- Eukaryotic gene structure and prediction are important areas of bioinformatics that focus on identifying genes and reconstructing their organization within eukaryotic genomes. Eukaryotic organisms, including animals, plants, fungi, and protists, generally have more complex genome structures than prokaryotes. Their genes commonly contain exons, introns, untranslated regions, promoters, regulatory elements, and multiple transcript variants. A single gene may produce several messenger RNA (mRNA) transcripts through alternative splicing, and these transcripts may encode different protein products. Computational gene prediction methods therefore need to identify not only potential coding sequences but also the complex structural features that determine how genes are transcribed and processed.
- A eukaryotic gene is a functional unit of DNA that contains information required to produce a protein or a functional RNA molecule. Protein-coding genes are transcribed into precursor messenger RNA (pre-mRNA), which undergoes processing before becoming mature mRNA. This processing commonly includes the removal of introns, joining of exons, addition of a 5′ cap, and formation of a 3′ polyadenylated end. Mature mRNA is subsequently translated into a protein by ribosomes. Non-coding RNA genes follow different processing and functional pathways and do not necessarily produce proteins. Examples include transfer RNA (tRNA), ribosomal RNA (rRNA), microRNA (miRNA), small nuclear RNA (snRNA), and long non-coding RNA (lncRNA).
- The basic structure of a typical eukaryotic protein-coding gene includes a promoter, transcription start site, 5′ untranslated region, coding exons, introns, translation termination region, 3′ untranslated region, and transcription termination and polyadenylation-associated signals. These components vary among genes and organisms, and not every gene contains the same regulatory elements or structural organization. A gene may contain a single exon or many exons separated by introns. Some genes also have alternative promoters, alternative transcription start sites, and multiple polyadenylation sites, producing different transcript structures.
- Promoters are regulatory DNA regions located near transcription start sites. They contain sequence elements that help recruit RNA polymerase and transcription factors to initiate transcription. Some eukaryotic promoters contain a TATA box, commonly located approximately 25–30 base pairs upstream of the transcription start site, although its position and presence vary. Other promoter-associated elements include initiator sequences, downstream promoter elements, and GC-rich regions. Many genes, particularly housekeeping genes, have promoters that lack a recognizable TATA box. Promoter prediction is therefore challenging because promoter architecture is diverse and sequence motifs alone do not reliably identify all transcription initiation sites.
- Enhancers and silencers are additional regulatory elements that influence gene expression. Enhancers can increase transcription through interactions with transcription factors and other regulatory proteins, while silencers can reduce transcriptional activity. These elements may be located upstream, downstream, within introns, or at considerable distances from the genes they regulate. Their regulatory effects can depend on tissue type, developmental stage, environmental conditions, and chromatin organization. Although enhancers and silencers are important components of gene regulation, their prediction is generally considered a related but distinct task from the prediction of gene structures.
- The transcription start site (TSS) marks the position where RNA polymerase begins synthesizing an RNA transcript. The region downstream of the TSS is transcribed into pre-mRNA, including the 5′ untranslated region and the coding sequence. Accurate TSS identification is important because alternative transcription initiation can produce different transcript variants from the same gene. These variants may differ in their 5′ UTRs and sometimes in their coding sequences. Experimental methods such as CAGE and other transcript-end sequencing approaches can provide evidence for transcription start sites.
- The 5′ untranslated region lies between the transcription start site and the translation initiation codon. It is transcribed into mRNA but is not translated into protein. The 5′ UTR can contain regulatory elements that influence mRNA stability, ribosome recruitment, and translation efficiency. Some 5′ UTRs contain upstream open reading frames (uORFs), which can regulate translation of the main protein-coding sequence. The length and sequence composition of the 5′ UTR vary considerably among genes and transcripts.
- Exons are genomic regions that remain in the mature RNA transcript after splicing. In protein-coding genes, some exonic regions encode amino acids, while others form untranslated regions. The coding portions of exons are joined together during RNA splicing to produce a continuous coding sequence. Exons may be short or long and can contain complete codons or partial codons at their boundaries. Consequently, exon boundaries do not necessarily coincide with codon boundaries. Gene prediction algorithms must preserve the correct reading frame when reconstructing coding sequences across multiple exons.
- Introns are intervening regions located between exons. They are transcribed into pre-mRNA but are generally removed during RNA splicing. Introns can vary greatly in length and may contain regulatory elements, repetitive sequences, non-coding RNA genes, and other functional or non-functional genomic regions. The number and length of introns differ among genes and organisms. Some eukaryotic genes contain very short introns, whereas others contain extremely long introns that make gene prediction more difficult.
