Machine learning ATR-FTIR spectroscopy data for the screening of collagen for ZooMS analysis and mtDNA in archaeological bone
Machine learning ATR-FTIR spectroscopy data for the screening of collagen for ZooMS analysis and mtDNA in archaeological bone
复制标题
DOI：
10.1016/j.jas.2020.105311
复制
发表时间：
2021-01-16
影响因子：
2.8
通讯作者：
Buckley, Michael
中科院分区：
文献类型：
--
作者：
Chowdhury, Manasij Pal;Choudhury, Kaustabh Datta;Buckley, Michael
考古遗址的动物遗骸可以识别动物物种，不仅从它们的形态信息，而且从这些遗迹中保存了数千年甚至数百万年的古代生物分子(脂肪、蛋白质和DNA)，能够更好地了解人与动物之间的关系。然而，由于古代生物分子分析所需的成本和努力，人们已经进行了大量的研究，以开发准确和有效的考古遗迹筛选方法。傅里叶变换红外光谱是一种被考虑用于筛选蛋白质的方法，但由于其预测精度可能因样品保存的程度和使用的仪器而有很大差异，这一事实阻碍了它的广泛使用。此外，用于古DNA(ADNA)分析的筛选方法很少。在这里，我们提出了一种新的方法，通过使用基于随机森林的机器学习，在ZOOM(动物考古学)和aDNA分析之前，极大地改进基于FTIR的筛选方法。为此，我们使用ATR-FTIR检查了三组考古骨骼组合，并对它们进行了放大分析(用于分类鉴定)。其中两个来自旧石器时代，以陆地动物为主，包括具有各种保存条件的标本。第三组由全新世动物遗骸组成，保存程度各不相同，以鲸目动物为主。使用全新世动物遗骸，我们能够更一致地评估基于ATR-FTIR的线粒体DNA筛查以及缩放成功。我们报告了机器学习在基于ATR-FTIR的古代mtDNA筛选中的潜力，我们的机器学习模型最终将基于ATR-FTIR的缩放筛选的准确率提高了20%？40%。结果还表明，这种方法潜在地允许一个通用的筛查系统，适用于多个地点，在很大程度上独立于所使用的光谱仪。
Faunal remains from archaeological sites allow for the identification of animal species that enables the better understanding of the relationships between humans and animals, not only from their morphological information, but also from the ancient biomolecules (lipids, proteins, and DNA) preserved in these remains for thousands and even millions of years. However, due to the costs and efforts required for ancient biomolecular analysis, there has been considerable research into development of accurate and efficient screening approaches for archaeological remains. FTIR spectroscopy is one such approach that has been considered for screening of proteins, but its widespread use has been hindered by the fact that its predictive accuracy can vary widely depending on the extent of sample preservation and the instrument used. Further, screening methods for ancient DNA (aDNA) analysis are scarce. Here we present a new approach to vastly improve upon FTIR-based screening methods prior to ZooMS (Zooarchaeology by Mass Spectrometry) and aDNA analysis through the use of random forest-based machine learning. To do so, we use ATR-FTIR to examine three sets of archaeological bone assemblages and analyse them by ZooMS (for taxonomic identification). Two of these are from Palaeolithic contexts, dominated by terrestrial fauna and include specimens with a variety of preservational conditions. The third set consists of Holocene faunal remains, with variable levels of preservation and is dominated by cetaceans. Using the Holocene faunal remains, we were able to more consistently evaluate ATR-FTIR-based screening for mtDNA as well as ZooMS success. We report on the potential of machine learning in ATR-FTIR-based screening for ancient mtDNA analysis, and our machine learning models conclusively improve the accuracy prior to usage of ATR-FTIR-based screening for ZooMS by 20?40%. The results also suggest this approach potentially allows for a universal screening system, applicable across multiple sites and largely independent of the spectrometers used.