*Article* **Providing a Forensic Expert Opinion on the "Degree of Force": Evidentiary Considerations**

**Hans H. de Boer 1,\* ,† , Charles E. H. Berger 2,3,† and Soren Blau 1,†**


**Simple Summary:** When giving evidence in court, forensic pathologists and anthropologists are often asked for their opinion on the amount, or degree of force required to cause a specific injury. Such 'degree of force' questions are considered difficult, if not impossible to answer due to many theoretical and practical issues. This paper explores these issues and provides a possible solution. First, the logical underpinnings of the question on the 'degree of force' are explored. Then the experimental research on 'degree of force' is reviewed and the limitations with applying this research to everyday forensic casework are discussed. In the second part of the paper, it is argued that these limitations do not, however, mean that a forensic pathologist or anthropologist cannot add anything of value to the discussion. The application of Bayes' theorem helps to circumvent many of the problems. The final part of the paper is dedicated to a detailed discussion of how it can be applied to the issue of 'degree of force'.

**Abstract:** Forensic pathologists and anthropologists are often asked in court for an opinion about the degree of force required to cause a specific injury. This paper examines and discusses the concept of 'degree of force' and why it is considered a pertinent issue in legal proceedings. This discussion identifies the implicit assumptions that often underpin questions about the 'degree of force'. The current knowledge base for opinions on the degree of force is then provided by means of a literature review. A critical appraisal of this literature shows that much of the results from experimental research is of limited value in routine casework. An alternative approach to addressing the issue is provided through a discussion of the application of Bayes' theorem, also called the likelihood ratio framework. It is argued that the use of this framework makes it possible for an expert to provide relevant and specific evidence, whilst maintaining the boundaries of their field of expertise.

**Keywords:** degree of force; skeletal trauma; forensic pathology; forensic anthropology; review; evidence; opinion; likelihood ratio; Bayes' theorem

## **1. Introduction**

*"* . . . *force alone is woefully inadequate and often (particularly in a legal environment) misleading in describing an impact".* [1], p. 283

The concept that skeletal trauma occurs when a force exceeds the strength or maximum threshold of bone elasticity is well established [2,3]. In forensic pathology and anthropology, descriptions of the application of a force to the body are typically divided into three groups of causation: blunt force, sharp force and high or low energy ballistic force. While the potential results of these forces on the human body have been well documented in the biomechanical [3–5] and forensic medical literature [1,6,7], correlating the amount

**Citation:** de Boer, H.H.; Berger, C.E.H.; Blau, S. Providing a Forensic Expert Opinion on the "Degree of Force": Evidentiary Considerations. *Biology* **2021**, *10*, 1336. https:// doi.org/10.3390/biology10121336

Academic Editors: Ann H. Ross and Eugénia Cunha

Received: 1 November 2021 Accepted: 14 December 2021 Published: 16 December 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

of force applied to the body to a specific injury or fracture outcome has proven more difficult. Nonetheless, when providing expert opinion in court about skeletal injury, forensic pathologists and anthropologists are often asked their opinion about the cause and specifically, the amount or the 'degree of force' that was required.

The relationship between injury morphology and applied force is complex, and opinions on the 'degree of force' are therefore fraught with difficulties. This paper provides an overview of the concept of 'degree of force' in forensic pathology and anthropology and in doing so, provides an aid for practitioners when giving evidence on this issue. While the paper focuses on skeletal injuries, much of the discussion also applies to the same issue when interpreting soft tissue injuries.

## **2. Why Is the 'Degree of Force' Considered to Be Important?**

In order to understand why the issue of 'degree of force' appears to be so pertinent in criminal cases it is necessary to reflect briefly on the purpose of a criminal court proceeding, and the role of the expert witness. Although differences between jurisdictions exist, a criminal court proceeding typically aims to determine whether enough evidence is available to convict the defendant for the alleged crime. This process is traditionally dialectic, with prosecution and defence both trying to convince the trier-of-fact of their respective positions, usually by presenting evidence. The opinion of the expert witness, like other evidence, can assist the trier-of-fact in weighing the competing positions of prosecution and defence.

Within this context, opinions on the 'degree of force' have been considered useful to help the trier-of-fact to reconstruct the events that led up to and resulted in death. In other words, such opinions are intended to help the trier-of-fact choose between various scenarios. The high frequency with which forensic pathologists [8–11]; forensic anthropologists, and forensic physicists [12] are confronted with the question suggests that such an opinion is considered particularly helpful by the court. This perceived value of opinions on the 'degree of force' appears to be based on three assumptions, namely that proportional relationships exist between:


If all three assumptions are valid, the conjecture is that knowing the 'degree of force' may help to differentiate between intentional or accidental injuries, and therefore, help to conclude if a crime was actually committed. Furthermore, since the seriousness of a crime ordinarily influences sentencing decisions [13], intent is an important aspect of culpability in many jurisdictions. An expert opinion that can inform on intent can therefore have an impact on sentencing [14].

#### **3. Forensic Expert Responses to the Question of 'Degree of Force'**

Questions relating to the 'degree of force' may be asked in various forms. Typically, however, the expert is asked the question in a simplistic form: "what degree of force is required to cause this skeletal injury?" The expert is subsequently expected to provide an estimate of that amount of force, based on the combination of observations, knowledge, and experience.

Anecdotal information, largely obtained from discussions amongst forensic experts, indicates there is variation in their responses. The general consensus is that a specific answer (i.e., including a number expressing the amount of force) cannot be provided. As a result, experts may provide a response along the lines of "I am unable to comment" or "I can comment, but without a degree of precision". Other types of responses include "the force was sufficient to result in a skeletal injury"; or "clearly there has been enough force to fracture a bone". Since such comments only reiterate the facts that are already known, it may be argued that these opinions are as uninformative as "no comment".

To simplify the issue, some experts choose to use a qualitative three- or four-point scale to describe the amount of applied force. This approach has also been described in the literature, with verbal descriptions such as "mild", "moderate" and "severe" force used by Nolan et al. [15]; and "mild", "moderate", "considerable" and "severe" force by Gilchrist et al. [16] and Sharkey et al. [17]. A definition of what these specific categories mean, or how the expert should choose between them, however, remains largely undiscussed. In a study pertaining to stab injuries, Gilchrist and colleagues [16] stated that a mild level of force would "typically" be associated with penetration of skin and soft tissue, moderate force with injuries that penetrate cartilage or rib bone, and severe force with injuries that penetrate dense bone and cause visible damage to the knife's blade. But these definitions are not generally accepted, and the limitations of using these vague and relative terms have been previously noted [8,15].

Overall, there is no consensus on how an expert should answer a question on the 'degree of force'. This lack of consensus has served as a justification for research which has sought to quantify the degree of force in various types of injuries.

#### **4. Evidence for the Relationship between Degree of Force and Injury Outcome**

A range of experimental studies have been undertaken to investigate and correlate the relationship between degree of force and injury outcome. This research ordinarily focuses on the method of injury, rather than on the type of tissue injured. For instance, research has included the investigation of degree of force and sharp force trauma involving knives [8,16,18], as well as stabbing involving other implements such as screwdrivers [19]. Such research has used pork skin [15,18] as well as synthetic materials such as foam [20,21], silicone rubbers [22,23] and modelling clay [24] as substitutes for human skin. Research investigating the relationship between degree of force and blunt force trauma has also been undertaken. This research has mostly focused on understanding the force required to cause head injuries, including brain injuries [25] and skull fractures [26]. As experimental models, researchers have used human skulls [27] as well as those of pigs [17,26] and monkeys [28], in addition to computer simulations [29].

