Structural variants (SVs) can disrupt gene function and contribute to pathogenesis of rare disorders. Here, we created a genome-wide knockout dataset across 125,730 individuals with genome sequencing data in the UK's National Genomic Research Library by leveraging the distinct read-depth signal associated with homozygous deletions. We curated 535,699 rare high-confidence homozygous deletion SVs, of which 48,735 were rare. These deletions collectively covered 213Mb or 6.92% of the human genome (4.58% of autosomal sequence), revealing substantial tolerance to complete sequence loss. From a subset of 58,022 individuals with rare disease, we identified 295 individuals with likely diagnostic homozygous deletions impacting protein-coding regions of known disease genes. A further 32 individuals had candidate non-coding SVs in or near to known disease genes, 19/32 (59.37%) of which disrupted 5'-UTR/promoter regions, revealing promoter deletion as an underappreciated cause of rare disorders. Finally, we identify 43 genes with no known rare-disease association but with exonic homozygous deletions in two or more individuals with consistent phenotypes. We describe in detail PDC (phosducin) in Leber Congenital Amaurosis, GCG (glucagon) for a syndromic neurodevelopmental disorder with gastrointestinal involvement, and ENTPD3 for intellectual disability with autism, as candidate novel disease-associated genes. Overall, we create a genome-wide map of homozygous deletions and demonstrate the power of this dataset for rare disease diagnosis and novel disease-gene discovery.