🧬 biology

In which cases does it make sense to classify a trace into one population, and what are the alternatives?

This study evaluates classification methods for biogeographical ancestry analysis using 104 AIMs, demonstrating that a recent naive Bayes (RNB) classifier effectively distinguishes between in-reference, admixed, and out-of-reference individuals, thereby addressing the limitations of assigning unknowns to a single reference population.

Original authors: Julika Schmitz, Jana Naue, Lukas Hoffmann, Zoe Lange, Suchanek Ferdinand, Sabine Lutz-Bonengel, Holger Dette, Peter Pfaffelhuber, Carola Sophia Heinzel

Published 2026-10-19

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Julika Schmitz, Jana Naue, Lukas Hoffmann, Zoe Lange, Suchanek Ferdinand, Sabine Lutz-Bonengel, Holger Dette, Peter Pfaffelhuber, Carola Sophia Heinzel

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

When a crime scene yields a tiny drop of blood or a single hair, forensic scientists often turn to DNA to learn who the person might be. Beyond simply matching a suspect, investigators can now look at genetic markers to guess where a person's ancestors came from, what color their eyes might be, or how old they are. This process, known as biogeographical ancestry analysis, works by comparing the DNA found in the evidence against a library of reference groups from around the world. Imagine these reference groups as distinct neighborhoods; the goal is to see which neighborhood the DNA profile fits into best. However, the real world is rarely as neat as a library catalog. People move, families mix, and populations exist that no one has sampled yet. If a person's parents come from two different continents, or if their family comes from a region never included in the database, forcing their DNA into a single "neighborhood" can lead to a wrong conclusion. The critical question for investigators is not just how to assign a label, but when it is safe to assign one at all.

A team of researchers from universities in Germany and the Ruhr area set out to solve this specific problem. They asked when it makes sense to classify a trace into one specific population and, just as importantly, when the system should admit it does not know. To answer this, they tested seven different statistical methods using a set of 104 genetic markers, a standard toolkit used in modern forensics. Their goal was to build a system that could do three things: confidently assign a person to a known population, identify if a person has parents from two different populations, or declare that the person's background is too different from the reference data to be classified.

The researchers found that the most reliable approach was a method they called the "recent naive Bayes" classifier. Unlike older methods that might force a guess even when the data is weak, this new system acts like a careful gatekeeper. It first calculates the most likely pair of ancestral populations for an individual. If the person is not mixed, the system predicts the same population for both parents. If the person is mixed, it predicts two different ones. Crucially, the system then checks a confidence score. If the genetic evidence is strong enough to pass a specific threshold, it reports the result. If the evidence is too ambiguous, the system refuses to guess and reports the sample as "out of reference." This refusal is not a failure; it is a safeguard against error.

In their tests, the researchers simulated thousands of scenarios to see how well each method performed. They created artificial children from parents in the reference database to test the system's ability to spot mixed ancestry. They also held back entire populations from the training data to see if the system could recognize when a person did not belong to any of the known groups. The best-performing method correctly identified people from unknown populations in nearly 85 percent of cases, effectively rejecting them rather than mislabeling them. When the system did accept a classification, it was highly accurate. For example, when the system identified a person as having parents from two different groups, it correctly named both parental origins in 64 percent of cases and got at least one right in almost all others.

The researchers also tested their method on real-world samples from 28 volunteers who provided their own family history. In this group, the system successfully classified 16 individuals, matching their self-reported ancestry with high precision. For the other 12 people, the system correctly declined to make a guess. In many of these rejected cases, the individuals had complex family histories, such as a parent who was already mixed or a grandparent from a distant region, which made the genetic signal too fuzzy for a single label. The system's refusal to guess in these difficult cases prevented it from making incorrect assignments that could mislead an investigation.

One of the key findings was that the size of the reference database matters. The study used a panel of 104 markers that is excellent for distinguishing between major continental groups like East Asia, Europe, and Africa. However, the researchers noted that some populations are so genetically similar that even the best markers struggle to tell them apart. For instance, distinguishing between people from the Middle East and certain parts of Europe proved difficult, leading the researchers to treat them as a single group in their tests. This highlights a limitation: the system can only say a person is "out of reference" if they are truly different enough from the groups in the database. If a missing population is very close to a known one, the system might still assign the person to the known group, which is a reasonable approximation but not a perfect match.

The study also addressed the issue of mixed ancestry. Traditional methods often struggle with people who have parents from different backgrounds, sometimes forcing them into a single category or failing to identify the mix entirely. The new method treats mixed ancestry as a natural part of the model. It does not need to be taught with fake mixed examples; instead, it understands that a person is a combination of two parents. This allows it to report a mixed result, such as "one parent from Europe and one from South Asia," rather than just saying "mixed." This level of detail is vital for forensic investigations, as it narrows down the pool of potential suspects more effectively than a vague label.

Ultimately, the paper demonstrates that the most honest answer in forensic genetics is sometimes "we do not know." By combining a sophisticated classification tool with a strict rule for when to accept or reject a guess, the researchers created a system that minimizes the risk of false leads. The system does not just look for the best fit; it evaluates whether that fit is good enough to be trusted. In the high-stakes world of criminal investigation, where a wrong guess could waste resources or point the wrong way, this ability to say "no" is just as valuable as the ability to say "yes." The work confirms that while we can learn a great deal from DNA, we must also respect the limits of our current knowledge and the complexity of human history.

