sample.txt

graphical ancestry. However, they have not been systematically compared to classifiers widely used in other 
disciplines. Noting that genetic data have a tabular form, this study addresses this gap by benchmarking 
forensic classifiers against TabPFN, a cutting-edge, general-purpose machine learning classifier for tabular 
data. The comparison evaluates performance using metrics such as accuracy – the proportion of correct 
classifications – and ROC AUC. We examine classification tasks for individuals at both the intracontinental 
and continental levels, based on a published dataset for training and testing. Our results reveal significant 
performance differences between methods, with TabPFN consistently achieving the best results for accuracy, 
ROC AUC and log loss. E.g., for accuracy, TabPFN improves SNIPPER from 84% to 93% on a continental scale 
using eight populations, and from 43% to 48% for inter-European classification with ten populations.
1. Introduction
Classification of individuals into ancestral populations, i.e., inferring 
biogeographical ancestry (BGA) from DNA traces, is a fundamental task 
in forensic science [1]. This process is particularly useful for identi
fying disaster victims [2] and in broader forensic investigations [3]. 
Additional applications include detecting population stratification [4
6]. Researchers commonly classify individuals into several continental 
populations such as Africa, Europe, East Asia, South Asia, Admixed 
Americans, and Oceania [7,8], or on an intracontinental level.
The statistical task of classification can be broken down into two 
steps. First, genetic markers have to be chosen which are able to dis
tinguish between populations, i.e. with frequency differentials between 
populations. Second, these Ancestry Informative Markers (or AIMs) 
are used as features in a classification task. Here, various algorithms 
are used, leveraging genetic distance, regression, or Bayesian infer
ence [7]. Tools like the naive Bayes classifier Snipper [9] have been used 
alongside multinomial logistic regression [10] and the Genetic Distance 
Algorithm [11]. Furthermore, the Admixture Model, as implemented 
in tools like Structure [12] or Admixture [13], is widely used for 
classification [7]. More recently, other machine learning methods such 
as XGBoost [14–16], Partial Least Squares-Discriminant Analysis (PLS
DA) [14], and random forests [16–18] have been applied in forensic 
genetics to ancestry prediction.
Several challenges must be taken into account for BGA. First, classi
fication becomes increasingly challenging for more genetically similar 
∗ Corresponding author.
E-mail address: carola.heinzel@stochastik.uni-freiburg.de (C.S. Heinzel).
https://doi.org/10.1016/j.fsigen.2025.103290
populations, which are frequently found within continents. Second, 
admixed individuals, such as Americans, present difficulties in classi
fication [19,20]. Third, marker selection significantly affects classifica
tion quality [7], with ideal markers showing large differences in allele 
frequencies between populations [21]. Fourth, selecting populations as 
classes must lead to results which are both, reliable and helpful in 
further investigations.
The choice of Ancestry Informative Markers (AIMs) is frequently 
performed by finding alleles with large frequency differentials be
tween populations using expert knowledge [9,22,23]. However, fea
ture selection for classification is also a well-known task in machine 
learning, which has been applied to the choice of AIMs [24,25]. The 
goal is to find a set which leads to a reduced sequencing effort, 
focusing on 50–200 informative markers (relative to the classes un
der study) instead of thousands [26]. However, Heinzel et al. [27] 
showed that marker quality criteria extend beyond allele frequency 
differences. Additionally, forensic contexts often face challenges with 
missing data [7].
Cheung et al. [7] compared several classification approaches, in
cluding Snipper [9], multinomial logistic regression [10], and the Ge
netic Distance Algorithm [11]. They also analyzed the Admixture 
Model, concluding that ‘‘STRUCTURE was the most accurate classifier 
for both complete and partial genotypes in non-admixed individuals across 
four reference populations, including populations with suspected admixture’’.
Received 31 January 2025; Received in revised form 26 March 2025; Accepted 28 April 2025
Available online 13 June 2025 
1872-4973/© 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ). 
C.S. Heinzel et al.
Forensic Science International: Genetics 79 (2025) 103290 
Fig. 1. (A) Overview of the number of individuals per population in case (1), i.e. for continental classification. In this case, there are a total of 4342 individuals. (B) Overview 
of the number of individuals per population for case (2), i.e. intra-European classification. In this case, there are a total of 635 individuals. Total sample sizes for all populations 
can be found in Tables S1 and S2 in the Supplement.
This study is complemented by a comparison to XGboost and PLS-DA 
in [14], who conclude that such machine learning tools have a higher 
accuracy when compared to the above methods. We complement this 
study by using TabPFN, a new foundational model which was shown to 
outperform existing models for tabular data [28].