- Splice-site signals provide important evidence for identifying exon–intron boundaries. Most conventional nuclear introns in eukaryotes follow the GT–AG rule in genomic DNA. The 5′ splice donor site commonly begins with GT, while the 3′ splice acceptor site ends with AG. During RNA processing, these correspond to GU and AG at the RNA level. The spliceosome recognizes these signals together with additional sequence and structural information. Other important splicing elements include the branch-point sequence, usually containing an essential adenine residue, and the polypyrimidine tract located near the 3′ splice site. Not all introns follow the GT–AG pattern; GC–AG and AT–AC introns also occur. Gene prediction methods use splice-site probabilities and surrounding sequence features to identify likely intron boundaries.
- Alternative splicing is one of the most important characteristics of eukaryotic gene organization. It allows different combinations of exons to be included in mature transcripts from the same gene. Common alternative splicing patterns include exon skipping, alternative 5′ splice sites, alternative 3′ splice sites, mutually exclusive exons, and intron retention. These events can produce multiple transcript isoforms with different coding sequences, protein domains, cellular localization signals, or regulatory properties. Some transcript variants may also contain premature termination codons and be degraded through nonsense-mediated decay. Accurate gene annotation must therefore distinguish between a gene locus and its individual transcript isoforms.
- The coding sequence (CDS) consists of the portions of exons that are translated into protein. It begins at a translation initiation codon and ends at a termination codon. In eukaryotic nuclear protein-coding genes, ATG is the common DNA initiation codon, corresponding to AUG in mRNA. The standard termination codons are TAA, TAG, and TGA. The CDS may be distributed across multiple exons, so its complete sequence must be reconstructed by joining the coding portions of exons in the correct order and orientation. Incorrect exon boundaries can change the reading frame and produce an incorrect predicted protein.
- The 3′ untranslated region lies downstream of the translation termination codon and remains part of mature mRNA without being translated. It may contain regulatory sequences involved in mRNA stability, localization, translation, and degradation. Many eukaryotic mRNAs contain a polyadenylation signal, commonly AAUAAA or a related variant, upstream of the cleavage and polyadenylation site. The addition of a poly(A) tail contributes to mRNA stability, export, and translation. Alternative polyadenylation can produce transcripts with different 3′ UTR lengths, affecting regulatory interactions and gene expression.
- Eukaryotic genes may also contain alternative promoters, alternative transcription termination sites, and overlapping transcriptional units. These features increase the complexity of gene annotation because a single genomic locus can produce several RNA products. Some genes have multiple independent promoters that initiate transcription from different positions, while others produce transcripts that overlap with neighbouring genes. In complex genomes, transcription can also occur from both DNA strands in overlapping regions. Gene prediction methods must therefore consider alternative transcript structures and genomic context rather than assuming that each gene has one simple linear structure.
- Eukaryotic gene prediction is the computational process of identifying genes and their structural components in genomic DNA. It involves locating candidate protein-coding regions, exons, introns, splice junctions, start and stop codons, transcription-related features, and possible transcript isoforms. The process is more complicated than prokaryotic gene prediction because eukaryotic genes commonly contain introns and may produce multiple alternatively spliced transcripts. A successful prediction must identify the correct genomic regions, strand orientation, exon boundaries, coding frame, and transcript organization.
- One of the fundamental principles of eukaryotic gene prediction is the identification of coding potential. Protein-coding exons exhibit characteristic statistical properties related to codon usage, nucleotide composition, amino acid requirements, and evolutionary constraints. Prediction algorithms examine codon frequencies, hexamer frequencies, GC content, reading-frame-specific patterns, and other sequence characteristics to distinguish coding regions from non-coding DNA. These properties are especially useful when identifying coding exons that may be separated by long introns.
- Splice-site recognition is another central principle. Since eukaryotic protein-coding genes commonly contain introns, prediction algorithms must identify the boundaries at which introns are removed. They examine donor and acceptor motifs, surrounding nucleotide composition, branch-point signals, and other features associated with RNA splicing. However, splice-site motifs are not perfectly specific. Similar sequences may occur by chance in non-splice regions, and genuine splice sites may deviate from consensus patterns. Statistical models and additional biological evidence are therefore necessary to improve boundary prediction.
- Gene prediction algorithms also use exon and intron length distributions. Different organisms and gene classes show characteristic patterns in exon and intron lengths. These distributions can help distinguish likely gene structures from random sequence arrangements. However, eukaryotic genes vary substantially in size, and individual genomes may contain both compact genes and genes with very long introns. Gene prediction systems therefore use statistical distributions learned from training data rather than relying on fixed length limits.