Technical Summary: Classification of Biogeographical Ancestry with Out-of-Reference and Admixed Detection

Problem Statement
In forensic biogeographical ancestry (BGA) analysis, unknown individuals are typically assigned to one of several reference populations using Ancestry Informative Markers (AIMs). However, this standard "closed-set" classification is often inadvisable or misleading in two specific scenarios: (1) the individual is admixed (having parents from two different populations), and (2) the individual originates from a population not represented in the reference database ("out-of-reference"). Existing methods often struggle to distinguish between these scenarios and cases where a single-population assignment is appropriate. Furthermore, many approaches rely on simulated training data for admixture or fail to provide a mechanism to reject assignments that do not fit the reference panel well. This study addresses the need for a classification framework that can simultaneously assign an individual to a single population, identify admixture by specifying parental population pairs, or reject the assignment as out-of-reference, all without prior knowledge of the specific class category.

Methodology
The study evaluated seven classification methods using 104 AIMs from the VISAGE Enhanced Tool. The workflow involved a three-way outcome: In-Reference (IR), Admixed (AD), and Out-of-Reference (OR).

Data: The reference panel consisted of 3,541 individuals from seven continental populations (East Asian, European, South Asian, African, Middle Eastern, Native American, Oceanian). To simulate the OR scenario, one population (or a pair of closely related populations, EUR and MEA) was held out per fold. Admixed individuals were simulated in silico as F1 offspring of two in-reference parents. An additional 800 individuals from seven groups absent from the reference panel and 28 newly sequenced samples with self-reported ancestry served as external validation.
Classifiers: Seven methods were compared:
1. Bayes (NB): Standard classification over populations plus a pooled AD class.
2. Recent Bayes (RNB): A novel classifier operating over unordered population pairs (including self-pairs for non-admixed cases). It models the individual's genotype as inheriting one allele per marker from each of two parents, where each parent belongs to a single reference population.
3. Admixture Model (AM): Estimates continuous ancestry proportions; a cutoff determines if an individual is admixed.
4. Recent Admixture Model (RAM): Estimates proportions separately for two parental halves.
5. TabPFN: A pre-trained tabular foundation model.
6. TVE/TVP: Goodness-of-fit statistics testing against single populations or population pairs.
Acceptance Rule: A critical component of the methodology is the "acceptance rule." Instead of forcing a decision, the classifier assigns a score (likelihood or posterior probability) to the best-matching class. This score is compared against a class-specific threshold calibrated on held-out data. If the score falls below the threshold, the assignment is rejected, and the individual is labeled OR. This allows the system to handle open-set recognition.
Evaluation Metric: Performance was measured using a selection criterion $M$ , defined as the minimum of three retention rates: correct IR assignment ( $r_{IR}$ ), correct AD assignment ( $r_{AD}$ ), and correct OR rejection ( $r_{OR}$ ). Maximizing $M$ ensures the method performs adequately across all three scenarios, preventing the trade-off of high accuracy in one area for poor performance in another.

Key Results

Best Performing Method: The Recent Bayes (RNB) classifier, using the likelihood score ( $s_{lik}$ ), achieved the highest selection criterion ( $M = 0.752$ ). It outperformed the RAM, AM, NB, TabPFN, and goodness-of-fit statistics (TVE/TVP).
Performance Metrics: RNB correctly rejected out-of-reference individuals in 84.9% of cases ( $r_{OR}$ ) while maintaining high accuracy for in-reference (76.8%) and admixed (75.2%) individuals.
Error Patterns: RNB exhibited the most evenly distributed errors, with no off-diagonal confusion exceeding 21%. In contrast, methods like TVE failed to identify admixture entirely ( $M=0$ ), and TabPFN rejected a high proportion of admixed individuals.
External Validation:
- When applied to 800 individuals from populations not in the reference panel, only 29.4% of individuals from four non-admixed groups were rejected. The authors note this is consistent with genetic distance; groups genetically close to reference populations (e.g., North African, Central Asian) were often assigned to the closest reference population rather than rejected.
- In a test set of 28 newly sequenced samples with self-reported ancestry, 16 were classified and 12 were rejected. Among the 16 classified, 12 were exact matches and 4 were partial matches. The rejection of 12 samples prevented partially incorrect classifications in 9 of those cases.
Limitations Observed: The model struggles with complex ancestry involving more than two parental populations (e.g., a sample with three distinct ancestral components was correctly rejected). Additionally, the distinction between closely related populations (e.g., EUR and MEA, $F_{ST} = 0.047$ ) remains challenging with the current 104-marker panel.

Significance and Claims
The paper claims that the primary contribution is the adaptation of the "reject option" (selective classification) to the forensic BGA setting, specifically utilizing a per-class threshold and a likelihood-based score that measures absolute fit rather than relative fit. The authors argue that RNB is the recommended method because it:

Structurally models admixture as a hypothesis over population pairs rather than relying on simulated training offspring.
Provides a deterministic, closed-form estimate for every possible pair of populations.
Balances performance across IR, AD, and OR outcomes, avoiding the pitfalls of methods that excel at one metric while failing at others (e.g., TVE's high rejection rate but zero admixture detection).

The authors modestly note that the recommendation is conditional on the specific evaluation design and the genetic resolution of the 104 AIMs used. They acknowledge that the "out-of-reference" detection is a matter of degree based on genetic distance; populations very close to the reference panel may not be rejected but instead assigned to the nearest reference group. The study concludes that a "not classified" result should not be interpreted as definitive proof of an unknown population, but rather as an indication that the genotype did not provide sufficient support for a specific assignment under the calibrated acceptance rule.

More like this