mHolmes improves cross anatomical cadaveric microbiome forecasting for postmortem interval estimation

Forensic microbiology leverages postmortem microbiome succession as a promising biomarker for estimating the postmortem interval (PMI). However, current methods are constrained by sparse sampling (typically 3–5 time points) and limited cross-anatomical generalizability, leading to imprecise PMI estimates with errors often exceeding ±3 days, particularly in cases of dismembered remains. To address these limitations, we developed mHolmes, a Transformer-based transfer learning framework for forecasting cadaveric microbiome dynamics. Trained on daily longitudinal data from 34 cadavers over 21 days, mHolmes achieves daily predictions of microbial dynamics, with reduced error (MAE < 2 days) in cross-anatomical forecasting tasks (e.g., hip to face). Shapley Additive exPlanations (SHAP) analysis supports interpretability by identifying seven bacterial classes as candidate PMI-associated features. This study supports mHolmes as a robust forecasting framework that addresses key limitations in sparse and cross anatomical microbiome data, improves PMI estimation from incomplete observations, and has potential forensic applications in body part matching and higher resolution timeline reconstruction. This study presents mHolmes, a transfer-learning framework that forecasts postmortem microbiome changes across body sites, improving postmortem interval estimation and highlighting microbial features with forensic potential.