A Systematic Review: Deep Learning for Analyzing Genomic Data to Discover Evolutionary Patterns Authors: Raha Hassanpour Faramoushjani, Sanam Ansari Published: 2026 ABSTRACT Deep learning has been increasingly applied to evolutionary genomics as genomic datasets have grown in scale and complexity. However, the literature encompasses heterogeneous biological objectives, modeling assumptions, and evaluation standards, often treated as a unified field despite important conceptual differences. This study presents a systematic review of research published between 2016 and 2025 on the use of deep learning to identify evolutionary patterns in genomic data. Following a structured screening process, 50 studies were selected for qualitative synthesis. The reviewed applications can be organized into three partially overlapping but conceptually distinct domains: (i) population genetic inference, (ii) phylogenetic reconstruction, and (iii) sequence representation learning using DNA and protein language models. In population genetics, deep learning is predominantly employed within simulation‐based inference frameworks. In phylogenetics, neural architectures are used to approximate or accelerate tree and model inference under defined conditions. In representation learning, models focus on extracting transferable sequence features for downstream evolutionary or functional analyses. Across domains, deep learning provides flexible modeling of complex genomic inputs. Nevertheless, recurring limitations include challenges in interpretability, sensitivity to training assumptions—particularly under simulation‐based settings—heterogeneous evaluation protocols, and substantial computational demands. By organizing the literature using a domain‐aligned framework, this review clarifies domain‐specific strengths, limitations, and research gaps, providing a structured basis for future methodological development in evolutionary genomics. 1. Introduction As a key resource in biology and medicine, genomic data enable deeper understanding of biological processes, including the evolution of species. Accurate analysis of these data supports the identification of evolutionary patterns—such as genetic change through time, mutation and adaptation, and phylogenetic relationships—which in turn informs genetics, personalized medicine, biotechnology, and biological resource management [1, 2]. The rapid adoption of next‐generation sequencing (NGS) has made large volumes of genomic data widely available, accelerating discovery while also creating practical challenges for analysis and interpretation [3]. High dimensionality, heterogeneous evolutionary regimes, and measurement noise can complicate inference. Classical statistical pipelines and traditional machine learning approaches remain essential, but they can face limitations in certain high‐dimensional or highly heterogeneous settings, particularly when strong model assumptions are violated or when flexible representation learning is required [4]. Deep learning has expanded the computational toolkit for analyzing complex, high‐volume data and has been increasingly applied in computational biology and genomics [5]. In evolutionary genomics, deep learning methods have been explored for tasks including selection detection, demographic inference, admixture and hybridization characterization, and phylogenetic reconstruction [6–9]. Importantly, evidence reported across studies is often context‐dependent: performance and computational feasibility depend on the chosen input representations, training regimes, and evaluation designs [10–13]. Despite progress, accurate inference of evolutionary patterns from genomic data remains challenging due to model misspecification risk, computational demands, and evolutionary heterogeneity [8, 14]. Many population‐genetic applications rely on simulation‐based supervised learning, where external validity depends on how well simulated training regimes approximate empirical complexity; distribution shift and demographic or recombination misspecification can therefore degrade generalization [6, 15, 16]. In addition, ongoing debates concern the trade‐off between predictive accuracy and interpretability and the robustness of models when training data or generative assumptions are uncertain [6, 17]. Motivated by these challenges, this review adopts a domain‐aligned perspective that separates biologically distinct objectives and evaluation practices. Applications of deep learning in evolutionary genomics can be grouped into three partially overlapping but conceptually distinct domains: (i) population genetic inference, (ii) phylogenetic reconstruction, and (iii) sequence representation learning through self‐supervised DNA and protein language models. While these domains share computational foundations, they address different biological questions and rely on different modeling and validation frameworks; in the reviewed literature, population genetic inference dominates empirical applications, whereas language models represent a complementary and emerging paradigm [6, 7, 18, 19]. The aim of this systematic review is to critically evaluate deep learning methods applied to discovering evolutionary patterns in genomic data, with emphasis on methodological choices, validation practices, strengths, limitations, and actionable research gaps [6, 9, 11, 20]. By synthesizing evidence across diverse studies and tasks, the review seeks to align computational advances with domain‐specific biological questions and to support more principled method selection and development [6, 7, 12, 16]. This review contributes to the literature in the following ways:1.Domain‐specific structuring: it organizes the literature into population genetic inference, phylogenetic reconstruction, and language/representation models to reduce cross‐domain conflation and clarify differences