- Protein-coding gene prediction is a fundamental process in genome analysis and structural genome annotation. It involves identifying genomic regions that contain genes encoding proteins and determining their organization, including coding sequences, exons, introns, splice sites, start codons, stop codons, and transcript boundaries. The predicted genes are subsequently used to infer protein sequences and support functional annotation, comparative genomics, evolutionary studies, and investigations of biological processes.
- A protein-coding gene is a DNA sequence that is transcribed into RNA and ultimately provides the information required to produce a protein. In most protein-coding genes, transcription produces a messenger RNA (mRNA) molecule that is processed and translated by ribosomes. In eukaryotes, genes commonly contain exons and introns. Introns are generally removed during RNA splicing, while exons are joined to form mature transcripts. In prokaryotes, protein-coding genes are usually more compact and generally lack the extensive intron–exon organization typical of eukaryotic nuclear genes.
- Protein-coding gene prediction aims to identify the location and structure of genes within a genomic DNA sequence. It determines where a gene begins and ends, which strand contains the gene, where its coding region starts and terminates, and, in eukaryotes, how coding exons are connected through splicing. A gene may produce several transcript isoforms through alternative splicing, alternative promoter usage, and alternative polyadenylation. Consequently, accurate gene prediction may require identifying multiple transcript structures rather than assigning only one coding sequence to each gene.
- The first important feature used in protein-coding gene prediction is the open reading frame (ORF). An ORF is a sequence of nucleotides that can be read in groups of three without encountering an in-frame stop codon. In a conventional protein-coding sequence, translation begins at a start codon and continues until a stop codon is reached. The common start codon is ATG in DNA, corresponding to AUG in mRNA. The standard genetic code includes three stop codons: TAA, TAG, and TGA in DNA. However, the presence of an ORF does not automatically demonstrate that a region is a functional protein-coding gene. Random genomic sequences can contain ORFs, and genuine genes may have incomplete or unusual coding structures. Additional evidence is therefore required to distinguish real genes from chance ORFs.
- Coding regions exhibit sequence characteristics that differ from many noncoding regions. Protein-coding sequences often show characteristic codon usage, nucleotide composition, codon-position patterns, and conservation of reading frames. Because codons are read in groups of three, coding regions frequently display periodic patterns in nucleotide composition. These patterns can be used by computational algorithms to identify likely coding sequences. Other important signals include start and stop codons, splice donor and acceptor sites, promoter-associated features, and sequence similarities to known proteins.
- Gene prediction methods are commonly divided into three broad categories: ab initio prediction, homology-based prediction, and evidence-based prediction. Ab initio methods identify genes using statistical properties and biological signals within the genomic sequence. Homology-based methods compare the genome with known genes or proteins from the same or related organisms. Evidence-based methods use experimental data, particularly RNA sequencing and protein evidence, to support gene structures. Modern annotation pipelines frequently combine all three approaches to improve accuracy.
- Ab initio gene prediction uses computational models to recognize gene structures without requiring direct experimental evidence from the organism being studied. These methods examine characteristics such as codon usage, GC content, coding potential, start and stop codons, splice sites, exon length, intron length, and gene organization. Statistical models learn patterns from known genes and use these patterns to identify potential coding regions in unannotated genomic sequences.
- Hidden Markov models (HMMs) and generalized hidden Markov models (GHMMs) have been widely used in ab initio gene prediction. These models represent different biological states, such as intergenic regions, coding exons, introns, start codons, and stop codons. The algorithm evaluates possible transitions between these states and identifies the most likely arrangement of gene features along a DNA sequence. Other approaches use neural networks, support vector machines, and deep learning models to recognize coding regions and gene structures. The effectiveness of these methods depends on the quality and suitability of their training data.