Despite the use of these various experimental models, different anatomical parts of the body, and different types and amounts of force, the results of these experimental studies are difficult to apply to forensic casework. This shortcoming becomes more apparent when reconsidering the previously mentioned three assumptions that underpin the alleged validity of the 'degree of force' question.

Experimental research has predominantly focused on the third assumption: the relationship between the force applied to the body, and the severity of injury. Consequently, such research only addresses one part of the issue at hand. Further, the highly controlled settings typical of experiments do not (and cannot) take into account the many intrinsic and extrinsic variables that influence the relationship between applied force and injury outcome. Intrinsic variables include the sex and age of the deceased, and the specific anatomical region impacted (e.g., head vs. chest). The anatomical region, and therefore the skeletal element, is also important to consider, as different bones differ in their density, flexibility, and design (e.g., the area of impact may be buttressed by other anatomical structures) [27]. In addition to the health status which affects bone plasticity [30], individual variation in bone morphology must also be considered (e.g., skull thickness [31–33]). Overall, while the results of experimental research may be interesting as a means of demonstrating the biomechanical properties of human (and non-human) tissue, they are not directly transferable to forensic casework.

Published research focused much less on the first and second assumptions that underpin the 'degree of force' question, that is, the relationship between 'intent' and 'force used', and between 'force used' and 'force transferred'. The difference between 'force used' and 'force transferred' is an often-overlooked issue in experimental settings but is, nonetheless, relevant in forensic casework. In many instances the relationship between these two forces is not proportional. Consider, for example, a situation in which a perpetrator exerts what

may be described as a 'relatively minimal' force (e.g., a gentle, yet intentional push) which nonetheless results in what is described as 'severe trauma' (e.g., multiple comminuted fractures due to a fall from height). In other settings the relationship between 'force used' and 'force transferred' may be proportional, but there is no way of knowing the extent to which one is influenced by the other. For example, in the case where a perpetrator uses an implement that modifies the force that is used (e.g., a baseball bat, a hammer, or a knife). Extrinsic variables such as the effect of the size, shape, elasticity, and mass of the impacting implement [17,27] are important in this regard. The directionality of the impact is also of interest [12] as well as its speed (because bone is viscoelastic, that is, responds differently depending on the speed at which a load is applied). The direction-ality and speed of impact relate directly to the relative position of perpetrator and victim and the dynamics of the event. When all these variables are considered, it becomes appar-ent that the same amount of force used by a perpetrator can, depending on the circumstances, result in different amounts of force being transferred (applied) to the body of the victim.

The assumed relationship between the intent of the perpetrator and the force that is used is also rarely considered in empirical research. Although it is intuitively true that the intention to inflict grievous bodily harm results in a large amount of force being used, it is not necessarily so that unintentional behaviour results in less force. Consider, for example, scenarios of self-defence, in which forcefully fending off an attack can cause serious harm to the attacker. Moreover, one study showed that when volunteers were asked to use 'mild', 'moderate' or 'severe' force, the resultant amounts of (stabbing) force were too similar to reliably infer the 'intent' of the volunteer [15]. The sex and age of the perpetrator have been noted as important variables to consider in this [12,15]. However, these are only two of a multitude of interacting variables that may be of relevance.

Overall, while the findings from experimental research can perhaps support claims about the potential effects of force on the human body, the data seem of limited use to provide informative opinions on the 'degree of force'. It may be argued that, in fact, the results from experimental research reinforce the idea that the 'degree of force' is an issue associated with great complexity and uncertainty, while its relevance is very limited.

#### **5. Taking a Different Approach: Applying Bayes' Theorem**

Given the complexity and uncertainty associated with providing an opinion on the 'degree of force', an alternative approach is to use probability, described as "a tool to handle uncertainty" [34]. The difficulties surrounding the issue can perhaps be addressed by applying the laws of logic and probability, using Bayes' theorem. This theorem describes the logical underpinnings of the process by which probabilities are updated based on observations [35]. Many textbooks and journal articles provide introductions to Bayes' theorem, and explain why its use is the logically correct way to interpret and present forensic evidence [34,36–38]. Bayes' theorem has been applied in a range of forensic disciplines including pathology [39,40], anthropology [41,42], entomology [43], biometrics [44], and biomechanics [45], and to address different questions such as time since death [46], manner of death [47], and identification [48,49], including disaster victim identification [50–52] and missing persons investigations [53]. To date, however, Bayes' theorem has not yet been applied to address the issues associated with opinions on the 'degree of force'.

Bayes' theorem, which in forensic science is also referred to as 'the likelihood ratio framework', is best explained by the equation in odds form:

$$\frac{P(H1)}{P(H2)} \times \frac{P(E|H1)}{P(E|H2)} = \frac{P(H1|E)}{P(H2|E)}$$

With:

P(Hx) = prior probability of proposition x

P(*E*|Hx) = probability of the evidence *E*, given proposition x

P(Hx|*E*) = posterior probability of proposition x, i.e., given the evidence *E*

This can also be written as:

prior odds × likelihood ratio = posterior odds

It should, however, be kept in mind that this equation only shows the logical relationship between the probabilities of observations and propositions. The theorem therefore remains valid in the absence of numerical data.

## *5.1. Prior Odds*

The prior odds are given by the ratio of the probability of proposition H1 and that of H2, without considering the expert's observations, that is, the evidence (*E*). Because the prior odds are based on all information outside the expert's evidence, assessing the prior odds would take the expert outside their area of expertise.

#### *5.2. Likelihood Ratio*

To provide an opinion while staying within their area of expertise, the expert needs to focus on the likelihood ratio (LR) only. The LR is the ratio of two probabilities: the probability of their observations (*E*) given one proposition is true, and the probability of the same observations given an alternative (mutually exclusive) proposition is true. Assessing these two probabilities relies directly on the experience and expertise of the expert. This process does not necessarily imply using statistics and calculations: the same logic applies with or without the use of numerical data.

#### *5.3. The Posterior Odds*

The posterior odds take all the evidence into account: they equal the prior odds multiplied by the LR. Since the posterior odds require the prior odds, the posterior odds are also outside the forensic pathologist or anthropologist's area of expertise.

## **6. A Hypothetical Case Example**

The utility of the application of this framework to the issue of 'degree of force' can be illustrated by the following hypothetical case. The partially skeletonized remains of an adult male were located at the bottom of a mine shaft. The individual's skull was fragmented. The remains were examined by a forensic pathologist and a forensic anthropologist. Reconstruction of the skull fragments revealed two concentric, patterned impact fractures: one in the left fronto-temporal region, and the other in the left temporoparietal region. There was also a linear defect on the right posterior aspect of the occipital bone. In their joint report, the forensic pathologist and anthropologist concluded that these observations indicated multiple impacts, and that the cranial trauma was a reasonable cause of death. Eventually, a person was arrested in relation to the matter and the case went to trial. In court, the experts were asked their opinion on the 'degree of force' required to produce this fragmentation and patterned injury.

While the experts can try to answer this question, as previously discussed, many limitations preclude the provision of a robust opinion. Using vague terms such as 'mild', 'moderate', 'severe' and 'extreme' to describe force does not overcome these limitations. These restrictions do not, however, mean the expert cannot add anything of value to the discussion.

#### **7. The Need for Propositions**

When applying Bayes' theorem, the expert's opinion is used as evidence to help give weight to one of two propositions, most often the positions of the prosecution and defence. For instance, in the hypothetical case outlined above, the prosecution may allege that the decedent was beaten to death with a shovel and then dumped in the mine shaft. In contrast, the defence may propose that the skull fractures were the result of a fall following a verbal altercation between the decedent and the defendant. As discussed previously, information about the 'degree of force' is just an intermediate step in addressing the larger issue: which

of the two propositions is correct. If the court is focused on one specific injury and only enquires about the 'degree of force' required, these propositions are not made explicit to the expert. Consequently, the full meaning of the pathological/anthropological findings cannot be borne out.