in objectives, data representations, and evaluation criteria.2.Comparative methodological synthesis: it summarizes typical inputs, training paradigms (including simulation‐based training where applicable), and evaluation practices in each domain, highlighting how these choices shape reported results [6–8, 14].3.Balanced assessment of strengths and limitations: it consolidates recurring limitations—such as distribution shift, computational demands, and interpretability constraints—without overstating generalizability, and it identifies where evidence is most sensitive to modeling assumptions [6, 15, 17].4.Benchmark‐aware outlook: it highlights the need for transparent reporting and more comparable evaluation designs to improve reproducibility and cross‐study interpretability, and it motivates future directions for robust and biologically credible deployment [9, 18, 20]. 1.Domain‐specific structuring: it organizes the literature into population genetic inference, phylogenetic reconstruction, and language/representation models to reduce cross‐domain conflation and clarify differences in objectives, data representations, and evaluation criteria. Domain‐specific structuring: it organizes the literature into population genetic inference, phylogenetic reconstruction, and language/representation models to reduce cross‐domain conflation and clarify differences in objectives, data representations, and evaluation criteria. 2.Comparative methodological synthesis: it summarizes typical inputs, training paradigms (including simulation‐based training where applicable), and evaluation practices in each domain, highlighting how these choices shape reported results [6–8, 14]. Comparative methodological synthesis: it summarizes typical inputs, training paradigms (including simulation‐based training where applicable), and evaluation practices in each domain, highlighting how these choices shape reported results [6–8, 14]. 3.Balanced assessment of strengths and limitations: it consolidates recurring limitations—such as distribution shift, computational demands, and interpretability constraints—without overstating generalizability, and it identifies where evidence is most sensitive to modeling assumptions [6, 15, 17]. Balanced assessment of strengths and limitations: it consolidates recurring limitations—such as distribution shift, computational demands, and interpretability constraints—without overstating generalizability, and it identifies where evidence is most sensitive to modeling assumptions [6, 15, 17]. 4.Benchmark‐aware outlook: it highlights the need for transparent reporting and more comparable evaluation designs to improve reproducibility and cross‐study interpretability, and it motivates future directions for robust and biologically credible deployment [9, 18, 20]. Benchmark‐aware outlook: it highlights the need for transparent reporting and more comparable evaluation designs to improve reproducibility and cross‐study interpretability, and it motivates future directions for robust and biologically credible deployment [9, 18, 20]. This systematic review follows a structured selection process guided by PRISMA and analyzes 50 studies published between 2016 and 2025. The remainder of the study is organized as follows: Section 2 describes the methodology; Section 3 reports domain‐specific results; Section 4 provides a critical synthesis; Section 5 discusses implications, limitations, and future directions; and Section 6 concludes the review. 2. Methodology This study employed a structured systematic review approach to identify and synthesize research on deep learning applications in evolutionary genomics. The review followed established reporting standards to ensure transparency and reproducibility in study selection. 2.1. Research Question The primary research question guiding this review was How have deep learning methods been applied to identify evolutionary patterns in genomic data, and what are their methodological strengths, limitations, and emerging directions? This question was operationalized through focused search terms targeting deep learning architectures, genomic data analysis, and evolutionary inference tasks. 2.2. Search Strategy and Databases A systematic literature search was conducted across PubMed/MEDLINE, Web of Science Core Collection, Scopus, IEEE Xplore, bioRxiv, arXiv, and Google Scholar to identify relevant studies published between 2016 and 2025. Boolean operators (AND, OR) and subject filters were applied to retrieve studies focusing on deep learning applications in evolutionary genomics. Studies were considered eligible if they applied deep learning architectures (e.g., convolutional, recurrent, residual, generative, or hybrid neural networks) to evolutionary genomics tasks, including phylogenetic reconstruction, natural selection detection, admixture inference, mutation rate prediction, or evolutionary‐oriented functional genomic analysis. Studies relying exclusively on traditional machine learning methods, lacking empirical validation, or not directly addressing evolutionary genomics were excluded. A total of 176 records were identified through database searches, with an additional 140 studies retrieved via citation chaining. After removing 49 duplicates and 18 ineligible records identified by automated screening, 249 articles remained for title and abstract screening. Of these, 118 were excluded for not meeting the inclusion criteria, leaving 131 studies for full‐text assessment. Six articles could not be retrieved, and 75 were excluded based on relevance and methodological quality. Ultimately, 50 studies were included in the final qualitative synthesis. The detailed selection process is illustrated in Figure 1. Systematic literature review with PRISMA format. 