- Homology-based gene prediction identifies genes by comparing genomic sequences with previously characterized genes, transcripts, or proteins. Protein sequences are often particularly useful because protein-coding regions may remain evolutionarily conserved even when the underlying DNA sequence has changed. Similarity searches using BLAST or DIAMOND can identify candidate homologous proteins and help locate coding regions in a new genome. More detailed protein-to-genome alignment tools can help determine exon boundaries, intron positions, and gene structures. However, evolutionary changes, gene duplication, pseudogenes, and incomplete reference annotations can complicate the interpretation of similarity results.
- Evidence-based gene prediction uses experimental information to support predicted gene structures. RNA sequencing provides evidence about which genomic regions are transcribed and where exon–exon junctions occur. When RNA-seq reads are aligned to a reference genome using splice-aware aligners such as STAR or HISAT2, they can reveal expressed exons, splice junctions, and transcript boundaries. Transcript assembly tools such as StringTie can reconstruct transcript models from aligned reads. Full-length transcript sequencing, including PacBio Iso-Seq and Oxford Nanopore technologies, can provide additional evidence for complete transcript structures and alternative isoforms.
- Protein evidence can also support gene prediction. Known protein sequences from the same or related organisms can be aligned against the genome to identify conserved coding regions. Protein evidence is particularly valuable when RNA-seq data are unavailable or when genes are expressed only in tissues or conditions not represented in the transcriptomic dataset. Tools such as Exonerate and miniprot can align protein sequences to genomic sequences and provide evidence for coding exon structures. Protein evidence should be interpreted carefully because evolutionary divergence and differences in gene structure may affect alignment accuracy.
- Prokaryotic protein-coding gene prediction differs from eukaryotic gene prediction because bacterial and archaeal genomes generally contain compact genes with relatively few introns. Prokaryotic gene prediction focuses on identifying coding regions, start and stop codons, ribosome-binding signals, and gene boundaries. Codon usage and sequence composition are important features, while operon organization provides additional biological context. Prodigal is a widely used tool for predicting protein-coding genes in prokaryotic genomes. Other tools include GeneMarkS-2 and Prokka, which can be used in broader prokaryotic genome annotation workflows.
- Eukaryotic protein-coding gene prediction is more complex because genes frequently contain multiple exons separated by introns. It must identify splice donor and acceptor sites, coding exon boundaries, alternative transcripts, and coding sequences distributed across multiple genomic segments. Eukaryotic genes may also contain long introns, alternative promoters, untranslated regions, and complex regulatory structures. Tools such as AUGUSTUS, GeneMark-ES, GeneMark-ET, and BRAKER are commonly used for eukaryotic gene prediction and annotation. Their performance depends on the organism, genome assembly, available evidence, and training configuration.
- Several specialized software tools are used for protein-coding gene prediction. AUGUSTUS is a gene prediction program that uses probabilistic models to identify gene structures in eukaryotic genomes. It can incorporate extrinsic evidence, such as transcript and protein alignments, to improve predictions. GeneMark provides several gene prediction approaches for different organism types and evidence conditions. Prodigal is widely used for prokaryotic coding sequence prediction. Glimmer is another established gene-finding program, particularly associated with microbial genome analysis. BRAKER combines gene prediction with RNA-seq and protein evidence, using tools such as GeneMark and AUGUSTUS in its annotation workflow. MAKER integrates gene predictions, sequence similarity, and transcript evidence into a genome annotation pipeline.
- A simplified example of prokaryotic gene prediction using Prodigal is shown below:
# Predict protein-coding genes in a prokaryotic genome
prodigal -i genome.fna \
-a predicted_proteins.faa \
-d predicted_cds.fna \
-f gff \
-o gene_predictions.gff \
-p single
- The input file contains the genome sequence in FASTA format. The -a option specifies the predicted protein sequences, while -d specifies the predicted coding DNA sequences. The -f gff option requests GFF-format output, and -p single selects a single-genome prediction mode. For metagenomic or multi-sequence datasets, the appropriate Prodigal mode and options should be selected according to the data.
