical differences in shape and size are often not discriminatory to the species level. And so while palynologists can readily get to like a family or a genus, it can sometimes be extremely difficult to get to that species level, which might be needed in some cases.

And then a second major limitation is that there's a very large lack of training when it comes to palynology, specifically within a forensic setting. And so the accessibility of pollen analysis to most crime labs is highly. And so that kind of leads into my work... If we can go to the next slide....which is, can we design a novel pollen characterization method that is able to resolve that species level which we need, is more compatible with the modern crime lab, and is ideally non-destructive in nature so that we can still bring in that morphological characterization? Next.

And so with pollen, like I just mentioned, we have our physical or morphological characterization that's been tried and tested since about the 50s. However, a lot of research, I want to say a lot, there has been a lot of more recent research looking into the chemical characterization of pollen. So next. And there's two main groupings, I guess you can put the chemical characterization pollen.

The first one would be like genetic characterization, namely through DNA barcoding. I don't want to overstep and say this is where the majority of research has been, but within a forensic context, most of the chemical characterization has revolved around the genetic analysis. And so the idea with DNA barcoding is that there are regions within the plant genome that we can target and sequence and use as markers for taxonomic or species identification. Along with genetic characterization, we also have our molecular analysis. So this has not been used in a forensic setting, to my knowledge. However, a lot of research in other areas have looked at how different techniques such as spectroscopy and spectrometry can be used to identify and characterize the different macromolecular compositions of pollen, namely lipids and proteins that are present in these grains. And so a couple of studies listed here are examples of how people have used these variations in the chemical composition as tools for pollen analysis. However, they have not yet been used as a tool for species identification.  

And so, next slide.  

And so that is where we wanted to kind of bring this in. Is it possible then to combine this genetic and molecular characterization along with the physical or morphological characterization using microscopy to generate these species-specific physicochemical profiles that we can use as tools for taxonomic identification? And so for today's presentation, I'm going to focus more on the chemical profiling using LCMS, just because it's a little bit more of the novel aspect of my work.  

Next slide.

And so for this project that we work with, we did start with a 15 species initial sample set. This consisted of 10 trees and then five grasses/weeds. And so these are all chosen for a couple reasons. Most of them are native to Central Florida, which is where this project was worked, or where this project has been conducted.

And so for comparative samples, we wanted to be able to access these species in the wild. Next. We then split up the sample set to include a wide variety of taxonomically related species. So six of them are all from the Quercus genus or oaks. Next. Two are pines. Next. You get to hear me say next a lot. Next. Two are ragweeds, and then the five remaining ones...

Next. ...are all in a taxonomically different. And so these were commercially sourced from a company that provides allergen extracts, as well as the source materials like the pollen for biomedical research. And so luckily, they have great quality control measures to make sure that these pollen have not been previously chemically treated. Next.

A general workflow for this project that we had was, can we first extract the chemical features, components like these lipids and proteins non-destructively? Once we have those, can we analyze them using LCMS with a very standard LCMS workflow, following an untargeted analyte detection, feature alignment normalization, and then perform a series of chemometric analyses in order to basically visualize and analyze this analysis? And then the final step would be, are these chemical profiles that we've obtained different enough to be used for species classification, which we performed using LDA? Next.

So I’ll briefly go over the extraction. This is not something that we sort of developed. We followed a Folch method, which if anyone's familiar with biochemical extractions, this has been used also since around the 50s. It's a chloroform methanol 2 to 1 mixture, which is commonly used to isolate lipophilic components from biological cells. And so we found that we could use this as well to extract these chemical components from pollen without having to rupture the grain wall.

And so we just did a series of vortexing and then centrifuging to pellet the pollen. And then finally, we can combine the extracts with water, which creates a biphasic system, which is really excellent for our work because that separates our lipophilic components, such as our lipids and proteins, into our organic phase. Meanwhile, the aqueous phase contains most of our polar charged species, such as the DNA. Once we were able to separate these out, we can then perform those orthogonal analyses.  

Next slide.

