Deep learning and satellite imagery predict genetic diversity and differentiation

Abstract

During the last decade, convolutional neural networks (CNNs) have revolutionised the application of deep learning (DL) methods to classification tasks and object recognition. These procedures can capture key features of image data that are not easily visible to the human eye and use them to classify and predict outcomes with exceptional precision. Here, we show for the first time that CNNs provide highly accurate predictions for small‐scale genetic differentiation and diversity in Ctenomys australis, a subterranean rodent from central Argentina. Using microsatellite genotypes and high‐resolution satellite imagery, we trained a simple CNN to predict local FST and mean allele richness. To identify landscape features with high impact on predicted values, we applied species distribution models to obtain the distribution of suitable habitat. Subsequent use of a machine learning algorithm (random forest) allowed us to identify the attributes that contribute the most to predictions of population genetic metrics. Predictions obtained from the CNN accounted for more than 98% of the variation observed both in FST and mean allele richness values. Random forest regression on landscape metrics indicated that features involving connectivity and consistent prevalence of suitable habitat promoted genetic diversity and reduced genetic differentiation in C. australis. Validation with synthetic data via simulations of genetic differentiation based on the landscape structure of the study area and of a nearby area showed that DL models are able to capture complex relationships between actual data and synthetic data in the same landscape and between synthetic data generated under different landscapes. Our approach represents an objective and powerful approach to landscape genetics because it can extract information from patterns that are not easily identified by humans. Spatial predictions from the CNN may assist in the identification of areas of interest for biodiversity conservation and management of populations.

Spectral-Based Classification of Genetically Differentiated Groups in Spring Wheat Grown under Contrasting Environments

“…In another study, Zhu et al [ 42 ] modeled a CNN using near-infrared hyperspectral imaging to classify three varieties of soybeans (i.e., “Zhonghuang37”, “Zhonghuang41”, and “Zhonghuang55”), and each variety has a classification accuracy of over 90%. Kittlein et al [ 43 ] performed a CNN model to provide a highly accurate prediction of the genetic diversity and differentiation of Ctenomys australis populations using molecular markers and high-resolution satellite imagery. CNN-based predictions accounted for about 98% of the variation observed in the genetic differentiation index and mean allele richness values, which may facilitate the identification of areas of interest for the conservation and management of populations.…”

Deep Learning in Population Genetics

“…Nevertheless, very recent developments have led to new approaches, DASP ( Ancona et al 2019 ) and G-DeepShap ( Chen et al 2022 ), that may scale up to population genomics datasets. For the moment, there are no applications of Shapley values to population genomics studies; there is only an application in population genetics but in the context of random forests ( Kittlein et al 2022 ).…”

An ensemble learning approach to map the genetic connectivity of the parasitoid Stethynium empoasca (Hymenoptera: Mymaridae) and identify the key influencing environmental and landscape factors

“…Further, with increasing access to remote sensing and climate data, machine learning methods can be used to explore the effects of multiple environmental factors on genetic variations at any sampling scale. Some machine learning approaches include deep learning (Kittlein et al, 2022) and random forest (Murphy et al, 2010;Sylvester et al, 2018;Shanley et al, 2021) been recently developed for their application in landscape genetics.…”

“…Despite increasing applications, current machine learning methods still show some limitations for landscape genetics. For example, a convolutional neural network (CNN), a deep learning method first introduced in landscape genetics by Kittlein et al (2022), usually performs poorly on small datasets (Elavarasan et al, 2018). This is a major limitation as the number of sample sites in most population-based landscape genetic studies is often <50.…”