A central challenge in human genomics is to understand the cellular, evolutionary, and clinical
significance of genetic variants. Here we introduce a unified population-genetic and machine-
learning model, called Linear Allele-Specific Selection InferencE (LASSIE), for estimating the
fitness effects of all potential single-nucleotide variants, based on polymorphism data and pre-
dictive genomic features. We applied LASSIE to 51 high-coverage genome sequences annotated
with 33 genomic features, and constructed a map of allele-specific selection coefficients across all
protein-coding sequences in the human genome. We show that this map is informative about both
human evolution and disease.