NIR Spectroscopy for Collagen Quantification in Archaeological Bone

Reproduction of the analytical workflow from:

Ryder, C. et al. (2026). Refining near-infrared spectroscopy for collagen quantification: A new predictive model for archaeological bone. Journal of Archaeological Science, 185, 106448.

The notebook implements the full pipeline — from raw reflectance spectra to collagen yield prediction — using Python open-source libraries.

---

Contents

File	Description	Open in Colab
NIR_Collagen_Prediction.ipynb	Main analysis notebook
1-s2.0-S0305440325002973-mmc2.csv	Supplementary data from the paper (176 samples, 2151 wavelengths)
plsr_2045.json	Serialized PLSR pipeline — IntensityConversion → SG → RangeCut(2030–2060 nm) → PLSRegression
rf_2045.json	Serialized Random Forest pipeline — IntensityConversion → SG → RangeCut(2030–2060 nm) → RandomForestRegressor

---

Dataset

The CSV file contains spectral data for 176 bone samples:

• 140 Reference samples — used for model calibration and validation. Already K-means balanced from an original set of 319 samples (see paper Section 2.3).
• 36 Zafarraya samples — independent external validation set from Zafarraya Cave, Spain.

Columns:

• Sample ID, Country of Origin, Extraction Technique (ORAU / MPI)
• Collagen Yield (%)
• Reflectance at 350–2500 nm (2151 variables)

---

Notebook Overview

The notebook reproduces the complete workflow in 13 sections:

Section	Content
1–2	Imports and data loading
3	Exploratory data analysis (yield distribution, raw spectra)
4	Spectral preprocessing: reflectance → pseudo-absorbance → Savitzky-Golay 2nd derivative
5	PCA for outlier detection (full range and NIR range)
6	Stratified calibration / validation split (100 / 40 samples)
7	Helper functions
8	PLSR across 11 wavelength ranges (reproduces Table 5)
9	Random Forest regression — full NIR and 2030–2060 nm (reproduces Sections 3.1–3.2)
10	Variable importance plots
11	Combined LOO-CV models on all 140 samples
12	Zafarraya external validation with PVA consolidant detection (reproduces Tables 6–8)
13	Model persistence to JSON using OpenModels

The preferred model (2030–2060 nm, 1 PLSR factor) targets the 2045 nm absorption feature associated with the 2nd overtone of C=O stretching and N-H stretching in collagen.

---

Dependencies

numpy
pandas
matplotlib
scikit-learn
chemotools
openmodels

Install with:

pip install -r requirements.txt

or manually:

pip install numpy pandas matplotlib scikit-learn chemotools openmodels

---

Usage

1. Clone the repository.
2. Install dependencies (see above).
3. Open NIR_Collagen_Prediction.ipynb in Jupyter Lab or Jupyter Notebook.
4. Run all cells in order (Kernel → Restart & Run All).

The notebook is self-contained: all preprocessing, modelling, and evaluation steps run sequentially without external configuration.

---

Prediction API

The serialized models are also available as a REST API, so you can predict collagen yield directly from your own spectra without running the notebook.

Try it online — no installation needed:

https://collagen-prediction.fastapicloud.dev

Run it locally — download the pre-built executable for your platform from the Releases page and run it:

# macOS / Linux
chmod +x nir-collagen-api-macos   # or nir-collagen-api-linux
./nir-collagen-api-macos

# Windows
nir-collagen-api-windows.exe

Then open http://localhost:8000/docs to explore the endpoints interactively.

The API accepts spectra as JSON (single or batch) or as a CSV file upload, and supports both pseudo-absorbance and raw reflectance inputs.

---

Key Results Reproduced

PLSR model performance across wavelength ranges (100-sample calibration / 40-sample validation split):

Range	Factors	Val R²	Val RMSE
2030–2060 nm (preferred)	1	0.876	1.78%
2030–2060 + 2244–2300 nm	3	0.890	1.67%
2000–2300 nm	3	0.883	1.73%
780–2500 nm (full NIR)	3	0.862	1.88%

Combined models — Leave-One-Out CV (all 140 reference samples):

Model	LOO-CV R²	LOO-CV RMSE
PLSR 2030–2060 nm	0.883	1.62%
RF 2030–2060 nm	0.894	1.54%
RF 780–2500 nm	0.919	1.35%

• The 2030–2060 nm range (1 PLSR factor, 31 features) delivers the best balance between parsimony and predictive accuracy — the preferred model for deployment.
• Restricting the spectral range to 2030–2060 nm avoids PVA consolidant absorption bands at 2135, 2250, and 2296 nm, making the model robust for museum collections.
• Random Forest (780–2500 nm) achieves the lowest LOO-CV RMSE (1.35%) but requires the full spectral range and is more sensitive to consolidant contamination.