And so just as like a little caveat, I keep saying it's non-destructive. So just like, trust me. So these are some microscopic images that we've taken. So like I showed you in the previous slide, we did retain the grains following extraction. So this is just a comparison of two of the species, bald cypress and reed canary grass, pre and post extraction. And so we saw very minimal occurrences of grain rupture during the vortex. The only noticeable observations that we saw was slight variations in size due to grain compression and retaining of that solution. But after drying them out, they returned to their normal size. Next slide. And so now getting into the analysis, so using that 15 species sample set, I performed 10 separate extractions for each species.

So that gave us a total of 150 extractions. And so using online platforms, XCMS, we were able to identify over 34,000 features from the LC profiles. And we wanted to see if we could use those as predictors for LDA and other machine learning algorithms. And so this heat map here is a common way that people that work with LCMS or metabolite feature, I guess, workflows like to visualize their data.

So each row here is a unique feature across the entire sample set. And we z-score normalized them so that the intensities of those features did not overwhelm if they were too high. So for example, here, the lower... Sorry, if a feature was presented higher in a class versus others, it would be red, so it would be standard deviations higher than the mean. Whereas if we have a blue feature, that would be standard deviations lower than the mean. And so the more abundant feature profiles are going to be higher in red, such as we see with the giant and short ragweed there kind of in the middle left. And then so using these features, we performed a very or a series of different chemometric analysis.

Next. The first one being an unsupervised sort of clustering model using HCA or hierarchical cluster analysis. And so this is really exciting because this is unsupervised. So we did not tell this model that the classes or the individual extractions belong to a specific class. It tries to basically find similarities or correlation profiles between the different classes. And so here, we were able to observe distinct taxonomic similarities and differences based on their chemical profiles.

For example, you can see that the pink bracket there showing the yellow pine and the loblolly pine have a very small correlation distance indicating they are extremely similar, which makes sense because they are highly related taxonomic species. Same thing seen for the giant ragweed and the short ragweed shown in yellow, as well as the six oaks shown in green.

And so here we do see that there are chemical features being retained within a genus. However, they are still varied enough to allow for pollen class or species separation. Next. And so the next step then was to see if we could classify unknown or sort of like mock samples using these feature lists. And so we did that using LDA. This is a supervised learning, or sorry, supervised machine learning algorithm, which basically you tell it what the pollen classes are, and it tries to best classify your test samples based on those classes. Next. And so using a stratified training model with a 30% holdout and 100%, or sorry, with a 30% holdout validation and a 70% training, we achieved 100% classification.

So that's like the red dotted line at the top. So every time we perform this model using the full 34,000 feature list, we had 100% classification accuracy, which sort of makes sense because if there are features that are specific to a class, you would expect that the sample to accurately classify to it. However, 34,000 features is a lot to work with.

And obviously, if we were to add more species, that would probably expand. And so we wanted to see if we could achieve a similar classification accuracy with a greater, greatly reduced data set. And so a common combination with LDA is the application of PCA as a data reduction method, which is also unsupervised data reduction, which uses a combination of linear, sorry...

Yeah, you use linear combinations of your variables to sort of maximize global variance. So think of like, it takes your entire sample set and tries to determine the most variance between classes and only use the features that correspond or give rise to that variance. And so we did that across the entire PCA, the principal component space, and we saw that we could obtain a similar accuracy at just about 18 PCs with a greater than 90% accuracy at just 13 PCs. So greatly reduced data sets are still able to achieve similar classification accuracy. Next slide. And so a common way of also visualizing that sort of analysis is if anyone's familiar with PCA, you might see a PCA score plot.  

However, that tends to, like I mentioned previously, show global variance. And so our entire 15 species set was highly dominated by a few species because they were so heavily abundant in features that when we did PCA score plots, the clustering was not the best for visualization. So we decided to apply T-distributed stochastic neighbor embedding, which is a non-linear embedded algorithm that actually utilizes the PCA reduced data sets, but it prioritizes local relationships or neighborhoods to better visualize the clustering that is occurring. And so this plot sort of just shows that there is tight clustering within all of those 10 individual extractions, suggesting that the chemo variability is inherent to these following classes and is not something that's being added in from like the analytical bias. And so the results of the classification can be better visualized, or I guess the sort of the behind the scenes of that clustering. And then I think that's about everything I have for the main one.  

