Machine Learning for collagen peptide biomarker determination in the taxonomic identification of archaeological fish remains

Abstract

Species identification of archaeofaunal remains can be informative of changing local and global ecosystems, as well as how we interacted with them in the past. However, with the vast majority of assemblages being dominated by morphologically indeterminate specimens, methods of biomolecular species identification are becoming more popular, such as the protein fingerprint-based identification approach known as ZooMS (Zooarchaeology by Mass Spectrometry). As larger datasets are being produced, Machine Learning techniques such as those using Random Forest algorithms have been considered to expedite the identification process. However, it has proven difficult to extract meaningful biomarkers from these processes alone and so far only tackled for mammals. Here, we introduce a novel approach based on the principles of the ID3 algorithm for biomarker identification from ZooMS spectral databases, focussing on archaeological fish remains from Europe and the Caribbean. We show that the tool is highly effective at generating comprehensible lists of family- and genus-level biomarkers. At family-level identification, the average sensitivity across five families of fish was ∼0.9, where a specificity of 1 would indicate a greatly effective algorithm that outperforms both traditional Random Forest and Naïve Bayes approaches; at the genus level the mean sensitivity reduced to ∼0.8 across the nine genera tested. However, some anomalous matches were produced, with accuracy dropping when distinguishing between genera of the same family, such as Epinephelus and Mycteroperca belonging to the Serranidae. Therefore, while this tool has value in rapidly producing lists of biomarkers and can efficaciously identify new ZooMS fingerprints based on those markers, it indicates that manual intervention remains a requirement at finer taxonomic resolutions.