- Ab initio gene prediction is an important approach in eukaryotic genome analysis. It predicts gene structures from genomic sequence characteristics without requiring direct experimental evidence for every gene. These methods use statistical models to recognize coding regions, splice sites, start and stop signals, and other gene-associated features. Hidden Markov models (HMMs) and generalized hidden Markov models are widely used in this context. In a typical model, hidden states represent biological components such as intergenic regions, coding exons, introns, and untranslated exons. Each state has characteristic emission probabilities, while transition probabilities describe possible movements between gene components.
- Generalized hidden Markov models are particularly useful because biological segments such as exons and introns have variable lengths. They can model the probability distributions of segment lengths and incorporate different sequence characteristics for coding and non-coding regions. Dynamic programming algorithms, including the Viterbi algorithm, can be used to identify the most likely sequence of hidden states. More advanced models may also incorporate information about reading frames, splice signals, coding statistics, and gene structure constraints.
- Another important computational approach uses probabilistic sequence models and machine learning. Traditional gene prediction systems rely heavily on statistical models trained on known genes, while newer approaches may incorporate support vector machines, neural networks, deep learning, and other classification methods. These techniques can learn complex sequence patterns associated with coding exons, splice sites, promoters, and regulatory regions. However, their accuracy depends on the quality and representativeness of training datasets. Models trained on one species may not perform reliably on another species with different gene structures and genomic characteristics.
- Evidence-based gene prediction uses biological information obtained from experiments or previously annotated sequences. RNA sequencing (RNA-seq) is one of the most important sources of evidence. RNA-seq reads are aligned to the reference genome to identify expressed regions, exon boundaries, and splice junctions. Reads that span exon–exon junctions provide evidence that two genomic regions belong to the same mature transcript. Transcript assembly can reconstruct candidate RNA molecules and help identify alternative splicing patterns. However, RNA-seq evidence depends on tissue type, developmental stage, environmental conditions, and sequencing depth. A gene that is not detected in one sample may still be present and functional.
- Full-length transcript sequencing can provide additional evidence for eukaryotic gene structures. Technologies such as PacBio Iso-Seq and Oxford Nanopore direct RNA or cDNA sequencing can generate long reads that cover much or all of individual transcripts. These data can help resolve complex isoforms, alternative transcription start sites, and alternative polyadenylation. However, long-read data may contain sequencing errors or incomplete transcript coverage, and computational processing is required before the evidence can be incorporated into genome annotation.
- Protein homology is another valuable source of evidence. Predicted genomic regions can be translated into protein sequences and compared with known proteins from related species. Similarity to conserved proteins can support the presence of a gene and help identify coding exons. Protein domain information can also provide evidence for gene structures and possible biological functions. Databases such as UniProt, RefSeq, Pfam, and InterPro support these analyses. Homology-based methods may fail to detect genes that are newly evolved, rapidly diverging, lineage-specific, or absent from available reference databases.
- Integrated gene prediction combines ab initio methods with transcript and protein evidence. This approach is particularly valuable for eukaryotic genomes because gene structures are complex and sequence signals alone may be insufficient. AUGUSTUS is a widely used eukaryotic gene prediction program that can incorporate external evidence. GeneMark provides statistical gene prediction methods, while SNAP uses hidden Markov model-based approaches and can be trained for particular organisms. BRAKER is an integrated genome annotation workflow that combines GeneMark-based prediction with AUGUSTUS and available RNA-seq or protein evidence, depending on the workflow configuration. These methods can improve predictions when suitable evidence and training data are available.
- AUGUSTUS is commonly used to predict gene structures from genomic sequences. A basic example is:
augustus \
–species=human \
genome.fasta \
> predicted_genes.gff
- The species parameter must refer to an available and appropriate model. This command illustrates a basic ab initio prediction and does not incorporate RNA-seq or protein evidence. The resulting annotations should be evaluated and, where possible, improved using supporting evidence.
- BRAKER can be used to combine genome sequence information with RNA-seq or protein evidence. A typical RNA-seq-supported workflow may be initiated as follows, depending on the installed version and configuration:
braker.pl \
–genome=genome.fasta \
–bam=rnaseq_alignments.bam \
–species=sample_species \
–workingdir=braker_output
- The BAM file should contain suitable RNA-seq alignments against the genome, prepared according to the workflow requirements. BRAKER options and evidence requirements vary by release, so the installed version’s documentation should be consulted before running a production analysis.
