Steffen Schmidt

Protein design aims to build novel proteins customized for specific purposes, thereby holding the potential to tackle many environmental and biomedical problems. Recent progress in Transformer-based architectures has enabled the implementation of language models capable of generating text with human-like capabilities. Here, motivated by this success, we describe ProtGPT2, a language model trained on the protein space that generates de novo protein sequences following the principles of natural ones. The generated proteins display natural amino acid propensities, while disorder predictions indicate that 88% of ProtGPT2-generated proteins are globular, in line with natural sequences. Sensitive sequence searches in protein databases show that ProtGPT2 sequences are distantly related to natural ones, and similarity networks further demonstrate that ProtGPT2 is sampling unexplored regions of protein space. AlphaFold prediction of ProtGPT2-sequences yields well-folded non-idealized structures with embodiments and large loops and reveals topologies not captured in current structure databases. ProtGPT2 generates sequences in a matter of seconds and is freely available.

Quantifying the distribution of fitness effects among newly arising mutations in the human genome is key to resolving important debates in medical and evolutionary genetics. Here, we present a method for inferring this distribution using Single Nucleotide Polymorphism (SNP) data from a population with non-stationary demographic history (such as that of modern humans). Application of our method to 47,576 coding SNPs found by direct resequencing of 11,404 protein coding-genes in 35 individuals (20 European Americans and 15 African Americans) allows us to assess the relative contribution of demographic and selective effects to patterning amino acid variation in the human genome. We find evidence of an ancient population expansion in the sample with African ancestry and a relatively recent bottleneck in the sample with European ancestry. After accounting for these demographic effects, we find strong evidence for great variability in the selective effects of new amino acid replacing mutations. In both populations, the patterns of variation are consistent with a leptokurtic distribution of selection coefficients (e.g., gamma or log-normal) peaked near neutrality. Specifically, we predict 27-29% of amino acid changing (nonsynonymous) mutations are neutral or nearly neutral (|s|<0.01%), 30-42% are moderately deleterious (0.01%<|s|<1%), and nearly all the remainder are highly deleterious or lethal (|s|>1%). Our results are consistent with 10-20% of amino acid differences between humans and chimpanzees having been fixed by positive selection with the remainder of differences being neutral or nearly neutral. Our analysis also predicts that many of the alleles identified via whole-genome association mapping may be selectively neutral or (formerly) positively selected, implying that deleterious genetic variation affecting disease phenotype may be missed by this widely used approach for mapping genes underlying complex traits.

Alu retrotransposons evolved from 7SL RNA approximately 65 million years ago and underwent several rounds of massive expansion in primate genomes. Consequently, the human genome currently harbors 1.1 million Alu copies. Some of these copies remain actively mobile and continue to produce both genetic variation and diseases by "jumping" to new genomic locations. However, it is unclear how many active Alu copies exist in the human genome and which Alu subfamilies harbor such copies. Here, we present a comprehensive functional analysis of Alu copies across the human genome. We cloned Alu copies from a variety of genomic locations and tested these copies in a plasmid-based mobilization assay. We show that functionally intact core Alu elements are highly abundant and far outnumber all other active transposons in humans. A range of Alu lineages were found to harbor such copies, including all modern AluY subfamilies and most AluS subfamilies. We also identified two major determinants of Alu activity: (1) The primary sequence of a given Alu copy, and (2) the ability of the encoded RNA to interact with SRP9/14 to form RNA/protein (RNP) complexes. We conclude that Alu elements pose the largest transposon-based mutagenic threat to the human genome. On the basis of our data, we have begun to identify Alu copies that are likely to produce genetic variation and diseases in humans.