Publication Date
Spring 2026
Degree Type
Master's Project
Degree Name
Master of Science in Computer Science (MSCS)
Department
Computer Science
First Advisor
Teng Moh
Second Advisor
Melody Moh
Third Advisor
Mark Barash
Keywords
Forensic DNA phenotyping, Genome Wide Association Study(GWAS), Neural Network, Single Nucleotide Polymorphism, Transformer
Abstract
Forensic DNA profiling is a powerful identification tool, but its utility in criminal casework is limited when no comparison profiles are available for direct matching. Forensic molecular phenotyping addresses this gap by inferring externally visible characteristics (EVCs) from biological evidence, offering investigative leads when traditional database searches fail. In this study, we present a pigmentation component of DNA2Face, a set of machine learning models that classify discrete pigmentation traits directly from single nucleotide polymorphisms (SNPs) selected from published association findings and the theoretical component of a DNA to 3D point cloud model of the human face. For classifying discrete pigmentation traits, We evaluate two modeling strategies: (1) compact linear neural networks trained on allele-dosage and one-hot encodings, and (2) an attention-based Transformer encoder that treats SNP alleles as token-like categorical inputs to learn cross-locus dependencies. Using a dataset of 277 individuals with genotype data and recorded phenotypes, we assess performance for eye, hair, and skin color classification under multiple SNP selection strategies, including candidate SNP panels motivated by prior GWAS and statistical filtering. Results show that small feed-forward neural networks provide the most stable performance under limited sample size, while attention-based models achieve above- chance discrimination, but are more sensitive to feature ordering and data constraints. Oversampling methods improve minority-class performance in more imbalanced traits, particularly hair color. Overall, DNA2Face illustrates both the promise and the current limitations of deep learning approaches for forensic phenotyping in small, local datasets, and motivates larger, ancestry-diverse training cohorts and biologically informed SNP organization for future work.
Recommended Citation
Kuei, Andy, "Classifying and Generating Externally Visible Characteristics from Human DNA using Transformer Architectures and Neural Networks" (2026). Master's Projects. 1828.
DOI: https://doi.org/10.31979/etd.rmn8-pyht
https://scholarworks.sjsu.edu/etd_projects/1828