Only when provided with propositions, can the expert provide the most relevant evidence. For instance, in the hypothetical case, the expert could clarify that the propositions provided by prosecution and defence both imply that substantial force was applied to the skull, and therefore, an opinion on the 'degree of force' is of no use to distinguish between the two propositions. Further, by focusing on 'degree of force', other observations made by the expert remain undisclosed. In the hypothetical case example such information includes the findings that the victim had a minimum number of three impacts, both sides of the skull were impacted, and that there were two patterned impression fractures and one linear fracture. These details are all potentially useful to the court proceedings, especially when the expert is provided with some case circumstances.

#### **8. How Does the Expert Assess Evidential Strength (An LR)?**

Instead of requesting an opinion on the 'degree of force', a more appropriate question for the expert may be: "to what extent do your observations support scenario A (that the decedent was assaulted with a shovel and dumped in the mineshaft) vs. scenario B (that the decedent fell into the mineshaft)?" When confronted with these propositions, the expert can apply Bayes' theorem, and therefore provide the evidential strength of their observations (an LR).

But how are experts supposed to assess an LR? Where do they get 'the numbers' from? It is important to remember that the use of probability does not imply statistics and calculations [34], and that a lack of data does not preclude the application of logic. LRs can be used qualitatively. However, the LR framework cannot mitigate gaps in scientific knowledge. If the expert thinks there is insufficient scientific knowledge to provide an opinion, it is their professional obligation to say so. In that situation the expert's opinion represents an LR of 1, which simply means that in the expert's opinion, their observations do not assist in distinguishing between the two propositions.

In the hypothetical case the observations of the two concentric, left-sided patterned impact fractures in the fronto-temporal and temporo-parietal regions, and the right-sided linear defect in the occipital region are the relevant evidence (*E*). The first question is, therefore, to what extent does the expert expect (or is surprised by) these observations if scenario A (H1) is true? How probable is the presence of a linear fracture when hit with a shovel? And would such an impact result in multiple concentric, patterned impact fractures? Moving to scenario B (H2), what is the probability of the observations if the deceased just fell in the shaft without being beaten? Answers to these questions rely on the expert's observations. Ideally, however, they would also be informed by some (preferably undisputed) information on the case circumstances. In this case this information would include details about the structure (walls and bottom), height, and width of the mine shaft. To obtain an LR the expert finally needs to relate the expectation for the observations under both propositions to one another, because it is their ratio that determines the evidential strength. It is important to remember that having a low expectation for the observations under one proposition does not imply support for the other proposition, since the observations could be even more improbable under the alternative proposition.

Suppose that in the hypothetical case the mine shaft was dug into soil, did not contain any rocks, and was six meters deep. In these circumstances there is a much higher expectation for the three fractures under scenario A (the decedent was assaulted with a shovel and dumped in the mineshaft) than under scenario B (the deceased just fell in the shaft without being beaten). Suppose the probability of the observations is assessed to be higher by a factor of hundreds for scenario A versus scenario B. If the opinion scale as defined in the 'Guideline for Evaluative Reporting' by the European Network of Forensic Science Institutes (ENFSI) [54] is used, LRs in the range between 100 and 1000

are represented as 'moderately strong support'. With reference to that scale, the expert would report that the observations offer 'moderately strong support' for scenario A over scenario B. The 'moderately strong support' is the qualitative LR in this example. This LR does not imply that scenario A is the most probable scenario, as other evidence (prior odds) could point to scenario B. It does, however, mean that this expert opinion offers moderately strong support for the case of the prosecution. How to best communicate (verbal) LRs is discussed in more detail in [55].

Note how the propositions enable the expert to use all their observations to answer the question, instead of focusing on one (often out of context) single element (i.e., the degree of force). This approach increases the amount of information that can be used for the opinion. Instead of being constrained to the limited empirical evidence for the relationship between force and injury morphology, the expert can now use other sources of information as well. For instance, the expert can refer to published literature which provides an evidence base for the types of skull fractures associated with different categories of trauma (e.g., [27,56]), or fracture patterns, i.e., the number, location and morphology of skull fractures in falls [57] vs. assaults [58].

Making the question explicit in the form of propositions allows the expert to provide an LR. It furthermore clarifies the issues most relevant to the court and therefore allows the expert to maximize the relevance of their evidence. Moreover, as previously discussed in various other publications dedicated to the application of Bayes' theorem in forensic science, it helps to maintain the separate roles of the trier-of-fact and the expert, and helps to interpret evidence in a logically correct way. Thus, when asked the right question, the expert can appropriately draw on their expertise and therefore, inform the court in the most meaningful way.

#### **9. Conclusions**

Questions relating to the 'degree of force' often implicitly assume that such an opinion assists the court in establishing whether an injury was caused accidentally or intentionally. As demonstrated in this paper, this assumption is flawed, since theoretical and practical limitations preclude a connection between the 'degree of force' and intent. Similar to forensic biomechanical injury assessment, providing an opinion about the 'degree of force' does not occur in a vacuum [45], that is, all lines of evidence must be considered. The use of Bayes' theorem helps to accomplish this, and therefore enables the expert to maximize the full potential of their evidence.

**Author Contributions:** All authors contributed equally to the paper. Conceptualization, H.H.d.B., C.E.H.B. and S.B.; writing—original draft preparation, H.H.d.B., C.E.H.B. and S.B.; writing—review and editing, H.H.d.B., C.E.H.B. and S.B. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Not applicable.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


**Helena Correia Dias 1,2,3,\*, Licínio Manco <sup>1</sup> , Francisco Corte Real 3,4 and Eugénia Cunha 2,3**


**Simple Summary:** DNA methylation age estimation is one of the hottest topics in forensic field nowadays. Age estimation can be improved under a multidisciplinary approach, the role of a forensic anthropologist and forensic epigeneticist being crucial in the establishment of new basis for age estimation. The development of epigenetic models for bones and tooth samples is crucial in this way. Moreover, developing models for age estimation using several samples can be a useful tool in forensics. In this study, we built two multi-tissue models for age estimation, combining blood, bones and tooth samples and using two different methodologies. Through the Sanger sequencing methodology, we built a model with seven age-correlated markers and a mean absolute deviation between predicted and chronological ages of 6.06 years. Using the SNaPshot assay, a model with three markers has been developed revealing a mean absolute deviation between predicted and chronological ages of 6.49 years. Our results showed the usefulness of DNA methylation age estimation in forensic contexts and brought new insights into the development of multi-tissue models applied to blood, bones and teeth. In the future, we expected that these procedures can be applied to the Medico-Legal facilities to use DNA methylation in routine practice for age estimation.

**Abstract:** The development of age prediction models (APMs) focusing on DNA methylation (DNAm) levels has revolutionized the forensic age estimation field. Meanwhile, the predictive ability of multi-tissue models with similar high accuracy needs to be explored. This study aimed to build multi-tissue APMs combining blood, bones and tooth samples, herein named blood–bone–tooth-APM (BBT-APM), using two different methodologies. A total of 185 and 168 bisulfite-converted DNA samples previously addressed by Sanger sequencing and SNaPshot methodologies, respectively, were considered for this study. The relationship between DNAm and age was assessed using simple and multiple linear regression models. Through the Sanger sequencing methodology, we built a BBT-APM with seven CpGs in genes *ELOVL2*, *EDARADD*, *PDE4C*, *FHL2* and *C1orf132*, allowing us to obtain a Mean Absolute Deviation (MAD) between chronological and predicted ages of 6.06 years, explaining 87.8% of the variation in age. Using the SNaPshot assay, we developed a BBT-APM with three CpGs at *ELOVL2, KLF14* and *C1orf132* genes with a MAD of 6.49 years, explaining 84.7% of the variation in age. Our results showed the usefulness of DNAm age in forensic contexts and brought new insights into the development of multi-tissue APMs applied to blood, bone and teeth.

