Is evolution predictable? Unlocking fundamental biological insights using new machine learning methods

进化是可预测的吗？如果我问你一种新的抗生素是否会面临耐药性，你会说会吗？如果是更大的变化呢？哪些动物物种将在一百万年或十年后灭绝？新的会是什么样子的？人类文化本身将如何演变？这些问题的答案对我们的生态系统和社会的适应能力有着深远的影响。但目前还不清楚这些问题在多大程度上是可以回答的。进化在理论和实践中是可预测的吗？从哪里来的？在多大程度上？这个未来领袖奖学金将授权我领导一个国际研究人员网络，为进化可预测性的基本问题提供具体答案。他们答案的一个关键是进化趋同，即相似表型或整体表型的重复进化，就像建议的(但也有争议的)胎盘和有袋类哺乳动物的体型重复。融合告诉我们，我们可以从进化中期待什么，是更多相同的东西，还是它的替代方案，全新的东西？但是，尽管许多趋同的例子都是假想的，但还没有研究衡量它们随着进化距离的不同而变化的全部程度，因此也没有测量它们的可预测性。第一次，机器学习的新计算方法将使我们能够测量整个可见表现体的进化收敛程度。这一奖学金将对进化趋同进行有史以来最全面的测试，对不同的进化群体进行两个平行的研究项目，蝴蝶和飞蛾以及哺乳动物--我们自己的进化分支。这将比较在数十万张鳞翅目昆虫照片和3D哺乳动物头骨扫描图像中可见的所有表型的融合程度和进化模式。这将检验在现实世界的宏观进化多样化中，进化最终或仅在局部上是可预测的程度。这将给出长期以来令人类着迷的可预测性问题的定量答案。除此之外，答案将告诉我们，我们可以在多大程度上预测新的进化事件。这些新的见解，以及为获得这些见解而开发的方法，将为实用的进化预测提供新的途径，具有从生物医学科学到技术创新的潜在应用。为了传递这些见解，作为未来领导者研究员，我将部署和推广机器学习方面的三项关键突破。首先，深度学习方法的新应用将在多维空间中嵌入图像，测量它们的相似性，使它们的第一个应用走向进化(HoYAL Cuthill等人，2019年，科学进步；2020年，自然)。其次，我将开发新的机器学习应用程序，通过进化距离直接衡量一种表型与另一种表型的可预测性。然后，第三项创新将在给定进化距离的情况下，结合产生式机器学习方法，开发表型图像预测的途径。为了实现这些目标，该奖学金将围绕进化生物学家詹妮弗·霍亚尔博士建立一个国际研究网络，由埃塞克斯大学生命科学学院主办，由领先的进化、生态和数据科学家指导。Collaborative Project Partners和Co-I将包括日本Cross Compass交叉实验室的工业计算机科学家，以及自然历史博物馆和剑桥大学动物博物馆的世界领先收藏品专家。因此，这一奖学金将在进化研究、开发和应用尖端技术方面产生翻天覆地的变化，为进化是可预测的这一基本科学问题提供新的答案？未来领导者奖学金的特殊范围使得洞察力的整合成为可能，这将为进化科学提供新的理论和预测框架。

Is evolution predictable? If I asked you whether a new antibiotic might face resistance, would you say yes? What about larger changes? Which animal species will be extinct in a million years, or ten? What will new ones look like? How will human culture itself have evolved? The answers to questions like these matter deeply for the resilience of our ecosystem and society. But it is not yet known to what degree such questions are answerable. Is evolution predictable in theory and practice? From what? To what extent?This Future Leaders Fellowship will empower me to lead an international network of researchers to give concrete answers to fundamental questions of evolutionary predictability.A key to their answers is evolutionary convergence, the repeated evolution of similar phenotypes, or overall phenomes, like the suggested (but also disputed) repetition of body forms in placental and marsupial mammals. Convergence offers to tell us what we can expect from evolution, more of the same, or its alternative, something entirely new? But, while many examples of convergence have been hypothesised, no study has measured the full extent to which they vary with evolutionary distance and, therefore, just how predictable they are. For the first time, new computational methods of machine learning will allow us to measure the extent of evolutionary convergence across entire visible phenomes. This fellowship will undertake the most comprehensive tests of evolutionary convergence ever performed, across two parallel research programs on diverse evolutionary groups, the butterflies and moths and the mammals, our own evolutionary clade. This will compare the extent and evolutionary patterns of convergence in all phenotype visible among hundreds of thousands of lepidopteran photographs and images from 3D mammal skull scans. This will test the extent to which evolution is ultimately, or only locally, predictable across real-world macroevolutionary diversifications.This will give quantitative answers to questions of predictability that have long fascinated humanity. Beyond this, the answers will tell us how far we can expect to predict new evolutionary events. These new insights, and the methods developed to gain them, will provide new avenues for practical evolutionary prediction, with potential applications from biomedical science to technological innovation.To deliver these insights, as a Future Leaders Fellow, I will deploy and extend three key breakthroughs in machine learning. First, new applications of deep-learning methods will embed images in multidimensional spaces, measuring their similarity, leading their first applications to evolution (Hoyal Cuthill et al., 2019, Science Advances; 2020, Nature). Second, I will develop new machine learning applications to directly measure the predictability of one phenotype from another with evolutionary distance. The third innovation will then develop avenues for phenotypic image prediction given evolutionary distance, incorporating generative machine learning methods.To achieve these aims, this fellowship will build an international network of researchers around the Fellow, evolutionary biologist Dr Jennifer Hoyal Cuthill, hosted by the School of Life Sciences at the University of Essex and mentored by leading evolutionary, ecological and data scientists. Collaborative Project Partners and Co-I will include industrial computer scientists at Cross Labs, Cross Compass, Japan and experts on world-leading collections at the Natural History Museum and University of Cambridge Zoology Museum. This fellowship will, thereby, resource a sea-change in evolutionary research, developing and applying cutting-edge technology to provide new answers to the fundamental scientific question, is evolution predictable? Integration of the insights, enabled by the exceptional scope of the Future Leaders Fellowship, will provide new theoretical and predictive frameworks for evolutionary science.