Simulation-Based Assessment of Cancer-Associated cfDNA Methylation and miRNA Profiles for Forensic Molecular Inference: A Proof-of-Concept Machine-Learning Study

Technical Summary: Simulation-Based Assessment of Cancer-Associated cfDNA Methylation and miRNA Profiles for Forensic Molecular Inference

Problem Statement
Forensic DNA phenotyping (FDP) is expanding beyond externally visible traits (e.g., pigmentation, ancestry) toward probabilistic molecular inference. However, inferring disease-associated states, such as cancer, presents unique challenges compared to physical trait prediction. Disease inference involves sensitive health-related information, raising significant ethical, legal, and evidentiary concerns regarding privacy and proportionality. Furthermore, translating clinical liquid biopsy biomarkers (circulating cell-free DNA [cfDNA] methylation and microRNA [miRNA]) to forensic contexts is complicated by environmental degradation, contamination, low-template inputs, and mixed-source traces. There is a lack of controlled computational benchmarks to evaluate whether combined cfDNA methylation and miRNA profiles can theoretically distinguish healthy from cancer-associated molecular states within a forensic framework, without overstating clinical or operational validity.

Methodology
This study is a proof-of-concept, forensic-oriented simulation designed to evaluate the internal discriminatory capacity of combined molecular profiles. It did not involve human participants, biological sample collection, or diagnostic testing.

• Data Generation: A balanced dataset of 1,000 simulated molecular profiles was generated (500 healthy, 500 cancer-associated) using 100 predictors (50 cfDNA methylation variables and 50 miRNA expression variables). The simulation was informed by publicly available lung cancer-related resources (e.g., GEO dataset GSE157152) and peer-reviewed literature to ensure biological plausibility.
• Simulation Parameters: The model incorporated biological heterogeneity, inter-feature correlation (via shared latent factors), stochastic variation, and partial distributional overlap between classes to avoid deterministic separation. A global effect parameter of 0.56 was applied to a subset of variables to simulate class effects.
• Forensic Perturbation Modeling: To assess robustness, the study introduced stress-test scenarios simulating adverse forensic conditions:
  • Degradation: Attenuation of class-related signal (20%, 40%, 60%).
  • Contamination: Partial profile mixing (10%, 20%, 40%).
  • Analytical Noise: Inflation of stochastic variance (25%, 50%, 100%).
  • Class Overlap: Shrinkage of between-group effect sizes (30%, 50%).
• Machine Learning Approach: Five supervised algorithms were compared: Logistic Regression, Support Vector Machine (RBF kernel), Random Forest, Gradient Boosting, and Extreme Gradient Boosting. Models were evaluated using stratified five-fold cross-validation with pooled out-of-fold predictions.
• Model Selection: Gradient Boosting was retained as the primary "conservative" classifier. Although Logistic Regression and SVM achieved higher internal AUCs, Gradient Boosting was chosen to avoid framing the study with maximal, potentially over-optimistic performance metrics.

Key Results

• Baseline Performance: Under controlled simulation assumptions, the combined cfDNA methylation and miRNA feature space demonstrated strong class separation. The primary Gradient Boosting model achieved an out-of-fold Area Under the Curve (AUC) of 0.9932, with an accuracy of 0.967, precision of 0.968, recall (sensitivity) of 0.966, and an F1-score of 0.967.
• Confusion Matrix: The primary model produced 484 true negatives, 483 true positives, 16 false positives, and 17 false negatives.
• Feature Importance: The model utilized information from both molecular layers, with top-ranking features including both methylation (MET) and miRNA (miR) variables, suggesting that combined profiles retain discriminative information.
• Sensitivity Analysis: Performance was highly sensitive to adverse conditions.
  • High Degradation (60% signal attenuation): AUC dropped to 0.8352.
  • High Analytical Noise (100% variance inflation): AUC dropped to 0.8972.
  • High Class Overlap (50% effect shrinkage): AUC dropped to 0.8876.
  • High Contamination (40% mixing): AUC dropped to 0.9340.
  • These results indicate that the classification signal is not invariant to forensic stress conditions.

Significance and Claims
The authors position this work strictly as a computational benchmark and a methodological starting point, explicitly rejecting claims of clinical diagnostic validity or operational forensic utility.

• Proof-of-Concept: The study demonstrates that, within a controlled simulation, combined cfDNA methylation and miRNA profiles possess strong internal discriminability between healthy and cancer-associated states.
• Methodological Transparency: By specifying simulation parameters, comparing multiple algorithms, and conducting stress tests, the study avoids presenting a single high AUC as sufficient evidence of forensic validity. It highlights that while the signal is strong under ideal assumptions, it degrades significantly under realistic forensic constraints (degradation, noise, contamination).
• Limitations and Boundaries: The paper emphasizes that the dataset is simulated, not derived from authentic forensic traces. It does not validate cancer detection from crime-scene evidence, does not establish donor diagnosis, and does not support investigative use.
• Future Requirements: The authors conclude that any progression toward practice requires wet-lab validation using authentic forensic materials (degraded, mixed, low-template), independent external datasets, and the establishment of explicit ethical, legal, and reporting safeguards before any investigative application could be considered.

In summary, the paper provides a conservative, transparent evaluation of the theoretical potential of cancer-associated molecular inference in forensics, while rigorously delineating the gap between computational simulation and operational reality.