**Keywords:** DNA methylation (DNAm); epigenetic age estimation; multi-tissue age prediction models (APMs); Sanger sequencing; SNaPshot

#### **1. Introduction**

Age estimation is one of the most important issues in forensic contexts. Among the parameters of the biological profile, the estimate of adult's age at death has always

**Citation:** Correia Dias, H.; Manco, L.; Corte Real, F.; Cunha, E. A Blood–Bone–Tooth Model for Age Prediction in Forensic Contexts. *Biology* **2021**, *10*, 1312. https:// doi.org/10.3390/biology10121312

Academic Editors: Ann H. Ross and Andrés Moya

Received: 29 October 2021 Accepted: 7 December 2021 Published: 10 December 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

been problematic in forensic anthropology since skeletal aging continues to be largely unknown, and all the available methods continue to fail in the approximation to the real age. In other words, there is a discrepancy between biological and chronological ages; the older, the worse. Despite significant research that has been conducted to face problems of adults' age at death, there is not a model of age prediction that can be considered very satisfactory. In particular, aging the elderly is lacking age indicators that can discriminate among individuals of seventy, eighty and ninety. Apart from that, the methods that can be applied always depend both on the state of completeness and preservation of the human remains. In forensic anthropology practice, there are many situations where the targeted age indicators are missing and where alternatives are needed. That is the case of some burned remains, dismembered bodies and incomplete bodies, among others. On the other hand, in the case of a fresh body of an unidentified victim, where physiognomic traits are no longer available and with no suspicion of identity, age is always a needed parameter. In those situations, an alternative is also required. Although imaging methods could be a good alternative, we here argue that the genetic approach by means of DNA methylation (DNAm) is also a good choice.

DNAm analysis for age estimation has emerged in the forensic field in recent years. Several age-related markers have been investigated in different tissues, including blood, saliva, buccal swabs, sperm, teeth and bones, allowing the development of tissue-specific age prediction models (APMs) with high accuracy [1]. The development of multi-tissue APMs brought many advantages in forensics, since they can be applied to several contexts with different types of samples. However, the discovery of universal biomarkers of age applied simultaneously to many tissue types can be a challenge, since it has been observed that only a few markers can work well as multi-tissue age predictive markers [2].

To our knowledge, only three reports addressed multi-tissue DNAm age prediction in human individuals. Horvath [3] assessed methylation information of 353 CpGs, developing a highly accurate multi-tissue age predictive model showing a strong correlation between predicted and chronological ages (R = 0.97), and revealing a median absolute difference between chronological and predicted ages of 2.9 years (training set) and 3.6 years (test set). The high accuracy can be explained by the larger number of CpGs included in the model. However, a high number of age markers can also bring a challenge for forensic casework application. Moreover, in the Horvath model a larger error (around 10 years) was observed in several tissues suggesting that the best markers for one tissue may not be the best for another. Using published databases, Alsaleh et al. [4] identified a small set of 10 CpG sites and built a multi-tissue model for blood, semen, saliva, menstrual blood and vaginal secretions with a Mean Absolute Deviation (MAD) from chronological age of 3.8 years. Jung et al. [2] developed a multi-tissue APM applied to blood, buccal swabs and saliva with DNAm captured by a SNaPshot assay using five CpGs located at *ELOVL2, FHL2, C1orf132, KLF14* and *TRIM59* genes. The multi-tissue model showed high accuracy with a MAD from chronological age of 3.553 years. This MAD value was similar to that reported in the same study when developing tissue-specific APMs (MAD = 3.17 years in blood; MAD = 3.82 years in buccal swabs; MAD = 3.29 years in saliva). In addition, Jung and colleagues [2] have observed that the *FHL2* gene is more tissue-specific, revealing strong positive age correlation values in saliva and blood, and a weak age correlation in buccal swabs. They observed also that *ELOVL2* and *TRIM59* seem to work as better multi-tissue markers than *FHL2, C1orf132* or *KLF14*.

Our group previously assessed the methylation information of age-correlated CpG sites in genes *ELOVL2*, *FHL2*, *EDARADD*, *PDE4C*, *C1orf132*, *TRIM59* and *KLF14,* captured by Sanger sequencing and SNaPshot methodologies [5–8]. Several tissue-specific APMs were developed, including for blood [5–7], teeth [8] and bones [8]. Considering the scarcity of multi-tissue APMs developed until now, the present study aimed to reexamine the obtained DNAm levels for these highly age-correlated genes combining the previously addressed tissues to test for a multi-tissue blood–bone–tooth age prediction model (BBT-APM).

#### **2. Materials and Methods**

#### *2.1. Population Sample*

A total of 185 samples (76 females, 109 males; aged 1–94 years old) from living and deceased individuals from blood, bones and teeth previously addressed for DNAm levels by Sanger sequencing in genes *ELOVL2* (9 CpGs), *EDARADD* (4 CpGs), *FHL2* (12 CpGs), *PDE4C* (12 CpGs) and *C1orf132* (6 CpGs) [5,6,8], and 168 samples (67 females, 101 males; 1–94 aged years old) from living and deceased individuals previously analyzed using a SNaPshot assay for 5 specific CpG sites in genes *ELOVL2, FHL2, KLF14*, *C1orf132* and *TRIM59* [7,8], were considered for this study. The same samples were addressed in both methodologies; however, some samples failed PCR amplification and were excluded from further analysis, which explains the difference in number between the two methods. The age distribution of each training set was shown in Table S1.

Peripheral blood samples from healthy living individuals of Portuguese ancestry were collected from users of Biobanco—Hospital Pediátrico de Coimbra and other hospitals; blood samples from deceased individuals were collected during routine autopsies, after consulting RENNDA (Registo Nacional de Não Dadores) in Serviço de Patologia Forense da Delegação do Centro do Instituto Nacional de Medicina Legal e Ciências Forenses (INMLCF) and from Bodies Donated to Science (BDS), before the embalming method in Departamento de Anatomia da Faculdade de Medicina da Universidade do Porto (FMUP). Fresh bone samples (rib) were collected, after consulting RENNDA, during autopsy in Serviço de Patologia Forense das Delegações do Centro e Sul do INMLCF. Tooth samples (molars) from living individuals were collected in dentist offices, after written informed consent, and tooth samples from deceased individuals (molars) were collected from BDS in Departamento de Anatomia da FMUP. We excluded individuals with known diseases or other clinical conditions that could influence DNAm levels. All blood and bone samples from dead bodies were collected within five days after death.

The herein developed multi-tissue APM using Sanger sequencing includes: 65 blood samples from healthy individuals (42 females, 23 males; aged 1–94 years old), 68 blood samples from deceased individuals (15 females, 53 males; aged 24–91 years old), 23 tooth samples (15 females, 8 males; aged 26–88 years old) and 29 bone samples (4 females, 25 males; aged 26–81 years old). For the multi-tissue APM developed by SNaPshot, 55 blood samples from healthy individuals (34 females, 21 males; aged 1–94 years old), 59 blood samples from deceased individuals (13 females, 46 males; aged 24–91 years old), 23 tooth samples (15 females, 8 males; aged 26–88 years old) and 31 bone samples (5 females, 26 males; aged 26–81 years old) were considered.