I'm going to briefly go through the next few slides or the next few slides because these were added on just if we had time. But it's sort of how we were transitioning from this original proof of concept into lower grain count.

So these were all done with about 100 milligrams of pollen, which doesn't sound like a lot, but is, and so...I kind of will just skip through them if you don't, if you don't mind, AJ, you can just go next. But essentially, we are looking at single grains using the chemical profiles. Still with mass spec, so this time using TIMS.

Next slide.  

And we looked at... You can go to the next slide. Yeah. We looked at the chemical profiles obtained using TIMS from just a single individual grain, which as you can imagine, very low concentrations. And while we are obtaining far fewer components than with the bulk, we are still obtaining profiles that do show species specificity, which is really exciting.  

Next slide.

Oh, and next, another profile. And then I'm not going to get too much into this because I think it's a whole another can of worms. But if anyone is interested in asking about how the DNA or coding side of it goes, feel free to reach out later. But essentially, we were also able to perform DNA quantitation from our single grain extraction.

So we were able to obtain viable DNA that we could amplify using PCR and sequence, which is really exciting. So if you want to go to the next slide as well. We were able to sequence these, so we sent these out for sequencing and cross-referenced them with the GenBank database. And that table kind of just briefly shows some of the results. So we were getting correct matches to both the genus and species level, again, from the single grain, which is really exciting. Next slide. So just finishing up, I kind of hope that I sort of demonstrated things that the chemical signatures do differ at these species level, and we were able to attain sort of taxonomic separation using them. Machine learning can definitely help to process these data sets, which do tend to be very large, especially with LC results and then we'll skip over this, the single grain, but just TLDR, just trust that the single grain was exciting work.  

Next slide.

Future work for this, we do want to evaluate the same models that we use for the bulk with the single grain to see if the features obtained from those can provide it. And then we are also working right now actually with comparing them with real samples. So like I mentioned previously, this is all done with standard pollen samples, but we are collecting environmental samples and we have gotten some pretty exciting results from that as well.

Next slide.  

So I know there was a lot to throw at everyone, but thank you all for listening. I want to thank the NIJ again for funding this work, as well as my research group at NCFS at the University of Central Florida. And yeah, thank you all for listening. If you have any questions, I guess I'll be happy to take them now.  

Tracey Johnson: Thank you very much, Jack. So we'll now move into the question and answer period. So if you have any questions, please submit them in the Q&A box. So starting off strong, Jack, can you talk about why LCMS was chosen as opposed to something like a GCMS? You know, because it seems like the LCMS was more useful for the species. But wouldn't in forensics be important to understand the geographical location? So using something like a GCMS, would that be more appropriate?  

Jack Trimble: Yeah, so originally, LC was chosen for a couple of things. One, a lot of the chemical components based on literature from the pollen coat, which surrounds the pollen, tend to not ionize well. With GC, there are larger molecular like phospholipids, proteins, things like that. So a lot of the studies I actually did do GCMS analysis of pollen had to add a derivatization step. And so we, it also kind of works with LC because we wanted to perform a one pot extraction. And so the DNA analysis with the liquid extraction and the LC, it just, I guess it was more efficient than separating out with the GC extracts would have had to have additional steps for analysis, if that sort of answered your question.  

Tracey Johnson: It does, thank you very much. All right, so moving in to the second question, do you anticipate cross-pollination being problematic, particularly when analyzing single pollen grains for DNA?  

Jack Trimble: um, yes, so if we're talking about cross-pollination with like the occurrence of...different pollen, or sorry, pollen from different species being left on a sample, or if we're referring to genetic hybrids, they both would be.