Ideally, a classifier predicts the population of every individual with 
100% ancestry of one population correctly with prediction probability 
1. Additionally, according to Cheung et al. [7], for admixed individuals 
the prediction probability should be equal to the ancestry proportions 
of the genome. Several measures of non-optimal classifiers, such as the 
misclassification rate, the logloss (sum of logs of misclassification prob
abilities), accuracy (fraction of correctly classified samples), the ROC 
AUC (the area under a curve of false positives against true positives) 
are among the most common statistics.
Machine learning has already seen applications in forensic science, 
including haplogroup prediction from Y-STR data [29], estimation of 
the number of contributors to a trace [30], and prediction of the age 
or post-mortem interval [31]. While machine learning methods have 
been employed to predict physical appearance from DNA [32] as well 
as BGA [24], the application of such methods have been criticized for 
their lack of fit to the specific needs in the forensic field [31]. Our work 
aims to address this gap by applying state-of-the-art machine learning 
techniques to the BGA classification problem.
2. Material and methods
2.1. Choice of the data and the marker set
There are different widely used marker sets in forensic genet
ics, (e.g. 9,22,23,33). Throughout our study, we adopt the dataset 
from [23] (their Supplementary Table S1 A), which uses the markers 
from the VISAGE Enhanced tool [33].  On the one hand, the choice of 
the markers influences the output of the classifiers. On the other hand, 
‘‘most of the BGA SNPs [in the VISAGE Enhanced Tool] are already 
well established for forensic use’’ [23]. So, we choose the VISAGE 
Enhanced tool, since this is the latest widely used marker set for 
ancestry prediction. This set of markers has 104 autosomal markers, 29 
of which are multiallelic. This data set consists of 2504 individuals from 
the 1000 genomes project [34], 929 HGDP-CEPH samples [35,36], 137 
samples from the Middle East [37], 130 samples from the SGDP [38], 
and 402 samples from the Estonian Biocentre human genome diversity 
panel [39].
We consider two cases:
(1) Continental level combining ‘AFRICAN’, ‘EAST AFRICAN’ and
‘ADMIXED AFRICAN’ to Africa (AFR), combining ‘MIDDLE EAST’
and ‘(SANGER) MIDDLE EAST’ to Middle East (ME), combining
‘EUROPEAN’ and ‘ROMA’ to Europe (EUR), East Asia (EAS), 
South-East Asia (SAS), Oceania (OCE), ‘CENTRAL ASIAN’ (CAS),
‘NORTH AFRICAN’ (NAF) and combining ‘ADMIXED AMERICAN’
and ‘AMERICAN’ to Admixed America (AMR). This makes a total 
of nine classes.
(2) Intra-European level, only considering the EUR classes from above 
with more than 20 individuals, using (CEPH) with N & W Eu
ropean ancestry’ (CEU), ‘Finnish in Finland’ (FIN), ‘Toscani in 
Italia’ (TSI), ‘Italy - Sardinian’ (SAR), ‘British in England and 
Scotland’ (GBR), ‘Iberian population in Spain’ (IBS), ‘Toscani in 
Italia’, ‘Russia - Russian’ (RUS), ‘France - French Basque’ (BAS),
‘France - French’ (FRA), and ‘Turkey’ (TUR) as classes. These are 
a total of ten classes. We did not combine BAS and FRA as [40] 
mentioned that there are big differences between basques and 
people from France. Similarly, we considered TSI and SAR as 
different populations as there genetic distance is significant [41].
The number of individuals is shown for both cases in Fig.  1, where we 
see that some classes are under-represented in the dataset. 
2.2. Classification
For classification, we compare the following methods: (i) Snipper, 
which is a version of a naive Bayes classifier [42]; (ii) the Admixture 
Model (AM) (in its supervised setting), which assumes that every allele 
in individual 𝑖 has a chance of 𝑞𝑖𝑘
 to come from population 𝑘 [12,13], 
and individual 𝑖 is classified into population 𝑘 if 𝑞𝑖𝑘
 is maximal; (iii) 
PLS-DA which has already been used by [14] in the forensic context; 
and (iv) TabPFN, a novel foundational model for tabular data (such as 
genetic data) based on a neural network [28], (v) XGBoost  [43] and 
(vi) Random Forest [44]. Note that the model-based methods (i) and (ii) 
treat markers as independent, while the machine-learning (black-box) 
methods (iii), (iv), (v) and (vi) incorporate joint segregation of markers.
Other classifiers, such as gradient-boosted decision trees [43,45,46], 
or logistic regression [47], would as well be applicable to genetic 
(i.e., tabular) data. However, Hollmann et al. [28] have shown that 
TabPFN dominantly outperforms other classifiers, including those men