Deep learning has shown its super power in demographic inference, especially when large number of summary statistics are involved. I created some scripts to test the performance of neural network regression and ABC in demographic inference. As expected, deep learning outperforms the traditional ABC in discriminatory power and prediction accuracy.
The most informative statistics for population structure inference
Information theory provides a spectrum of information measures, which can be used to infer the signals of evolutionary changes from DNA. I tested 127 information-based statistics for population geography inference using deep learning. I identified the most informative summary statistics for population demographic inference. Results can be found here.
A new approach for inferring population structure and their geographic origin
I proposed a new method for inferring population structure and the geographic origin of individuals. I have compared the performance with PCA, t-SNE and UMAP. Notably, my approach outperforms the current existing methods in population structure inference, especially for admixed populations with subtle structure. Tools will be available soon.