Cross-clade limits of melanosome-based color prediction: a cautionary machine learning study
Abstract
Reconstructing dinosaur coloration from fossilized melanosomes has revolutionized paleontology, but most predictive models are trained exclusively on modern birds. Whether these models generalize across clade boundaries to non-theropod dinosaurs remains a critical, untested question. We trained Quadratic Discriminant Analysis (QDA), Support Vector Machine (SVM), and Random Forest (RF) classifiers on 342 modern bird melanosome measurements spanning five color categories (black, brown, gray, iridescent, and yellow) and validated them on eight theropod fossils with known coloration. QDA achieved 75% validation accuracy (6/8 correct), SVM 62.5% (5/8), and RF 12.5% (1/8) this reflects the destabilizing effect of a severely under-sampled minority class. When applied to four non-theropod specimens (Psittacosaurus and juvenile Diplodocus), all three models collapsed to predicting "yellow", with SVM and RF confidence probabilities between 0.46 and 0.70. This uniform collapse is a synthetic artifact: the yellow class contained only three original training samples, and Synthetic Minority Oversampling Technique (SMOTE) inflated it to parity with the largest class (~173 samples), creating a spurious decision region that intersects the fossil morphospace. Our results demonstrate that aspect-ratio-based classifiers are not only limited within theropods but are actively misleading across clade boundaries when minority-class oversampling is applied naively. We establish a cautionary, reproducible benchmark for deep-time color reconstruction and identify minority-class oversampling as a previously unrecognized source of systematic error in paleocolor machine learning.
Related
- Leakage-controlled benchmarking reveals generalization limits of deep learning for protein-ligand binding affinity prediction
- Regional climate variability and rice yields prediction in Bangladesh using machine learning and deep learning models
- Integrating Multisource Target Evidence with Machine-Learning Potency Prediction for Alzheimer's Disease Drug Discovery
- An Explainable Hybrid Machine Learning Approach Based on MPSO-XGBoost for Pest Density Prediction in Almond Orchards
- Species Distribution, Antifungal Susceptibility, and Machine Learning-Based Prediction of Optimal Therapies Against Candida Species
BibTeX
@misc{chowdhury2026crossclade,
title = {{Cross-clade limits of melanosome-based color prediction: a cautionary machine learning study}},
author = {Md Ahnaf Usaid Chowdhury and Tanvir Shahriar},
howpublished = {PaleorXiv},
year = {2026},
month = oct,
doi = {10.31233/osf.io/69js3_v1},
url = {https://osf.io/preprints/paleorxiv/69js3_v1/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Cross-clade limits of melanosome-based color prediction: a cautionary machine learning study
AU - Md Ahnaf Usaid Chowdhury
AU - Tanvir Shahriar
PY - 2026
DA - 2026/10/22
PB - PaleorXiv
DO - 10.31233/osf.io/69js3_v1
UR - https://osf.io/preprints/paleorxiv/69js3_v1/
AB - Reconstructing dinosaur coloration from fossilized melanosomes has revolutionized paleontology, but most predictive models are trained exclusively on modern birds. Whether these models generalize across clade boundaries to non-theropod dinosaurs remains a critical, untested question. We trained Quadratic Discriminant Analysis (QDA), Support Vector Machine (SVM), and Random Forest (RF) classifiers on 342 modern bird melanosome measurements spanning five color categories (black, brown, gray, iridescent, and yellow) and validated them on eight theropod fossils with known coloration. QDA achieved 75% validation accuracy (6/8 correct), SVM 62.5% (5/8), and RF 12.5% (1/8) this reflects the destabilizing effect of a severely under-sampled minority class. When applied to four non-theropod specimens (Psittacosaurus and juvenile Diplodocus), all three models collapsed to predicting "yellow", with SVM and RF confidence probabilities between 0.46 and 0.70. This uniform collapse is a synthetic artifact: the yellow class contained only three original training samples, and Synthetic Minority Oversampling Technique (SMOTE) inflated it to parity with the largest class (~173 samples), creating a spurious decision region that intersects the fossil morphospace. Our results demonstrate that aspect-ratio-based classifiers are not only limited within theropods but are actively misleading across clade boundaries when minority-class oversampling is applied naively. We establish a cautionary, reproducible benchmark for deep-time color reconstruction and identify minority-class oversampling as a previously unrecognized source of systematic error in paleocolor machine learning.
ER -