So I know we've worked with, or we've seen issues where because of pollinators such as bees bringing pollen from different species onto one plant, that could be an issue because you think that you're looking at one species when you have the pollen of another, but also a hybrid species. Their genetic sequences might not be different enough. And so I will say I'm not a... I focus more on the LC, like the molecular side than the DNA barcoding. That is work that's kind of been incorporated into my project based on research from previous years. It has its own limitations and where the sequences are sometimes not varied enough at the species level still. So you are getting, you can get sequences, but those sequences do not differ enough at the species level.  

Tracey Johnson: Sounds like you got some challenges in front of you, Jack, with continuing this project, but always the case in forensics. All right, one more question. And so that question is, you described the need for reference libraries for pollen identification. Can you speak to the development of, access to, or necessary updates to pollen reference libraries?  

Jack Trimble: So for image libraries, I probably won't speak too much. I'm not too familiar with the available ones. So there are, I would say the major lab that performs this work, which is in Houston, they might have their own reference library. There are some available image libraries online that are not as extensive as we would like. However, with the DNA, that one, the GenBank database that I use, which is freely available, that one is massive, and that one does contain a wide variety of DNA sequences. There are no libraries, I guess, to my knowledge, containing like chemical information. And the few studies that have looked at analyzing the chemical components of pollen tend to all work with the same few species as like a model. And so there is kind of a large gap between having an extensive and robust database that encompasses several species, which is why we worked with our samples that just kind of like to begin.  

Tracey Johnson: All right. Thank you very much, Jack, for your presentation and the information. We'll now hear from Akshita Verma.  

Akshita Verma: Hi, good afternoon, everyone. Thank you for joining and thank you, NIJ, for giving this opportunity. And I'm very excited to be presenting a part of my PhD project on rapid LC-HRMS non-targeted screening method for emerging nitazenes in human urine samples. Next. So what are nitazenes? We know this is a very hot topic in the illicit drug market right now.

And these are very unusually dynamic class of synthetic opioids, which have re-emerged now since 2019. But they were originally created in 1950 as a potential pain relief medication. But because they have a high risk of overdose and addiction, they were never approved for medical use. But we are now seeing them again in the illicit unregulated drug market. And they are known to have a significantly very high potency, even more than morphine and some of the variants even higher than fentanyl.

So that's a very major toxicological concern that we don't know what the safe dose is and they are normally created in illegal labs. We cannot even track that toxicological aspect of them. And because they are very potent, they are normally abused in a counterfeit pills or adulteration or mixed in heroin, cocaine or fentanyl powder as well.

So apart from being a very potent significance, they just like any other NPS, they are very easily modified, modifiable. So a very small alteration in the chemical structure, we can create a lot of different analogs and isomers, which is becoming a problem for analytical chemists to keep a track on time. And that's another surveillance which we are required to track on them. Apart from nitazenes, there is another category of synthetic opioids, which is also gaining name called as orphines, which are having a similar effect on the whole toxicological pattern. And we try to implement both of these opioid families in our study. Next.

So the gap majorly is that because new analogs are emerging and the market shift is changing at a very fast pace, we normally are not including them in our screening methods in our target list, which is a fallback when we have to identify them. It's also a problem because they are changing so fast, we don't have available reference standards on time, which is also another problem in analyzing them on time. And that's a problem for the analytical methods in general that when the new analogs are coming, we normally are not able to track potentially the ones which we are not looking for. Because of which, the main concern is developing an acquisition method which can find analogs which we are not even looking for. So there comes a role of non-targeted methods and the importance of developing them and keeping a track of this set of NPS compounds. Next.

So the core challenge is because they are analogs, they are very closely related. They are modifiable with a very simple alteration. Normally they are isomeric in nature. So a lot of compounds are sharing the same molecular formula and thus the same mass. And they also give the identical fragmentation behavior through our analytical techniques using high-res mass spec. So the problem is how to differentiate between these isomers. So normally the chromatography becomes our last resort to help differentiating between these compounds. Next.  

