Protein Annotation

The GENCODE gene set is created by merging the results of manual and computational gene annotation methods. Manual gene annotation has two major modes of operation: clone-by-clone and targeted annotation. ‘Clone-by-clone’ annotation involves ‘walking’ across a genomic region, investigating the sequence, aligned expression data and computational predictions for each BAC clone. In doing so, an expert annotator investigates all possible genic features and considers all possible annotations and biotypes simultaneously. We believe this approach carries substantial advantages. For example, the decision to annotate a locus as protein-coding or pseudogenic benefits from being able to weigh both possibilities in light of all available evidence. This process helps prevent false positive and false negative misclassifications. Targeted annotation is designed to answer specific questions such as ‘is there an unannotated protein-coding gene in this position?’ Ranked target lists are generated by computational analysis based, for example, on transcriptomic data, shotgun proteomic data or conservation measures. Over the last two years mouse annotation has been dominated by the clone-by-clone approach while the human genome has been refined entirely via targeted reannotation except for the annotation of human assembly patches and haplotypes released by the Genome Reference Consortium (15 (link)), which take a clone-by-clone approach.
Over the last two years, we have focused on two broad areas: completing the first pass manual annotation across the entire mouse reference genome and a dedicated effort to improve the annotation of protein-coding genes in human and mouse.
We have completed the annotation of novel protein-coding genes, lncRNAs and pseudogenes, plus QC and updating previous annotation where necessary for mouse chromosomes 9, 10, 11, 12, 13, 14, 15, 16 and 17. These updates bring the fraction of the mouse genome with completed first pass manual annotation to approximately 97%. In addition, we have continued to work with the NCBI and Mouse Genome Informatics project at the Jackson Laboratory to resolve annotation differences for protein-coding, pseudogene and lncRNA loci. For protein-coding genes this is under the umbrella of the Consensus Coding Sequence (CCDS) project (16 (link)).
We have also manually investigated unannotated regions of high protein-coding potential identified by whole genome analysis using PhyloCSF (17 (link)) (a tool described in more detail below). In human, this led to the addition of 144 novel protein-coding genes and 271 pseudogenes (of which 42 were unitary pseudogenes). In mouse, we annotated orthologous loci for all but 11 of the 144 human protein-coding genes. We have also revisited the annotation of all olfactory receptor loci in both human and mouse, using RNAseq data to define 5′ and 3′ UTR sequences for ∼1400 loci. In human we have also targeted a ‘deep dive’ manual reannotation of genes on clinical panels for paediatric neurological disorders to identify missing functional alternative splicing. Incorporating second and third generation transcriptomic data, we reannotated ∼190 genes and added more than 3600 alternatively spliced transcripts, including ∼1400 entirely novel exons and an additional ∼30kb of CDS. We have also completed an effort to capture all recently described unannotated microexons (18 (link)) into GENCODE, and further added an additional 146 novel microexons mined from public SLRseq data (19 (link)).
As part of the CCDS collaboration with RefSeq, we have checked a large subset of human loci where there was disagreement over gene biotype. Similarly, we have checked all UniProt manually annotated and reviewed (i.e. Swiss-Prot) accessions that lack an equivalent in GENCODE. As a result, we added 32 novel protein-coding loci to GENCODE and rejected more than 200 putative coding loci. Finally, we are manually reviewing genes previously annotated as protein-coding, but with weak or no support based on a method incorporating UniProt, APPRIS, PhyloCSF, Ensembl comparative genomics, RNA-seq, mass spectrometry and variation data (20 (link),21 (link)). Of the 821 loci investigated to date, 54 have had their coding status removed while a further 110 potentially dubious cases remain under review.
The approach taken reflects in the kinds of updates captured in the annotation. For example, the targeted reannotation in human leads to the annotation of few novel protein-coding loci but many novel transcripts at updated protein-coding and lncRNA loci. Conversely, in mouse the emphasis on clone-by-clone annotation identifies many more novel loci and transcripts across a broader range of biotypes (Figure 1).

The reads quality check was performed using FASTQC ([58 ]). A trimming step of the low-quality bases at 5’ and 3’ was performed using Trimmomatic ([59 (link)]). Low-quality nucleotides were trimmed from the ends of the reads (first 8 bases), setting the minimum quality per base at a Phread score of 20 and minimum and maximum length of the reads after cleaning at 25 bp and 240 bp, respectively. Cleaned reads were assembled into transcript sequences using Trinity v.2.11.0 ([60 (link)]) with in silico read normalization, setting the -min_kmer_cov parameter at 2. The clustering of the transcriptome was performed using the CD-hit-est software (v. 4.6.8, [61 (link)],) with 90% identity threshold in order to remove transcriptome redundancy. The whole transcriptome was aligned with BLASTx software ([62 (link)]) versus the Uniprot SwissProt database (downloaded in July 2020), setting the e-value threshold to 1e⁻³. A filtering step was performed at this stage for removing all the matches against bacterial sequences from the transcriptome.
The prediction of the encoded proteins from the assembled transcripts was obtained via TransDecoder v 5.3.0 (https://github.com/TransDecoder/TransDecoder/releases). Coding sequences were identified by the software based on: 1) a minimum length Open Reading Frame (100 by default to minimize the number of false positives); 2) an internal score system; 3) if a candidate ORF is entirely included within the coordinates of another candidate ORF, the longer one is reported. The functional annotation of the predicted proteins was performed by InterProScan (version 5.33) ([63 (link)]).

Open reading frame (ORF) identification and, subsequently, prokaryote-predicted protein product annotations were performed with Prodigal v2.6.1 [66 (link)] implemented in Prokka [75 (link)]. Selected ORFs were also aligned against the NCBI non-redundant (nr) database (accessed in April 2021) using BLASTp for closest homologue taxonomy and functional annotation supplementation. Targeted single gene homologue searches within our data were also performed using BLASTp 2.2.30+ (E-value threshold = 1E−30, identity = 50%) against predicted protein sequences inferred from (i) metatranscriptomic assemblies, (ii) unbinned metagenomic contigs, and (iii) MAGs.

The proteins identified as differentially expressed proteins (DEPs) should satisfy the threshold of p ≤ 0.05 and |Fold Change|≥1.3. The enrichment analysis of DEPs used the clusterProfiler (version 3.4.4) for GO function annotation and KEGG pathways (Yu et al., 2012 (link)). The GO terms and KEGG pathways were regarded as significant protein enrichment annotations using the DEPs.