52/100

Moderate evidence

Cite
More
EmailLinkedInX

Soil microbes traced to their location with up to 97.9% accuracy in forensic study

Study title: Supervised machine learning enables forensic microbiome source attribution in soil and touch samples with evidence-type-specific relic DNA effects

Evidence type
Benchmark or lab evaluation
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: Supervised machine learning enables forensic microbiome source attribution in soil and touch samples with evidence-type-specific relic DNA effects

A study in the Journal of Forensic Sciences tested whether supervised machine learning can link microbial trace evidence to its source. XGBoost classifiers were applied to 16S rRNA amplicon profiles from mock soil traces from four locations and from human touch traces transferred under controlled conditions.

Read the study abstract

Microbial trace evidence is increasingly proposed for forensic source attribution, yet most studies lack performance benchmarks across evidence types and rarely test robustness across sample preprocessing conditions. Here, we evaluated supervised machine learning models for attributing (i) mock soil trace evidence to four source locations and (ii) human touch trace evidence to the correct donor using paired samples collected under controlled transfer conditions. 16S rRNA amplicon profiles were used to assess mic…

Supervised machine learning enables forensic microbiome source attribution in soil and touch samples with evidence-type-specific relic DNA effects
"Chih Cell Block of Chiayi Old Prison (Taiwan)" by Mearchan, CC BY-SA 4.0.

The gain

In a forensic lab study, machine learning classifiers trained on 16S rRNA microbiome profiles matched mock soil trace evidence to one of four source locations with accuracies above 84%, reaching 97.9%.

In brief

For soil, the classifiers placed samples at the right location with accuracies above 84%, reaching 97.9%. Location was the main driver of microbial community structure, and the signal held up across enzymatic treatment conditions.

Touch evidence was harder. Donor attribution worked but was more sensitive to conditions: the best top-3 accuracy was 86% for control-trained models on control PVC sinks, and it fell to 68% for sinks depleted of relic DNA. The authors say the results do not support routine relic DNA depletion, particularly for touch evidence.

Main points

  1. Models attributed mock soil trace evidence to one of four source locations with accuracies above 84%, up to 97.9%.
  2. Microbial community structure in soil samples was strongly driven by location; relic DNA depletion and sample type had smaller, location-specific effects.
  3. Donor attribution from human touch traces was feasible but weaker and more sensitive to conditions than soil attribution.
  4. The best touch result was 86% top-3 accuracy, for control-trained models predicting control PVC sinks.
  5. Top-3 accuracy dropped to 68% for relic-DNA-depleted sinks under the same training scheme.
  6. The authors conclude the results do not support routine relic DNA depletion, especially for touch evidence.

The full story

The researchers used paired samples collected under controlled transfer conditions. They profiled microbial communities with 16S rRNA amplicon sequencing and trained XGBoost classifiers under within-condition and cross-condition train/test schemes.

They also tested how removing extracellular (relic) DNA changed microbiome diversity and attribution performance. In soil, this treatment had only small, location-specific effects, so attribution stayed strong across preprocessing conditions.

For touch evidence, which is a lower-complexity microbial signal, depleting relic DNA reduced attribution performance. The authors present the work as a benchmark across evidence types and preprocessing conditions, not as proof of casework readiness.

Sources

1 on file
  1. Journal of Forensic Sciencesdoi:10.1111/1556-4029.70504
    Peer-reviewed2026-11-16

Talk it over

Three questions to start on, two more if you have time
Classifiers matched mock soil traces to one of four locations with accuracies above 84%, reaching 97.9%. What would you want to see before a court relied on a match like that?

Start with a question

Soil microbes traced to their location with up to 97.9% accuracy in forensic studyhttps://truvace.com/item/supervised-machine-learning-enables-forensic-microbiome-source-attribution-in-so

  1. Classifiers matched mock soil traces to one of four locations with accuracies above 84%, reaching 97.9%. What would you want to see before a court relied on a match like that?

    Answer in the thread
  2. The study used mock soil traces from four locations and controlled touch transfers. How might real casework samples, with more locations and messier conditions, change the results?

    Answer in the thread
  3. Touch attribution reached 86% top-3 accuracy for control-trained models on control PVC sinks and fell to 68% for sinks depleted of relic DNA. How do you weigh strong soil results against weaker touch results?

    Answer in the thread
If you have more time
  • The authors say the results do not support routine relic DNA depletion, especially for touch evidence. Who in a forensic lab should decide whether to change its sample preparation, and on what evidence?

    Answer in the thread
  • The authors present the work as a benchmark, not proof of casework readiness. What would a test on real cases need to include to show readiness?

    Answer in the thread

What readers say