Bridging biology and paleontology – a novel combined machine-learning approach to species delimitation

在林奈定义第一个物种之后的250多年里，人们仍然对如何定义和划分物种存在分歧。这种差异在比较化石和活物种时尤为明显，因为它们通常有不同类型和数量的信息，这导致了几个世纪以来许多不同的物种概念。为了比较过去、现在和预测的生物多样性周转率，重建地理模式，并现实地推断进化过程，需要一个标准化的物种分类系统，平等地处理化石和活类群。机器学习（ML）和图像识别的最新发展提供了一个独特的机会，可以在现代分析框架中共同界定化石和最近的物种。我们将开发一种新的ML方法，该方法使用最近物种的图像数据，基于分子数据建立物种边界，以允许相关化石物种的标准化和统一的物种划界。我们将使用Siamese卷积神经网络，它需要相对较少的图像来学习相似性，可以应用于未标记的数据而无需重新训练，甚至可以处理未知的类。作为模型组，我们将使用淡水腹足类的家庭Viviparidae，其中最近和化石的目标群体包括一个可比的，高形态可塑性，在过去造成的分类混乱。此外，我们将i）通过分类学家进行的独立划界来评估基于ML的物种划界系统与传统分类学的价值，ii）测试物种划界系统在不同类型和化石保存程度方面的局限性，iii）通过提供可变质量的系统图像来评估可靠地划界化石物种所需的细节水平，包括简单的图纸从文献中，最后iv）应用新推断的物种边界重建准确的生物多样性模式和估计化石物种组的多样化过程。我们新的机器学习衍生方法将在不同的分类学群体中广泛使用，并成为使物种在空间和时间上具有可比性的重要起点。一个标准化的物种划界系统，适用于现存和灭绝的物种是必要的，以比较整个地质时期的周转事件和生物多样性危机的路径，并最终提供更现实的人类世生物多样性危机的前景。

After more than 250 years after the definition of the first species by Linnaeus, there is still disagreement about how to define and delimit species. This discrepancy is especially evident when comparing fossil and living species, for which typically different types and amounts of information are available, which has led to a multitude of different species concepts over the centuries. To compare past, present and predicted rates of biodiversity turnover, reconstruct biogeographic patterns and infer evolutionary processes realistically, a standardized species classification system is needed that deals with fossils and living taxa equally. Recent developments in machine learning (ML) and image recognition provide a unique opportunity to jointly delimit fossil and recent species in a modern analytical framework. We will develop a novel ML approach that uses image data for recent species, for which species boundaries were established based on molecular data, to allow for a standardized and unified species delimitation of related fossil species. We will use Siamese Convolutional Neural Networks, which require relatively few images to learn similarities, can be applied to unlabeled data without re-training and can even deal with unknown classes. As model group we will use freshwater gastropods of the family Viviparidae, where the recent and fossil target groups encompass a comparable, high morphological plasticity that has caused taxonomic confusion in the past. In addition, we will i) assess the value of a ML-based species delimitation system versus traditional taxonomy by carrying out independent delimitations made by taxonomists, ii) test the limits of the species delimitation system with regard to different types and degrees of fossil preservation, iii) assess the level of detail that is required to reliably delimit fossil species, by feeding the system images of variable quality, including simple drawings from the literature, and finally iv) apply the newly inferred species boundaries to reconstruct accurate biodiversity patterns and estimate diversification processes for the fossil species group. Our new machine-learning-derived approach will be widely usable across different taxonomic groups and form an important starting point for making species comparable through space and time. A standardized species delimitation system that is applicable to extant and extinct species is imperative to compare pathways of turnover events and biodiversity crises throughout geological time and finally provide more realistic outlooks on the Anthropocene Biodiversity Crisis.