- Gene prediction results are commonly stored in GFF3 or GTF annotation files. These formats describe genomic features such as genes, transcripts, exons, coding sequences, and untranslated regions. A gene may have several transcript records, each with its own exon structure and coding sequence. The annotation must preserve parent–child relationships between genes, transcripts, and their component features. Predicted coding sequences and protein products are often stored in FASTA files. Correct strand information and coordinate conventions are essential because errors in these details can produce incorrect sequence extraction and protein translation.
- Gene prediction accuracy is evaluated using reference annotations and statistical performance measures. Sensitivity, or recall, measures the proportion of genuine features that are correctly predicted. Precision measures the proportion of predicted features that are correct, while specificity measures the ability to identify non-gene regions correctly. The F1-score combines precision and recall. Evaluation may be performed at nucleotide, exon, splice-site, transcript, or complete gene level. This distinction is important because a prediction may identify most coding nucleotides correctly but still have inaccurate exon boundaries or transcript structures.
- False positives and false negatives are major challenges in eukaryotic gene prediction. False positives occur when non-coding regions, pseudogenes, or random ORFs are incorrectly predicted as genes. False negatives occur when genuine genes are missed. Short genes, genes with unusual exon structures, genes expressed only in specific tissues, and genes located in repetitive or poorly assembled regions can be difficult to identify. Alternative splicing further complicates prediction because different transcripts may share some exons but contain different combinations of others. Incomplete genome assemblies and sequencing errors can also disrupt predicted gene structures.
- Pseudogenes require particular attention during eukaryotic genome annotation. They are sequences derived from functional genes that have lost some or all of their original function. Processed pseudogenes often originate from reverse-transcribed mRNA and may lack introns and promoter regions, while duplicated pseudogenes may retain exon–intron structures. Some pseudogenes can resemble functional protein-coding genes and may therefore be incorrectly predicted as intact genes. Additional evidence, including coding integrity, evolutionary conservation, transcription, and genomic context, helps distinguish pseudogenes from functional genes.
- Non-coding RNA genes must also be identified separately from protein-coding genes. Eukaryotic genomes contain many functional RNA molecules that do not encode proteins. These include transfer RNA, ribosomal RNA, microRNA, small nuclear RNA, small nucleolar RNA, and long non-coding RNA. Their identification requires specialized sequence and structure-based methods. Tools such as tRNAscan-SE are used for transfer RNA prediction, while Infernal and covariance models can identify conserved structured RNA families. Dedicated resources and tools are also available for microRNA and other non-coding RNA annotations.
- Eukaryotic gene prediction is closely related to genome annotation and functional genomics. Once gene structures have been identified, predicted proteins can be analysed using sequence similarity searches, conserved domain detection, protein family classification, and functional annotation. BLAST can identify similarities to known proteins, while InterProScan and Pfam can identify conserved domains and protein families. Gene Ontology annotations can describe molecular functions, biological processes, and cellular components. Pathway resources such as KEGG and Reactome can help identify the biological pathways in which predicted gene products may participate.
- Gene prediction is widely applied in human genomics, plant genomics, animal genomics, fungal research, evolutionary biology, agricultural biotechnology, and disease studies. In human genomics, it supports the annotation of protein-coding genes and transcript isoforms, helping researchers interpret genetic variation and gene regulation. In plant and animal genomics, gene prediction helps identify genes associated with development, reproduction, productivity, disease resistance, and environmental adaptation. In comparative genomics, predicted genes can be compared across species to identify conserved genes, gene duplications, gene losses, and lineage-specific evolutionary changes.
- Reliable eukaryotic gene prediction depends on genome assembly quality, appropriate training data, suitable computational methods, and the availability of supporting biological evidence. Researchers should consider genome completeness, sequencing accuracy, repeat content, expected gene structure, tissue-specific expression, and evolutionary relationships. Software versions, reference assemblies, training parameters, transcript evidence, protein databases, and filtering criteria should be documented. Genome annotation should be treated as an evolving process because new transcript evidence, improved assemblies, and updated computational models may change gene boundaries and transcript structures.
- Eukaryotic gene structure and prediction provide the foundation for understanding the organization and function of complex genomes. The identification of promoters, exons, introns, splice sites, untranslated regions, coding sequences, and transcript isoforms requires a combination of sequence analysis, statistical modelling, transcript evidence, and protein homology. No single prediction method can identify every gene and transcript accurately across all eukaryotic organisms. Integrating complementary evidence and validating predictions at multiple levels produces more reliable genome annotations and supports downstream functional, comparative, and evolutionary analyses.