The study protocol was approved by the ethical Committee of Faculdade de Medicina da Universidade de Coimbra (n◦ 038-CE-2017). For living individuals, written informed consent was previously obtained from adult participants and from children's parents under the age of 18 years.

#### *2.2. Sanger Sequencing of C1orf132 in Blood Samples from Living Individuals*

As the *C1orf132* gene was not previously addressed in blood samples from living individuals using the Sanger sequencing methodology, the genomic DNA extracted from blood samples of living individuals using the *QIAamp DNA Mini Kit* (Qiagen, Hilden, Germany) was bisulfite converted using the EZ DNA Methylation-Gold Kit (Zymo Research, Irvine, CA, USA), and submitted to polymerase chain reaction (PCR) amplification using the Qiagen Multiplex PCR kit (Qiagen, Hilden, Germany) for a selected region of *C1orf132*, as previously described [5]. Sequencing was performed in the ABI 3130 sequencer (Applied Biosystems, Foster City) with Big-Dye Terminator v1.1 Cycle Sequencing kit (Applied Biosystems), using primers and conditions previously described [5].

#### *2.3. Statistical Analyses*

Statistical analyses were performed using IBM SPSS statistics software for Windows, version 24.0 (IBM Corporation, Armonk, NY, USA). Linear regression models were used

to analyze the relationships between DNAm levels at CpG sites and chronological age. The simple linear regression coefficients from the highest age-correlated CpGs from each gene for Sanger sequencing data, and from each age-correlated CpG site addressed by SNaPshot, were used to predict the age of individuals in the combined set of blood, bone and tooth samples. For both methodologies, all the statistically significant age-correlated CpG sites were combined for analysis using the stepwise regression approach for selection of the relevant variables to be included in a multi-locus BBT-APM. We calculated the Spearman correlation value, the mean absolute deviation (MAD) and the root mean square error (RMSE) between chronological and predicted ages for the combined training set of samples in both methodologies. For both the training sets, each obtained MAD value was interpreted as either correct or incorrect using a cutoff value according to the standard error (SE) of the estimate calculated for each APM.

In addition, the MAD values were calculated for subsets of four distinct age categories (<30 years, 31–55 years, 56–79 years, >80 years) for each training set used in Sanger sequencing and SNaPshot methodologies.

Validation of the BBT-APMs was performed by 3-fold cross-validation that consists of randomly removing a set of samples from the training set and to develop three independent multiple linear regressions on the remaining samples. Subsequently, each model is used to predict the age of the removed samples assigned as validation sets. An additional validation was performed by splitting the complete data set into two subsets (training and validation sets) and independent regression was calculated for the training set and applied to the validation set. All the independent linear regression equations developed for validation purposes included the same CpG sites that have been selected for development of the final multi-tissue APM for each methodology.

#### **3. Results**

#### *3.1. Multi-Tissue BBT-APM using Sanger Sequencing*

DNAm levels of 43 CpGs located at *ELOVL2* (9 CpGs), *EDARADD* (4 CpGs), *FHL2* (12 CpGs), *PDE4C* (12 CpGs) and *C1orf132* (6 CpGs) were assessed in a combined training set of 185 samples, including blood, teeth and bones from Portuguese individuals (76 females, 109 males; aged 1–94 years) using the bisulfite PCR sequencing methodology. The simple linear regression analysis showed that the strongest age-correlated site in each gene was: *ELOVL2* CpG6 (R = 0.759, *<sup>p</sup>*-value = 6.87 <sup>×</sup> <sup>10</sup>−36), explaining 57.3% of the variation in age; *FHL2* CpG1 (R = 0.692, *<sup>p</sup>*-value = 1.11 <sup>×</sup> <sup>10</sup>−27), explaining 47.6% of the variation in age; *EDARADD* CpG3 (R = <sup>−</sup>0.682, *<sup>p</sup>*-value = 1.21 <sup>×</sup> <sup>10</sup>−26), explaining 46.2% of the variation in age; *C1orf132* CpG1 (R = <sup>−</sup>0.654, *<sup>p</sup>*-value = 5.67 <sup>×</sup> <sup>10</sup>−24), explaining 42.5% of the variation in age and *PDE4C* CpG2 (R = 0.613, *<sup>p</sup>*-value = 1.79 <sup>×</sup> <sup>10</sup>−20), explaining 37.2% of the variation in age (Table 1 and Supplementary Table S2). A clear positive age correlation was observed for *ELOVL2* CpG6, *PDE4C* CpG2 and *FHL2* CpG1 markers, and a clear negative age correlation was observed for *EDARADD* CpG3 and *C1orf132* CpG1 markers (Supplementary Figure S1). The predicted age of individuals was calculated using the simple linear regression coefficients for the individual strongest age-associated markers allowing us to obtain MAD values of 12.01 years for *ELOVL2* CpG6, 13.23 years for *C1orf132* CpG1, 13.52 years for *EDARADD* CpG3, 13.16 years for *FHL2* CpG1 and 13.58 years for *PDE4C* CpG2 (Table 1).

Simultaneously testing the 35 significant age-associated CpGs from *ELOVL2* (nine CpGs), *EDARADD* (three CpGs)*, FHL2* (nine CpGs), *PDE4C* (eight CpGs) and *C1orf132* (six CpGs) using stepwise regression analysis allowed us to select a multi-locus APM combining seven CpGs (*EDARADD* CpG3, *FHL2* CpG5, *FHL2* CpG11, *ELOVL2* CpG5, *PDE4C* CpG5, *PDE4C* CpG9, *C1orf132* CpG3). The multiple regression analysis combining these CpGs enabled an age correlation (R) value of 0.940 (*p*-value = 7.36 <sup>×</sup> <sup>10</sup>−79), explaining 87.8% of the variation in age (corrected R<sup>2</sup> = 0.878) (Table 1). The formula to predict age of individuals built with the multiple linear regression coefficients (Supplementary Table S3) was as follows: 26.852 − 24.767 × DNAm level *EDARADD* CpG3 + 68.537 × DNAm

level *FHL2* CpG5 − 51.319 × DNAm level *FHL2* CpG11 + 57.461 × DNAm level *ELOVL2* CpG5 + 41.449 × DNAm level *PDE4C* CpG5 − 66.397 × DNAm level *PDE4C* CpG9 − 27.418 × DNAm level *C1orf132* CpG3. The correlation between predicted and chronological ages was 0.915 (Spearman correlation coefficient) with a MAD from chronological age of 6.06 years (RMSE = 7.60) (Figure 1). Correct predictions were 73%, assuming that chronological and predicted ages match around eight years, according to the standard error of estimate calculated for the final APM (SE = 7.86). *Biology* **2021**, *10*, x FOR PEER REVIEW 6 of 15

**Figure 1.** Predicted age versus chronological age using the multi-locus multi-tissue APM developed for *ELOVL2*, *FHL2*, *EDARADD*, *PDE4C* and *C1orf132* genes including blood samples from living individuals (1), blood samples from deceased individuals (2), bone samples (3), tooth samples from living individuals (4) and tooth samples from deceased individuals (5). The corresponding Spearman correlation coefficients (r) are depicted inside each plot. **Figure 1.** Predicted age versus chronological age using the multi-locus multi-tissue APM developed for *ELOVL2*, *FHL2*, *EDARADD*, *PDE4C* and *C1orf132* genes including blood samples from living individuals (1), blood samples from deceased individuals (2), bone samples (3), tooth samples from living individuals (4) and tooth samples from deceased individuals (5). The corresponding Spearman correlation coefficients (r) are depicted inside each plot.

