Enhancing late postmortem interval prediction: a pilot study integrating proteomics and machine learning to distinguish human bone remains over 15 years

Abstract

Background: Determining the postmortem interval (PMI) accurately remains a significant challenge in forensic sciences, especially for intervals greater than 5 years (late PMI). Traditional methods often fail due to the extensive degradation of soft tissues, necessitating reliance on bone material examinations. The precision in estimating PMIs diminishes with time, particularly for intervals between 1 and 5 years, dropping to about 50% accuracy. This study aims to address this issue by identifying key protein biomarkers through proteomics and machine learning, ultimately enhancing the accuracy of PMI estimation for intervals exceeding 15 years.

Methods: Proteomic analysis was conducted using LC-MS/MS on skeletal remains, specifically focusing on the tibia and ribs. Protein identification was performed using two strategies: a tryptic-specific search and a semitryptic search, the latter being particularly beneficial in cases of natural protein degradation. The Random Forest algorithm was used to model protein abundance data, enabling the prediction of PMI. A thorough screening process, combining importance scores and SHAP values, was employed to identify the most informative proteins for model's training and accuracy.

Results: A minimal set of three biomarkers-K1C13, PGS1, and CO3A1-was identified, significantly improving the prediction accuracy between PMIs of 15 and 20 years. The model, based on protein abundance data from semitryptic peptides in tibia samples, achieved sustained 100% accuracy across 100 iterations. In contrast, non-supervised methods like PCA and MCA did not yield comparable results. Additionally, the use of semitryptic peptides outperformed tryptic peptides, particularly in tibia proteomes, suggesting their potential reliability in late PMI prediction.

Conclusions: Despite limitations such as sample size and PMI range, this study demonstrates the feasibility of combining proteomics and machine learning for accurate late PMI predictions. Future research should focus on broader PMI ranges and various bone types to further refine and standardize forensic proteomic methodologies for PMI estimation.

Keywords: Biomarker discovery; Forensic science; Machine learning; Postmortem interval; Proteomics.

Fig. 1

Characterization of proteomes in bone remains. A Protein identification using tryptic search. B Protein identification using semitryptic search. C Comparison of proteomes by PMI and bone segments for proteins identified through tryptic search. D Comparison of proteomes by PMI and bone segments for proteins identified through semitryptic search. E In silico characterization of molecular weight distribution between PMI < 1 year and PMI > 15 years. F Percentage of acidic amino acids in proteins from PMI < 1 year and PMI > 15 years. G Percentage of basic amino acids in proteins from PMI < 1 year and PMI > 15 years

Fig. 2

Global visualization of proteomes through principal component analysis. A Three-dimensional visualization of the first three principal components for proteins identified through tryptic search. B Three-dimensional visualization of the first three principal components for proteins identified through semitryptic search. C Variable loadings for the tryptic search. D Variable loadings for the semitryptic search

Fig. 3

Model optimization via variable screening for tibia late PMI using semitryptic proteins. A Initial model performance using the identified hyperparameters and the complete set of representative proteins. B Importance score for full set of proteins in model A. C Model performance after selecting 11 proteins with an importance score greater than 4%. Proteins were then iteratively removed based on their SHAP values. D SHAP value distributions for the most important proteins contributing to the classification of PMI 15. E SHAP value distributions for the most important proteins contributing to the classification of PMI 20. F Final model performance after reducing the set to three key proteins: PGS1, K1C13, and CO3A1

Fig. 4

Potential protein biomarkers for distinguishing Early and Late PMI. A Prediction performance using proteins K1C13, CO3A1, and PGS1 to classify PMI < 1 year, PMI 15 years, and PMI 20 years. B Importance score for the model in A. C–E Distribution of SHAP values for the three protein biomarkers across 100 iterations for rib samples with PMI < 1 year (C), tibia samples with PMI 15 years (D), and tibia samples with PMI 20 years (E). F–H Semitryptic protein abundance (intensity) for K1C13, CO3A1, and PGS1 across 100 iterations for rib samples with PMI < 1 year (F), tibia samples with PMI 15 years (G), and tibia samples with PMI 20 years (H)