Sex Classification and Identity Verification from GC×GC of Human Scent Samples

Importance of the Topic

Human scent carries a rich signature of volatile organic compounds (VOCs) that reflect physiological and metabolic states. Advanced separation and detection techniques such as comprehensive two-dimensional gas chromatography coupled to time-of-flight mass spectrometry (GC×GC–ToF–MS) enable high-resolution profiling of these complex mixtures. Automated analysis of scent profiles can transform forensic identification, non-invasive diagnostics, security screening and personalized health monitoring.

Objectives and Overview of the Study

This work presents an end-to-end, open-source pipeline for two key biometric tasks using raw GC×GC–ToF–MS data: biological sex classification and individual identity verification. The authors aim to eliminate manual intervention by developing an automated alignment procedure and leveraging deep learning on tensorial spectral data interpreted as images. The approach is validated on a dataset of 2 528 human scent samples from 252 individuals.

Methodology and Instrumentation

• Sample Data: Raw GC×GC–ToF–MS outputs are treated as three-dimensional tensors (mass spectral dimension × first-dimension retention time × second-dimension retention time), approximately 4 GB per sample.
• Data Variants: Two representations are evaluated—FULL (complete spectral tensor) and SAS (sum along the mass spectral axis reducing dimensionality).
• Alignment: An unsupervised registration step detects 24 reference compounds and applies a piecewise linear transformation to align all samples to a common chromatographic frame, requiring zero manual annotation.
• Machine Learning: Convolutional neural networks (FCN architecture) process the registered tensors/images. Models are trained with gradient-descent optimization to minimize task-specific loss.
• Computational Resources: All experiments were executed on the LUMI supercomputer, consuming approximately 120 000 GPU hours (estimated cost ∼ 250 000 USD on rented hardware).

Main Results and Discussion

• Sex Classification: On an 80:20 training/validation split with 10-fold cross-validation, the FCN achieved high accuracy distinguishing male from female scent profiles using both FULL and SAS representations.
• Identity Verification: Pairwise comparison of samples by the same individual versus different individuals yielded robust performance, with area under the ROC curve (AUC) demonstrating strong discrimination capability.
• Computational Efficiency: Despite the large data volume, the automated alignment and neural models scale effectively without human intervention and maintain reproducibility across runs.

Benefits and Practical Applications

The proposed pipeline offers several practical advantages:

• Zero Human Labor: Fully automated alignment and feature extraction eliminate subjective bias and reduce labor costs.
• Reproducibility: Standardized processing ensures consistent results across laboratories and time points.
• Open-Source Implementation: Python code and pretrained models will be publicly available to accelerate adoption in forensic, clinical and security contexts.

Future Trends and Potential Applications

Emerging directions include integration with portable GC×GC–ToF–MS platforms for field deployment, multimodal data fusion combining scent profiles with other biometric signals, real-time monitoring of metabolic biomarkers, and privacy-preserving neural architectures to protect individual data. Continuous improvements in deep learning and alignment algorithms may further enhance sensitivity and specificity.

Conclusion

This study demonstrates that raw GC×GC–ToF–MS data, when interpreted as images and processed through automated registration and convolutional neural networks, can accurately classify biological sex and verify individual identity from human scent. The open-source, zero-human-intervention pipeline paves the way for scalable, reproducible biometric analysis using complex chromatographic datasets.

Reference

No explicit literature references were provided in the original text.