AUGUSTUS can be used for eukaryotic gene prediction. A simplified example is:
# Predict genes in a eukaryotic genome
augustus –species=human \
–gff3=on \
genome.fa \
> augustus_predictions.gff3
- The selected species parameter must correspond to a suitable AUGUSTUS species model. The example uses a human model only as an illustration; it should not be used for another organism without an appropriate model or training procedure. The accuracy of predictions can be improved by supplying suitable evidence and configuring the analysis for the target species.
- BRAKER provides an integrated workflow for eukaryotic gene prediction using genomic sequence and available evidence. A simplified command is:
# Run BRAKER with RNA-seq alignment evidence
braker.pl –genome=genome.fa \
–bam=rnaseq_alignments.bam \
–species=target_species \
–gff3
- The BAM file should contain suitable RNA-seq alignments against the same genome assembly. The species name is used for the relevant prediction and training configuration, depending on the BRAKER version and setup. Protein evidence can also be incorporated using the supported options and input formats. Researchers should consult the documentation for the installed version because command-line options and workflow requirements may change.
- Gene prediction pipelines generally produce several output files, including predicted gene coordinates, transcript models, coding DNA sequences, and translated protein sequences. These outputs are commonly represented using FASTA, GFF3, or GTF formats. FASTA files store nucleotide or protein sequences, while GFF3 and GTF files describe genomic features and their relationships. A simplified GFF3 example is shown below:
##gff-version 3
chr1 prediction gene 1000 2500 . + . ID=gene1
chr1 prediction mRNA 1000 2500 . + . ID=transcript1;Parent=gene1
chr1 prediction exon 1000 1200 . + . Parent=transcript1
chr1 prediction CDS 1100 1200 . + 0 Parent=transcript1
chr1 prediction exon 1500 1800 . + . Parent=transcript1
chr1 prediction CDS 1500 1800 . + 0 Parent=transcript1
chr1 prediction exon 2200 2500 . + . Parent=transcript1
chr1 prediction CDS 2200 2300 . + 0 Parent=transcript1
- This simplified example represents one predicted gene with a single transcript containing three exons and three coding segments. The first exon includes an untranslated portion before the coding sequence, while the final exon contains sequence after the coding region. The CDS phase values are illustrative and must be calculated consistently for a real annotation. Actual output may contain additional attributes, transcript identifiers, support information, and other feature relationships.
- After gene prediction, functional annotation is generally performed to determine the likely biological roles of the predicted proteins. Protein sequences can be compared against databases such as UniProt and NCBI RefSeq using BLAST or DIAMOND. Protein domains and conserved motifs can be identified using InterProScan, Pfam, and related resources. Gene Ontology terms can provide information about molecular functions, biological processes, and cellular components, while pathway databases such as KEGG and Reactome can support pathway-level interpretation. Functional annotation is distinct from gene prediction: the former assigns possible biological meaning to predicted genes, while the latter identifies their genomic structures.
- Quality assessment is essential because predicted genes are not necessarily correct or complete. One important measure is the completeness of expected conserved genes. BUSCO evaluates genome or gene-set completeness using sets of conserved single-copy orthologs appropriate for particular taxonomic groups. It reports genes detected as complete, duplicated, fragmented, or missing. BUSCO provides useful evidence about gene-space completeness, but it does not independently verify every predicted gene or exon boundary.
- Other validation methods include comparing predicted transcripts with RNA-seq alignments, checking protein sequence similarity, examining conserved protein domains, and inspecting gene structures in genome browsers such as IGV and JBrowse. Predicted proteins can be examined for unexpected internal stop codons, unusual lengths, incomplete conserved domains, or implausible sequence composition. Tools such as AGAT can help process and validate GFF3 and GTF annotations, while gffread can extract transcript and coding sequences and perform several annotation-related checks.
