Prescreening Collagen for Archaeological Bone

Collagen is essential in archaeology for reconstructing past diets, environments, and interactions, but it often degrades through diagenesis. Traditional prescreening methods require destructive sampling and lab access. A recent study published in the Journal of Archaeological Science advances near-infrared (NIR) spectroscopy as a rapid, non-destructive, and portable alternative.^1,2 Using partial least squares regression (PLSR) and random forest (RF) models, researchers accurately predicted collagen preservation, outperforming %N methods and enabling efficient, high-throughput screening while minimizing unnecessary damage to valuable archaeological samples.^1,2

In this brief Q&A interview, Christina Ryder, who is a postdoctoral researcher at Texas A&M University and the lead author of this study, discusses her team’s findings.

What were the key technical advances that allowed NIR to assess subsurface collagen more effectively than FT-IR or Raman approaches?

In archeology, two techniques are primarily used, which are near-infrared (NIR), Fourier transform infrared (FT-IR), and Raman spectroscopy. The key technical advantage of NIR spectroscopy, relative to the other techniques for assessing collagen and archeological bone, is that it allows for substantially greater penetration depth into a scattering, heterogeneous, and biogenic material, such as bone. MIR and Raman techniques only probe the immediate surface, so typically in the order of microns. As a result, both of these approaches generally require powdered subsamples or freshly exposed cross sections. We do see studies that apply attenuated total reflectance FT-IR (ATR-FTIR) to whole bone, but they only have marginal success when they are applied to the outer surface and not a freshly exposed cross section. In contrast, the lower molar absorptivity of NIR allows the light to penetrate several millimeters into this highly scattering material using a strong NIR absorber as an optical barrier. I documented signal contribution from dust exceeding five millimeters within full intact bone. This shows that NIR is interrogating the bulk tissue, rather than just what's only on the surface. When you're working in an archeological or paleontological space, a non-destructive prescreening method is critical. So even if something's minimally destructive or minimally invasive, even if you need a small amount of powder, that's oftentimes a no-go for curators or for permits to get things out of certain countries and whatnot.

In your study, your team trained both Partial Least Squares Regression (PLSR) and Random Forest (RF) models on bones with known collagen yields; how did the two modeling strategies compare in terms of accuracy, robustness, and interpretability for archaeological applications?

In our paper, we presented two predictive modeling strategies: partial least square regression (PLSR) and random forest (RF) models. And then, we applied them in two different ways. We applied them to a classic validation set, which was 40 samples randomly extracted from our calibration validation set, and then to an external archeological data set. A real-world application of this technique in the clean set, which is clean, meaning free from consolidants or adhesives, which are very common. In the archeological and paleontological world, RF outperforms the PLSR models.

For the two RF models we used, one being the entire NIR range from 780 to 2500 nm, and then the second being the restricted range from 2030 to 2060 nm. I think our peak root mean square error (RMSE) was approximately 0.7, and the RF models correctly classified 39 of the 40 validation samples. So, it correctly identified samples that are suitable for radiocarbon dating, whereas the PLSR model had a higher root mean square prediction of about 1.62, and it correctly classified only 34 of the 40 samples. So, in a controlled setting, the RF model clearly had the lower error.

However, as when we applied these models to an external data set, which was an archeological collection from a late Pleistocene Neanderthal locality, the picture shifted so the restricted PLSR model correctly classified 18 of the 19 Zaria samples, which is approximately a 95% accuracy. The RF model, on the other hand, had a correction classification and success rate of about 58% and then the restricted RF improved upon the original RF model. The classification success improved to about 89%, but it still did not exceed the PLSR model in classification reliability, and this difference is likely reflected by overfitting in the high-dimensional RF model.

What we think is happening is that there are spectral regions sensitive to consolidants that the RF model is picking up on that we can restrict in a more controlled setting with the PLSR. The PLSR model focused on that collagen specific absorption band, which improved the prediction and stability when applied to this archeologically complex material. So overall, we're still continuing to improve our RF and our PLSR models as I collect more data and build increased robustness and variation in our models. But right now, the model I use almost every single day is the PLSR model, which we'll see if that continues in the future. But I just find it to be very reliable and obviously very supervised.

References

1. Ryder, C.; Celis, G.; Devièse, T. et al. Refining Near-infrared Spectroscopy for Collagen Quantification: A New Predictive Model for Archaeological Bone. J. Arch. Sci. 2026, 185, 106448. DOI: 10.1016/j.jas.2025.106448
2. Wetzel, W.; Spectroscopy Staff. Collagen Preservation in Archaeological Bone Using NIR Spectroscopy. Spectroscopy. Available at: https://www.spectroscopyonline.com/view/collagen-preservation-in-archaeological-bone-using-nir-spectroscopy (accessed 2026-03-31).