The accuracy of the developed BBT-APM was tested through a threefold cross validation in the training set of 185 samples showing a MAD of 6.27 years (RMSE = 6.27) (mean value obtained for the three test sets). This value was very close to the MAD of 6.06 (RMSE = 7.60) obtained in the whole training set. The validation by splitting the overall training set into two sets of 117 and 68 samples (training and validation sets) allowed us to obtain an independent MAD value for the training set of 6.09 years (RMSE = 7.55); applying the model on the validation set, a MAD of 6.08 years (RMSE = 7.64) was obtained. Both inde-The accuracy of the developed BBT-APM was tested through a threefold cross validation in the training set of 185 samples showing a MAD of 6.27 years (RMSE = 6.27) (mean value obtained for the three test sets). This value was very close to the MAD of 6.06 (RMSE = 7.60) obtained in the whole training set. The validation by splitting the overall training set into two sets of 117 and 68 samples (training and validation sets) allowed us to obtain an independent MAD value for the training set of 6.09 years (RMSE = 7.55); applying the model on the validation set, a MAD of 6.08 years (RMSE = 7.64) was obtained. Both independent MAD values were very close to the MAD of 6.06 (RMSE = 7.60) obtained in the whole training set.

#### pendent MAD values were very close to the MAD of 6.06 (RMSE = 7.60) obtained in the whole training set. *3.2. Multi-Tissue BBT-APM Using SNaPshot Methodology*

*3.2. Multi-Tissue BBT-APM Using SNaPshot Methodology*  DNAm levels at five CpG sites from the *ELOVL2, FHL2, KLF14*, *C1orf132* and *TRIM59* genes obtained through a SNaPshot assay were simultaneously addressed in a combined

DNAm levels at five CpG sites from the *ELOVL2, FHL2, KLF14*, *C1orf132* and *TRIM59* 

set of 168 samples, including blood, bones and teeth (67 females, 101 males; 1–94 aged years old). DNAm levels of *ELOVL2*, *FHL2*, *KLF14* and *TRIM59* genes revealed a positive correlation with age, and DNAm levels of *C1orf132* showed a negative correlation (Supplementary Figure S2). Testing the individual DNAm association with chronological age for the five CpG sites, the strongest correlation was observed for *ELOVL2* (R = 0.772, *p*value = 1.54 × 10−34), explaining 59.4% of the variation in age, followed by *C1orf132* (R = −0.693, *p-*value = 2.49 × 10−25), explaining 47.7% of the variation in age, *FHL2* (R = 0.686, *p*value = 1.36 × 10−24), explaining 46.8% of the variation in age, *KLF14* (R = 0.677, *p-*value = 6.57 × 10−24), explaining 45.6% of the variation in age and *TRIM59* (R = 0.584, *p-*value = 1.17

set of 168 samples, including blood, bones and teeth (67 females, 101 males; 1–94 aged years old). DNAm levels of *ELOVL2*, *FHL2*, *KLF14* and *TRIM59* genes revealed a positive correlation with age, and DNAm levels of *C1orf132* showed a negative correlation (Supplementary Figure S2). Testing the individual DNAm association with chronological age for the five CpG sites, the strongest correlation was observed for *ELOVL2* (R = 0.772, *<sup>p</sup>*-value = 1.54 <sup>×</sup> <sup>10</sup>−34), explaining 59.4% of the variation in age, followed by *C1orf132* (R = <sup>−</sup>0.693, *<sup>p</sup>*-value = 2.49 <sup>×</sup> <sup>10</sup>−25), explaining 47.7% of the variation in age, *FHL2* (R = 0.686, *<sup>p</sup>*-value = 1.36 <sup>×</sup> <sup>10</sup>−24), explaining 46.8% of the variation in age, *KLF14* (R = 0.677, *<sup>p</sup>*-value = 6.57 <sup>×</sup> <sup>10</sup>−24), explaining 45.6% of the variation in age and *TRIM59* (R = 0.584, *<sup>p</sup>*-value = 1.17 <sup>×</sup> <sup>10</sup>−16), explaining 33.7% of the variation in age (Table 2). The simple APMs for each CpG site allowed us to obtain MAD values from a chronological age of 10.95 years for *ELOVL2*, 12.10 years for *C1orf132*, 12.63 years for *FHL2,* 12.74 years for *KLF14* and 13.64 years for *TRIM59* (Table 2).

Applying the stepwise regression approach to the five CpG sites, only the CpGs located at *ELOVL2, KLF14* and *C1orf132* genes were selected for the development of a final multi-locus APM. The three selected CpGs revealed in the multiple regression analysis a very strong correlation with age, R = 0.922 (*p*-value = 3.14 <sup>×</sup> <sup>10</sup>−67), explaining 84.7% of the variation in age (corrected R<sup>2</sup> = 0.847) (Table 2). Predicted age through the multivariate regression coefficients was as follows (Supplementary Table S4): 29.220 + 96.850 × DNAm level *ELOVL2* + 208.747 × DNAm level *KLF14* − 33.437 × DNAm level *C1orf132.* This BBT-APM allowed us to obtain a MAD from chronological age of 6.49 years (RMSE = 8.42) (Table 2). Correct predictions were 73.8% considering the cutoff of 9 years, according to the standard error of estimate calculated for the final APM (SE = 8.53). The obtained Spearman correlation value between predicted and chronological ages was 0.893 (Figure 2). *Biology* **2021**, *10*, x FOR PEER REVIEW 8 of 15

**Figure 2.** Predicted age versus chronological age using the multi-tissue APM developed for *ELOVL2*, *C1orf132* and *KLF14* genes including blood samples from living individuals (1), blood samples from deceased individuals (2), bone samples (3), tooth samples from living individuals (4) and tooth samples from deceased individuals (5). The corresponding Spearman correlation coefficients (r) are depicted inside each plot. **Figure 2.** Predicted age versus chronological age using the multi-tissue APM developed for *ELOVL2*, *C1orf132* and *KLF14* genes including blood samples from living individuals (1), blood samples from deceased individuals (2), bone samples (3), tooth samples from living individuals (4) and tooth samples from deceased individuals (5). The corresponding Spearman correlation coefficients (r) are depicted inside each plot.

The model accuracy of the final APM with DNAm levels of *ELOVL2*, *KLF14* and *C1orf132* markers was evaluated through a threefold cross validation in the training set of

Evaluating the model performance obtained with the two developed multi-tissue BBT-APMs according to different age ranges (Table 3), we observed an increase in the MAD values between predicted and chronological ages with the increase in age of individuals. For both Sanger sequencing and SNaPshot methodologies, the value of MAD was the largest for the age group >80 years and the smallest for age group <30 years (Table 3).

the whole training set. The validation by splitting the overall training set into two sets of 113 and 55 samples (training and validation sets) allowed us to obtain an independent MAD value for the training set of 6.06 years (RMSE = 7.81). Applying the model on the

validation set, a MAD of 7.45 years (RMSE = 9.60) was obtained.

*3.3. Differences between Predicted and Chronological Ages with an Increase in Age*





The model accuracy of the final APM with DNAm levels of *ELOVL2*, *KLF14* and *C1orf132* markers was evaluated through a threefold cross validation in the training set of 168 samples, producing a MAD (mean value obtained for the three test sets) of 6.73 years (RMSE = 6.75). This value was very close to the MAD of 6.49 (RMSE = 8.42) obtained in the whole training set. The validation by splitting the overall training set into two sets of 113 and 55 samples (training and validation sets) allowed us to obtain an independent MAD value for the training set of 6.06 years (RMSE = 7.81). Applying the model on the validation set, a MAD of 7.45 years (RMSE = 9.60) was obtained.