So the main four study aims considering all of these issues with this drug family was to confidently differentiate isomeric compounds. In our, this specific study, we actually are working on 20 different nitazene compounds and all of them are isomeric with each other. And our target is to differentiate between as much isomer compounds as we can. Then the next aim was to actually use a very compressed runtime so that we are on track of the backlogs of any new compound coming in the market. So we are developing a method which is much faster and much easier to imply. The thought was because we are using a urine sample, we wanted to implement a very minimal sample prep type of the methodology. So we used a dilute-and-shoot by simply diluting using a water without any cleanup methods, without any evaporation step. Again, to use or adopt a very minimal preparation step in our analysis.  

And lastly, like I mentioned, that we wanted to use the non-targeted approach so that if there are new analogs coming in the future, we can still use the same method we have developed now and try to find the unknown compounds in the future. Next.

So, the methodology was simple. We used the dilute-and-shoot for the human samples, and we used Agilent LCQTOF method for this, where we used a data-dependent acquisition method, and we implied an inclusion list of our compounds. So, we worked with 20 nitazenes and three different orphines. So, we had like 23 analytes, and we had one internal standard to match with them. And we created and validated the method using the forensic tox 036 standard. And these are the exact parameters which we implemented in this study. And we were able to create an LC gradient method of 4 1/2 minute, the mean runtime, and one minute re-equilibration, so a total of 5 1/2. So our target was under six or seven minutes, which we were able to achieve for all these 23 analytes. And we used a data dependent Q-TOF method and we added the list and we used three different collision energies to actually implement the fragmentation and understand how all of these compounds are behaving for their fragmented product ions as well. Next.

So few of the results here, I'm showing the chromatogram of all these analytes. And like I mentioned, all of these are isomeric in nature, like all the nitazenes are. So there are 20 nitazenes which are isomeric. And we were able to differentiate 19 of the 20 nitazenes. The only pair which we were not able to actually differentiate was butonitazene and isobutonitazene, which is heat number 23 and 24 if you see on that chromatographic image. But apart from these two analogs, we were able to actually separate all of these through the gradient method which we applied, which was just 5 1/2 minute of the runtime. Next.

So we actually tried to keep working on maybe trying different gradient methods to differentiate that butonitazene triad. So we actually have 3 isomers in this pair, the sec butoni-azene, buto, and isobutonitazene. And we tried a bunch of gradient methods to see if we are able to differentiate. So when I injected them, separately using a normal isocratic method of 12 minutes. I was not seeing much of the difference, but I tried a bunch of different variations.

If you see the second image, between the 2nd and the third peak, we were still able to see some small separation, but this was only because I was injecting them separately. But the main aim of our study was to actually use all these 23 analytes in one single test mix and then inject in our urine sample. So our purpose was to inject them together and not individually while we are creating the method. So that fell short and we had to stop the optimization and just go ahead with the method which we developed previously. Next.

So we have like 9 isomeric groups and we could differentiate them on the basis of either the fragmentation behavior or majorly the retention time. So out of nine, the five groups, we could not differentiate them on their fragmentation MS2 patterns because they gave the same diagnostic fragment ions, but we were able to still thankfully separate them on the basis of retention time. But apart from that, butonitazene and isobutonitazene pair, we were able to separate that. Next.

So while we were developing this LC method and when I analyzed all the drugs, we saw a specific, very unique trend of the retention order of the different families of this compound. So we saw majorly two trends. The first was that we have a lot of iso-- and proto versions of these isomeric compounds. So I saw that all iso-, which is the branched version of the compound, they were eluting first then the linear proto- or propoxy- counterpart. That was consistent with all the isomeric pairs, which is a very unique and a consistent result, which I saw.

Then the second trend, which I also noticed that there was the elution order of a pyrrolidino and the parent and the piperidinyl analog behavior where they were occurring in this specific order. So these two trends were actually really highlight worthy so that we can keep considering them in our future when we add more nitazenes in our study method, or even if we are looking for any unknown compound in the future, because these were reproducible, they were occurring again and again. We could rely on this specific trend that theiso- analog is always coming first or eluting first in the gradient method which we apply. So we want to keep looking into this trend when we are improving or widening our search panel in future. Next.

