Driven by competition, automation, and technology, the genomics community has far exceeded its ambition to sequence the human genome by 2005. By analyzing mammalian genomes, we have shed light on the history of our DNA sequence, determined that alternatively spliced RNAs and retroposed pseudogenes are incredibly abundant, and glimpsed the apparently huge number of non-coding RNAs that play significant roles in gene regulation. Ultimately, genome science is likely to provide comprehensive catalogs of these elements. However, the methods we have been using for most of the last 10 years will not yield even one complete open reading frame (ORF) for every gene—the first plateau on the long climb toward a comprehensive catalog. These strategies—sequencing randomly selected cDNA clones, aligning protein sequences identified in other organisms, sequencing more genomes, and manual curation—will have to be supplemented by large-scale amplification and sequencing of specific predicted mRNAs. The steady improvements in gene prediction that have occurred over the last 10 years have increased the efficacy of this approach and decreased its cost. In this Perspective, I review the state of gene prediction roughly 10 years ago, summarize the progress that has been made since, argue that the primary ORF identification methods we have relied on so far are inadequate, and recommend a path toward completing the Catalog of Protein Coding Genes, Version 1.0.

A primary motivation for sequencing the mouse genome was to accelerate the discovery of mammalian genes by using sequence conservation between mouse and human to identify coding exons. Achieving this goal proved challenging because of the large proportion of the mouse and human genomes that is apparently conserved but apparently does not code for protein. We developed a two-stage procedure that exploits the mouse and human genome sequences to produce a set of genes with a much higher rate of experimental verification than previously reported prediction methods. RT-PCR amplification and direct sequencing applied to an initial sample of mouse predictions that do not overlap previously known genes verified the regions flanking one intron in 139 predictions, with verification rates reaching 76%. On average, the confirmed predictions show more restricted expression patterns than the mouse orthologs of known human genes, and two-thirds lack homologs in fish genomes, demonstrating the sensitivity of this dual-genome approach to hard-to-find genes. We verified 112 previously unknown homologs of known proteins, including two homeobox proteins relevant to developmental biology, an aquaporin, and a homolog of dystrophin. We estimate that transcription and splicing can be verified for >1,000 gene predictions identified by this method that do not overlap known genes. This is likely to constitute a significant fraction of the previously unknown, multiexon mammalian genes.

The availability of draft sequences for both the mouse and human genomes makes it possible, for the first time, to annotate whole mammalian genomes using comparative methods. TWINSCAN is a gene-prediction system that combines the methods of single-genome predictors like GENSCAN with information derived from genome comparison, thereby improving accuracy. Because TWINSCAN uses genomic sequence only, it is less biased toward highly and/or ubiquitously expressed genes than GENEWISE, GENOMESCAN, and other methods based on evidence derived from transcripts. We show that TWINSCAN improves gene prediction in human using intermediate products from various stages of the sequencing and analysis of the mouse genome, from low-redundancy, whole-genome shotgun reads to the draft assembly and the synteny map. TWINSCAN improves on the prior state of the art even when alignments from only 1X coverage of the mouse genome are available. Gene prediction accuracy improves steadily from 1X through 3X, more slowly from 3X to 4X, and relatively little thereafter. The assembly and the synteny map greatly speed the computations, however. Our human annotation using the mouse assembly is conservative, predicting only 25,622 genes, and appears to be one of the best de novo annotations of the human genome to date.

Correct gene prediction is impaired by the presence of processed pseudogenes: nonfunctional, intronless copies of real genes found elsewhere in the genome. Gene prediction programs frequently mistake processed pseudogenes for real genes or exons, leading to biologically irrelevant gene predictions. While methods exist to identify processed pseudogenes in genomes, no attempt has been made to integrate pseudogene removal with gene prediction, or even to provide a freestanding tool that identifies such erroneous gene predictions. We have created PPFINDER (for Processed Pseudogene finder), a program that integrates several methods of processed pseudogene finding in mammalian gene annotations. We used PPFINDER to remove pseudogenes from N-SCAN gene predictions, and show that gene prediction improves substantially when gene prediction and pseudogene masking are interleaved. In addition, we used PPFINDER with gene predictions as a parent database, eliminating the need for libraries of known genes. This allows us to run the gene prediction/PPFINDER procedure on newly sequenced genomes for which few genes are known.