#### *3.3. Differences between Predicted and Chronological Ages with an Increase in Age*

Evaluating the model performance obtained with the two developed multi-tissue BBT-APMs according to different age ranges (Table 3), we observed an increase in the MAD values between predicted and chronological ages with the increase in age of individuals. For both Sanger sequencing and SNaPshot methodologies, the value of MAD was the largest for the age group >80 years and the smallest for age group <30 years (Table 3).

**Table 3.** Evaluation of mean absolute deviation (MAD) between chronological and predicted ages according to four age-range groups in the training set of blood, bone and tooth samples using both methodologies.


#### **4. Discussion**

In the past decade, several specific epigenetic clocks with high accuracy have been developed using many tissue types [9–16]. However, the discovery of DNAm age-related markers with similarly high accuracy across different types of tissues (universal markers) remains a challenging task in the forensic field [17]. Evidence from previous studies shows that each age-correlated marker reveals a specific ability to predict chronological age, as each tissue type can be affected by different intrinsic or environmental factors. Eipel et al. [16] reported that using a specific APM with methylation information of age-correlated markers selected in one tissue-specific type can lead to a decrease in model accuracy in age prediction if applied to a different tissue. This should be related to the tissue-specific differences in epigenetic patterns [18–20]. Thus, a careful selection of age-associated CpGs and the validation of these proposed markers in each tissue type should be the first step for the development of multi-tissue APMs.

In fact, until now, only a few studies have explored the predictive ability of multitissue APMs [2–4]. In this study, we re-examined DNAm levels of *ELOVL2, FHL2, PDE4C, EDARADD, C1orf132, TRIM59* and *KLF14* genes, previously captured in different tissue types (blood samples from living and deceased individuals; tooth samples from living and deceased individuals; fresh bone samples collected during autopsies) by Sanger sequencing and SNaPshot methodologies to build multi-tissue APMs. We developed simple linear regression APMs for the best-selected CpG sites from each gene, and multilocus multi-tissues APMs using the best combination of CpGs selected by the stepwise regression approach.

DNAm levels captured by bisulfite Sanger sequencing allowed the development of a final APM with seven CpGs (*EDARADD* CpG3, *FHL2* CpG5, *FHL2* CpG11, *ELOVL2* CpG5, *PDE4C* CpG5, *PDE4C* CpG9, *C1orf132* CpG3), revealing a very strong age correlation value (R = 0.940), highly significant (*p*-value = 7.36 <sup>×</sup> <sup>10</sup>−79) and explaining 87.8% of the variation

in age. The BBT-APM developed with 185 Portuguese individuals (aged 1–94 years old) allows us to predict age with a moderate accuracy showing a MAD from chronological age of 6.06 years.

Regarding methylation information captured by the SNaPshot methodology, the final multi-locus multi-tissue APM combines three CpG sites located at *ELOVL2, KLF14* and *C1orf132* genes. This BBT-APM developed in 168 samples revealed a very strong age correlation value (R = 0.922), with a MAD from chronological age of 6.49 years.

In Table 4, we resume in brief the difference in results obtained with both methodologies. The multi-tissue APMs developed herein allows prediction of age of the individuals based on evaluation of DNAm levels captured from several types of samples, including blood, bone and teeth. The final models revealed an accuracy (MAD value) of about 6 years, being more accurate than the majority of anthropological approaches applied to adults' age estimation. When comparing the results with the ones retrieved by anthropological methods, it becomes clear that our method has clear benefits in relation to methods such as Suchey–Brooks', where age ranges are particularly large, mainly for old individuals.

**Table 4.** Comparison between Sanger sequencing and SNaPshot methodologies.


Comparing the herein developed multi-tissue APMs with the tissue-specific APMs previously developed by our group, we can observe that through Sanger sequencing, the blood-living APM [6] revealed a MAD of 5.35 years, which is a slightly lower value comparing with the BBT-APM (MAD = 6.06 years). However, for blood samples from deceased individuals [5], the tissue-specific APM revealed a similar accuracy with a MAD of 6.08 years. The tissue-specific APMs developed through the SNaPshot assay for blood samples revealed MAD values of 4.25 and 5.36 years for living and deceased individuals, respectively [7]. However, although these models have a better accuracy than the herein developed BBT-APM using the SNaPshot methodology (MAD = 6.49), they can only be applied to blood samples.

Regarding bones, we have previously obtained through Sanger sequencing and SNaPshot methodologies MAD values of 2.56 and 7.18 years, respectively [8]. Thus, we can observe that for age prediction in bones using Sanger sequencing, it is more advantageous to apply the tissue-specific model compared with the BBT-APM (MAD = 6.06 years). However, using the SNaPshot methodology we obtained a similar prediction accuracy for both the specific bone-APM (MAD = 7.18 years) and the BBT-APM (MAD = 6.49 years). In regards to tooth samples, the tissue-specific models previously developed [8] revealed MAD values of 11.35 years and 7.07 years using Sanger sequencing and SNaPshot methodologies, respectively, which is a lower accuracy in comparison with the BBT-APMs developed in this present study (MAD = 6.06 and 6.49 years, respectively).

Previous reports using DNAm levels for the development of multi-tissues APMs [2–4] showed higher prediction accuracy in age estimation (MAD values of 2.9, 3.55 and 3.8 years). In our study, the obtained higher MAD values (6.06 years in Sanger sequencing and 6.49 years in SNaPshot) can be explained by sample size, population variability or the laboratory methodologies for DNAm assessment. Of note, both developed BBT-APMs

included CpGs from the *ELOVL2* gene revealing the powerful of this age-associated gene for the development of multi-tissue APMs in forensic contexts. It has been shown that *ELOVL2* is a stable epigenetic marker, revealing a high performance as a multi-tissue predictor [2,13,14,21]. This locus has been used as a powerful age-correlated marker in many tissue-specific APMs developed for blood, tooth, bones and buccal swabs, revealing similar patterns of high accuracy in all APMs [2,10–15,22–30]. Moreover, it has been shown that CpGs from the other genes addressed in the present study also revealed higher age correlation values in blood samples [2,5–7,10–12,23,24,26,28–30], bones [8,13,14] and tooth samples [8,15,23,27], being promising markers to be selected for development of universal APMs.

Several aspects should be highlighted for future potential applicability of the hereindeveloped multi-tissues APMs.

In this study, both BBT-APMs revealed a general decrease in model accuracy (increase in MAD value) with the increase in age, in accordance with previous studies [3,11,12,26,30], revealing that age estimation based on DNAm levels can have a better performance in younger age ranges. Indeed, younger individuals show lower values of MAD reflecting a high accuracy in the APMs, comparing to older ages. This reflects larger differences between biological and chronological ages with the increase in age, related to the accumulation of specific alterations in DNAm patterns of each individual with aging due the stochastic or environmental factors, being accepted as the epigenetic drift contribution [31–33].