Now on the fragmentation pattern, we also saw trends that we saw mean diagnostic patterns among majority of the compounds, which was 100 or 110 or 72 or 44. So these 72 and 44 were the mean diagnostic among common among all the nitazenes. And for pyrrolidino analogs, we saw 98 and 56 as the main product ions. So we could have specific diagnostic fragment ions for each core of the nitazene. And orphines gave their own fragments, which all of these matched with the other literature, which is out there on them, studying on the fragmentation. So that matched. And these are a good quantifier because we were able to actually match them with other literature.

Next.

Now, coming on to the actual validation method and its results, we saw a very good mass accuracy of all the analytes, which was below 2 ppm. It's all good range of LODs with 0.15 to 2 nanogram for majority of the analytes. And the linearity was good for majority of the compounds. It was even above 0.998 value, which gave us the confidence that the calibration plots are good. And then the bias and precision study, they behaved very well. Majority of them were under the threshold value, which is given 20 according to the forensic tox standards, but majority of them are below 10, which was a very good finding. And I think I forgot to mention, but the calibration curve, we used eight different points of the calibration curve, ranging from 0.5 to 100 nanogram per mil. And for quality control for precision and matrix effects, we use three different levels of quality control at 1.5, 40, and 80 nanogram per mil. So that was the result for these parameters. And we also went ahead with checking any interferences or carryover, and we were positive in all our, those portions of the studies as well. Next.

So for matrix effects, the actual validation threshold is 25, but we saw a lot of compounds above 25 because probably we were using a very simple dilute and shoot method. But other literature have also reported this high increased values. So we were not concerned about that. But at low concentration level of 1.5, our results were good, which is also our primary concerning concentration because these compounds are normally very toxic.

So they are present at very low concentration levels. So our focus was on them behaving good at lower quality control sample. And for matrix effect, also, we saw a very interesting trend that majority of the nitazenes, they were showing ionization enhancement, but orphines, which we had in our study, we had orphine, we had its active metabolite, nor-orphine, and we also had cyclorphine, which is also gaining a name right now in the market. These 3 orphines we saw being on the suppression end of the ionization effect, and this was consistent with our repeated analysis as well. So this is also a trend which we are keeping into account to make a note if we are featuring or improving this method for other analytes to include in this panel as well.

Next.

So the main discussion out of this project was that we only had one butyl pair which was coeluting and apart from that, everything was resolvable within 5 1/2 minutes of the runtime. Although we used a single internal standard against all 23, but we were able to successfully identify all the compounds very accurately and they were below a good threshold values according to the forensic tox standard. And our next steps actually is to implement this validated method on the real authentic case specimens and see how the method is performing and behaving. Next.

So in conclusion, we were able to develop a single injection full panel of 23 analytes within a five and a half minute run. We met all the toxicological standard covering linearity, sensitivity, and all the performance metrics. And because we saw specific trends, we know this is a transferable parameter which can be taken into account for another studies by other laboratories or even us if we are modifying anything in the future. And for the fragmentation behavior, we also saw that there are some structural rules which are getting followed with some diagnostic product ions within the same nitazene family as well.

Next. So yeah, these were a few of the references which I accounted for this study. Next. And lastly, I would want to acknowledge my advisor, the facility at FIU, and of course, NIJ for helping fund this project and for us to be able to actually implement and work on this set of NPS compounds. And if you have any questions or if you want to reach out to me in future, this is my LinkedIn profile, and you can, you can contact me for this. And I'm open to any questions if you have.  

Tracey Johnson: Thank you very much, Akshita. We probably have time for one question. So are there plans to look at biological matrices other than urine?  

Akshita Verma: So as of now, we don't have, but if we have time to use the study, we might want to look into blood samples, but urine was our main target for the current method.  

Tracey Johnson: All right, thank you very much. So that's a wrap on the Graduate Research Fellow session and a wrap on the NIJ Forensic Science and Research Development Webinar Series. Thank you for dedicating your afternoon to today's webinar. We hope the information provided was helpful and informative. For any follow-up questions, please contact us at [email protected]. A recording of the webinar will be posted to the NIJ website in the coming days and we look forward to your participation in future events. Bye, everybody.