- A simplified example of extracting predicted coding sequences and proteins using gffread is:
# Extract transcript sequences
gffread gene_predictions.gff3 \
-g genome.fa \
-w transcripts.fa
# Extract coding sequences
gffread gene_predictions.gff3 \
-g genome.fa \
-x predicted_cds.fa
# Extract predicted proteins
gffread gene_predictions.gff3 \
-g genome.fa \
-y predicted_proteins.fa
- The annotation and reference genome must correspond to the same assembly. Extracted sequences can then be examined for completeness, reading-frame consistency, unexpected stop codons, and agreement with available transcript or protein evidence. Partial genes and biologically unusual coding sequences should be interpreted according to the organism and annotation context rather than automatically treated as errors.
- Protein-coding gene prediction has many applications in genomics and biological research. In comparative genomics, predicted genes allow researchers to identify orthologs and paralogs, compare gene content, investigate gene duplication and loss, and study genome evolution. In functional genomics, gene models provide the basis for transcript quantification, differential expression analysis, and gene-level interpretation. In evolutionary biology, gene structures and protein sequences can be used to reconstruct evolutionary relationships and examine conserved biological functions. In medical research, accurate gene annotation supports the identification and interpretation of disease-associated variants. In agriculture, gene prediction helps identify candidate genes associated with productivity, development, stress tolerance, and disease resistance in crops and livestock.
- Gene prediction also supports the analysis of newly sequenced organisms. When a reference genome is available but gene annotations are incomplete or absent, computational prediction provides an initial catalogue of protein-coding genes. These predictions can guide subsequent transcriptomic, proteomic, and functional studies. In metagenomics, gene prediction identifies coding regions within assembled microbial sequences and provides protein sequences for taxonomic and functional analysis. Tools such as Prodigal and other metagenomic gene callers are used to identify coding sequences in individual contigs or metagenome assemblies.
- Several challenges can affect protein-coding gene prediction accuracy. Fragmented genome assemblies may split genes across multiple contigs, while gaps and sequencing errors can disrupt coding regions. Repetitive elements can generate misleading gene predictions or interfere with sequence alignments. Highly divergent genes may not show strong similarity to known proteins, and lineage-specific genes may lack reference evidence. Short-read RNA-seq data may fail to resolve complex alternative splicing patterns, while incomplete transcript evidence can lead to missing exons or incorrect transcript boundaries.
- Gene prediction can also be affected by incomplete training data and differences between organisms. A model trained on one species may not accurately recognize gene structures in another, particularly when intron lengths, GC content, codon usage, or splice-site patterns differ. Some organisms have unusual genetic codes, noncanonical translation initiation sites, alternative splicing patterns, or specialized gene structures. Mitochondrial and chloroplast genomes may also require different prediction settings from nuclear genomes. Researchers should therefore select tools and models appropriate to the organism and genome type.
- Another important challenge is distinguishing protein-coding genes from pseudogenes and noncoding transcripts. Pseudogenes may resemble functional genes but contain disabling mutations, disrupted reading frames, or other features that prevent production of a functional protein. Some noncoding transcripts may contain short ORFs without producing biologically meaningful proteins, while some genuine protein-coding genes encode very short peptides that are difficult to identify. Coding-potential prediction should therefore be supported by transcript evidence, protein homology, conserved domains, evolutionary conservation, ribosome profiling, or proteomics data where available.
- For reproducible gene prediction, researchers should record the reference genome assembly and version, software names and versions, species models, training datasets, input evidence, and command-line parameters. They should preserve the original prediction files and document any manual corrections or filtering. If multiple annotation tools are used, the methods for combining their predictions should be clearly described. Gene identifiers and transcript identifiers should remain consistent across genome annotation, functional annotation, expression analysis, and downstream interpretation.
- Protein-coding gene prediction is a central step in transforming a genome sequence into a biologically interpretable resource. It identifies candidate genes, reconstructs their coding structures, and provides the sequences needed for protein prediction and functional analysis. Reliable results depend on integrating ab initio algorithms, homology-based searches, RNA and protein evidence, appropriate genome-specific models, and systematic quality assessment. Accurate gene prediction provides the foundation for genome annotation, comparative genomics, functional genomics, evolutionary research, and the interpretation of genetic variation.