The possibility that postmortem changes can alter the methylation status among specific loci should also be hypothesized, and this issue needs future clarification. As reported in previous studies from our group, comparing blood samples from living and deceased individuals [6,7], it is important for forensic casework application to know the healthy status of the sample donor. This is a paramount issue because the most developed APMs until now have been built using samples of living individuals. It has been observed that ancient DNA (aDNA) can suffer postmortem miscoding lesions, as deamination [34,35]. Postmortem deamination is a spontaneously chemical process that occurs due to the hydrolytic deamination of cytosine (C) residues into uracils (U) [34]. If DNA damage in the form of deamination occurs, the expected residues in PCR amplification could be different after bisulfite conversion. Bisulfite conversion is a chemical modification, which mediates the deamination of unmethylated C to U, appearing after PCR amplification as thymine (T), but leaves methylated C (5mC) intact. Therefore, if postmortem cytosine deamination occurs, both unmethylated C and 5mC appear as T after PCR amplification of bisulfite-converted samples, which could disturb the measurement of DNAm levels. As hydroxymethylcytosine (5hmC) is an oxidative product of demethylation of 5mC [36,37], in case of postmortem deamination, the 5hmC concentration can also be affected. Despite this, the stability of 5mC patterns in aDNA has been reported, when preserved aDNA samples were analyzed [38,39]. Moreover, Pedersen et al. [40] assessed to DNAm levels of permafrost hair samples collected from a Paleo-Eskimo with 4000 years old, and predicted age at death. This reveals the reliability on the assessment of DNAm levels to predict age in ancient samples.

An additional important issue for forensic practice is the effect of postmortem interval (PMI) on DNAm levels captured from aged forensic samples of different tissues. Data obtained from such forensic samples should be interpreted with caution due to the very low amount and degradation of the obtained DNA. A previous study developed by Zbie´c-Piekarska et al. [24] showed the stability of prediction accuracy using bloodstains that differed in time of storage. The authors evaluated DNA concentrations from bloodstains that had been deposited previously on tissue paper and kept at room temperature during 5, 10 and 15 years, observing a significant decrease in DNA concentration, a decrease in number of positive PCR amplifications and an increase in average degradation index. However, they did not observe an effect on the rate of corrected predictions, reporting that "the prediction success rate seemed not to correlate inversely with increasing time of storage" [24]. Hence, it seems that DNA degradation affects DNA concentration and, consequently, the rate of positive PCR amplifications; however, the accuracy of age prediction is not affected in positive PCR amplification samples.

The major drawback of our study was the limited number of samples, mainly in bones and teeth. We recognize that larger sample sets have greater statistical power and may be more representative of DNAm changes related to different age groups and different types of tissues, leading to the development of more accurate APMs. Another relevant factor that should be considered is the existence of some diseases or clinical conditions or even some life routines such as smoking or drinking, which may interfere with methylation data. For samples of deceased individuals, despite having access to medical reports of each case, information related to possible clinical conditions was unknown in many cases. Lastly, the use of different methodologies for evaluation of DNAm levels across studies can influence the accuracy of APMs. In particular, bisulfite sequencing or SNaPshot methodologies are semi-quantitative methods and thus may not be the optimal tool for precise DNAm analysis.

DNAm analysis is considered a promising method for age estimation in the future. If we question how easy it is to use it and how long it takes to apply it, we argue that in those laboratories supported by genetic facilities provided with the needed equipment, the results can be retrieved in two or three days. In comparison with the more traditional approaches, it takes longer, but in terms of the delivery of the final report, it does not imply any delay. Furthermore, it should be noted that any method that involves DNA analysis turns out to be more expensive, but it also tends to be more reliable. However, it should be emphasized that the development of universal APMs based on DNAm levels is at the beginning of age estimation research and, therefore, the herein proposed BBT-APMs can be a starting point for future research.

#### **5. Conclusions**

In conclusion, in this study we re-examined DNAm levels of *ELOVL2, FHL2, PDE4C, EDARADD, C1orf132, TRIM59* and *KLF14* genes previously captured by Sanger sequencing and SNaPshot methodologies across several tissues. Two multi-tissue BBT-APMs were developed using blood, tooth and bone samples from Portuguese individuals. To the best of our knowledge, the two BBT-APMs developed herein for the Portuguese population are the first multi-tissue APMs using bones and teeth. Moreover, despite being very often found in forensic contexts, the development of tissue-specific APMs using bones or teeth is scarce in forensic research. By Sanger sequencing, a moderate accuracy of 6.06 years was obtained in the BBT-APM using seven CpGs from genes *ELOVL2, FHL2, PDE4C, EDARADD* and *C1orf132*. Using the SNaPshot assay, the BBT-APM developed with methylation data from *C1orf132*, *ELOVL2* and *KLF14* genes revealed a MAD from chronological age of 6.49 years. Both methodologies revealed similar accuracy for use in multi-tissue APMs being both simple, rapid, cost-effective and easily available in forensic laboratories. Therefore, both BBT-APMs developed herein can be a promising tool for age estimation in forensic contexts.

This article, a priori, could appear too technical and a little far away from the forensic anthropology reality. However, we argue that a bridge between forensic genetics and forensic anthropology can be achieved, once the needed complicities between the experts involved are well established. In practical terms, what we here advise is an integrated evaluation of the case by the forensic anthropologist, along with the pathologist in charge of the case. If, for instance, the case is a fresh body without any physiognomic traits and where identification is unknown, blood is the best option for DNAm age estimation. If, on the other hand, blood is no longer available due to the state of decomposition of the body, a decision can be made to recover both bone and teeth to perform DNAm studies. What does that imply in practical terms? It means that the result will take 2 or 3 days to be known, that the needed equipment is necessary as well as the adequate kits. While those ones are more expensive than the blood ones, it is a good option in particular when the most adequate skeletal age indicators are damaged or no longer available. Having said

that, we argue that we should strive to implement the procedures here described in the Medico–Legal facilities in order to turn DNAm a routine practice for age estimation.

**Supplementary Materials:** The following are available online at https://www.mdpi.com/article/10.3 390/biology10121312/s1: Figure S1: Correlations between DNAm levels and chronological age in 185 samples including blood samples from living and deceased individuals, bone samples collected from autopsies and teeth from living and deceased individuals, obtained through Sanger sequencing methodology. Figure S2: Correlations between DNAm levels and chronological age in 168 samples, including blood samples from living and deceased individuals, bone samples collected from autopsies and teeth from living and deceased individuals, obtained through SNaPshot methodology. Table S1: Age distribution of the sample sets analyzed by Sanger sequencing and SNaPshot methodologies., Table S2: Univariate linear regression analysis of the 43 CpG sites in *ELOVL2*, *FHL2*, *EDARADD*, *PDE4C* and *C1orf132* loci in 185 samples including blood from living and deceased individuals, teeth from living and deceased individuals and bone collected during autopsies. Table S3: Statistical parameters obtained in a multiple regression model with the seven CpGs in genes *ELOVL2*, *FHL2*, *EDARADD*, *PDE4C* and *C1orf132* selected by stepwise regression approach, in blood, bone and tooth samples. Table S4: Statistical parameters obtained in a multiple regression model with the three CpGs in genes *ELOVL2*, *C1orf132* and *KLF14*, selected by stepwise regression approach, in blood, bone and tooth samples.

**Author Contributions:** Conceptualization, H.C.D., L.M. and E.C.; data curation, H.C.D.; investigation, H.C.D. and L.M.; methodology, H.C.D. and L.M.; supervision, L.M., F.C.R. and E.C.; writing original draft, H.C.D.; writing—review and editing, L.M., F.C.R. and E.C. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research has funding by Fundação para a Ciência e a Tecnologia (FCT) (UIDB/00283/2020). H.C.D. had a PhD grant from FCT (SFRH/BD/117022/2016).

**Institutional Review Board Statement:** The study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the Ethics Committee of Faculdade de Medicina da Universidade de Coimbra (n◦ 038-CE-2017, 26/06/2017) and by the Instituto Nacional de Medicina Legal e Ciências Forenses (INMLCF).

**Informed Consent Statement:** Informed consent was obtained from all subjects involved in the study.

**Data Availability Statement:** Not applicable.

**Acknowledgments:** The authors would like to thank to all the volunteers that participated in this study.

**Conflicts of Interest:** The authors declare no conflict of interest.

## **References**

