This work sits alongside research that uses molecular data for forensic identification, from analyzing blood from a single molecule to predicting diseases from routine tests. Neighboring studies focus on individual biomarkers or narrow panels, whereas this study covers five fluid types at once and, for the first time, systematically compares three fundamentally different proteomic strategies on a single cohort. It fills the gap between rapid preliminary tests and expensive confirmatory methods, offering a scalable approach for mixed samples.
Blood, saliva, semen, urine, and vaginal fluid samples were collected from 20 volunteers on sterile swabs. Proteins were extracted using the SP3 protocol, digested with trypsin, and analyzed by LC-MS/MS on a Q Exactive mass spectrometer in DDA mode. For the first test, 526 peptides specific to each fluid were selected with 80% sensitivity and 90% specificity. The second test used 250 peptide intensity ratios selected by ANOVA. The third test was built on a Classifier Chain Random Forest model trained on 1110 features selected by the ReliefF algorithm.
The specific peptide test achieved 100% accuracy for blood and saliva and over 90% for the other fluids, but showed false positives in mixtures, especially for urine. The peptide ratio test demonstrated accuracy above 90% for all classes, with mixtures recognized better than pure samples (75.6% versus 68.0%). The machine learning model performed best: subset accuracy of 95.9% and 100% for pure fluids. The three methods were found to use largely non-overlapping feature sets: only 15 peptides are common to all approaches, confirming their complementarity.
Unlike neighboring studies that focus on a single class of molecules or a single fluid type, this research directly compares three proteomic strategies on a unified dataset and shows that combining biomarkers with machine learning outperforms traditional approaches. Importantly, the machine learning model does not rely on known specific proteins but discovers its own patterns, which could reveal new diagnostic markers. This is a first step toward standardizing proteomic analysis in forensic practice.
The topic could evolve toward analyzing real crime scene samples, including degraded stains on various substrates and mixtures with unknown proportions. Expanding the taxonomic scope is promising, for example, incorporating microbiota or animal proteins to determine the source of a trace. User-friendly software is also likely to emerge, allowing experts without specialized training to interpret proteomic data.
The method will impact forensic examination, criminalistics, and biomedical diagnostics, where accurate identification of biological fluids in mixed or contaminated samples is required.
Next steps include validation on real samples with known composition and on simulated stains under realistic conditions, such as degradation and varying component ratios.
This work directly addresses an unresolved problem in forensic proteomics: how to reliably identify biological fluids in mixtures where traditional immunochromatographic tests give false results due to cross-reactivity. Another issue is the high variability of protein composition between individuals and the influence of sample storage conditions, which limits the transferability of models trained on laboratory data.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors developed and compared three LC-MS/MS-based proteomic methods for identifying five biological fluids: a panel of specific peptides, peptide ratio analysis, and a random forest model. On a cohort of 20 volunteers, the machine learning model achieved 95.9% accuracy for the subset and 100% for pure fluids, outperforming traditional approaches. The methods were shown to use nearly non-overlapping feature sets, confirming their complementarity.
Where its strength is
A systematic comparison of three fundamentally different strategies on a unified dataset, demonstrating that machine learning finds its own patterns without relying on pre-known biomarkers.
Where it could go next
Worth trying
Why this may matter
Standardizing proteomic analysis of biological fluids could substantially enhance the evidentiary value of forensic examinations, especially in cases with mixed stains.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.56, against a usual 0.64. It looks like few have drilled nearby.
Among the 526 specific peptides are fragments of proteins already used in routine immunochromatographic tests: hemoglobin, α-amylase, prostate-specific antigen, semenogelin, and uromodulin. But the proteomic approach expands the panel to hundreds of peptides, greatly increasing specificity.
In the TV series 'CSI: Crime Scene Investigation', experts often determine the type of biological fluid with a single quick test. This work brings reality closer to fiction, showing that mass spectrometry and machine learning can provide an equally fast and far more reliable answer, even for mixtures.