2.3. Data Extraction and Coding For each included study, the following information was systematically extracted:•Model architecture•Genomic data representation•Target evolutionary task•Evaluation metrics•Computational characteristics (training and inference considerations)•Interpretability strategies•Reported robustness to evolutionary complexity •Model architecture Model architecture •Genomic data representation Genomic data representation •Target evolutionary task Target evolutionary task •Evaluation metrics Evaluation metrics •Computational characteristics (training and inference considerations) Computational characteristics (training and inference considerations) •Interpretability strategies Interpretability strategies •Reported robustness to evolutionary complexity Reported robustness to evolutionary complexity The analysis was conducted qualitatively and thematically, focusing on methodological trends, comparative strengths, and limitations across studies rather than quantitative meta‐analysis. Risk‐of‐bias assessment results are provided in Appendix A. 3. Result The results of this systematic review synthesize findings from the selected studies and highlight the main methodological and empirical trends in applying deep learning to evolutionary genomics. Overall, the analysis demonstrates significant progress in model accuracy, scalability, and robustness, while also revealing persistent challenges such as interpretability and generalizability. In the following subsections, the results are presented in detail, beginning with a descriptive overview of the included studies and continuing with a critical synthesis and thematic analysis. 3.1. Descriptive Analysis This section outlines the research landscape of the literature on the use of deep learning to discover evolutionary patterns in genomic data and includes a diverse collection of studies that apply different deep learning architectures to evolutionary genomics problems. The reviewed works cover applications such as phylogenetic inference, natural selection detection, logging, mutation rate prediction, and identification of functional genomic elements using convolutional, recurrent, residual, and generative neural networks. The comparative analysis highlights methodological trends, including the integration of simulation‐based training, diverse data representations, and efforts to increase interpretability and robustness. This combination is crucial for addressing research questions about model capabilities, interpretability, and the management of evolutionary complexity, thereby guiding future methodological developments. The included studies can be broadly categorized into three primary domains within evolutionary genomics. The first domain, population genetics, comprises studies focused on detecting natural selection, inferring demographic history, admixture, introgression, and polygenic adaptation using deep learning approaches. The second domain, phylogenetic, includes works aimed at inferring evolutionary relationships, reconstructing tree topologies, estimating substitution patterns, and modeling diversification processes. The third domain, referred to here as language and representation learning models, encompasses studies that apply sequence‐based deep learning architectures—often including self‐supervised or large‐scale models—to learn representations from DNA or protein sequences for tasks such as motif discovery, functional element identification, conservation analysis, and sequence‐to‐function prediction. While these domains share methodological foundations, they address distinct biological questions and rely on different evaluation frameworks. Table 1 shows the descriptive analysis of studies. TABLE 1Descriptive summary of studies.StudyModel accuracyComputational efficiencyData representationInterpretabilityRobustness to evolutionary complexityPrimary domainKorfmann et al. [6]High accuracy in detecting selection and demographyModerate training time; uses simulators for training dataPopulation genetic data encoded as haplotypes and temporal dataModerate interpretability; some efforts on network explainabilityHandles balancing selection and demographic scenarios wellPopulation geneticsYang et al. [21]High accuracy in tracing genealogy origins across diverse geographic populationsRelatively efficient; deep learning framework optimized for large‐scale genomic datasets.Genomic variation encoded using sequence‐based and feature extraction approachesModerate interpretability; results linked to biologically meaningful evolutionary relationshipsStrong robustness; performs well under complex evolutionary patterns such as admixture and mutationPopulation geneticsZhang et al. [10]Effective in inferring introgression and hybridizationEfficient inference on whole genomesSequence alignments from population genomics and assembliesLimited interpretability; focuses on prediction accuracyRobust to admixture and hybridization eventsPopulation geneticsBlischak et al. [12]Accurate hybrid speciation inference using CNNsComputationally efficient with linked SNPs as imagesSNPs encoded as binary images of pairwise divergenceModerate interpretability via model selection among scenariosHandles linked data and admixture effectivelyPopulation geneticsZou et al. [8]Outperforms traditional methods on heterogeneous dataGPU‐accelerated, faster than ML and BI methodsOne‐hot encoded amino acid sequencesGood interpretability; residual networks capture complex signalsRobust to substitution heterogeneity and LBA treesPhylogeneticSuvorov et al. [7]High accuracy in quartet topology inferenceFast inference; robust to gapped and ungapped dataMultiple sequence alignmentsConfidence scores more reliable than bootstrapRobust to bias‐inducing parameter zonesPhylogeneticSilvestro et al. [14]Matches or exceeds likelihood‐based inferenceHighly scalable to large genomic datasetsMultiple sequence alignments directly analyzedProvides biologically interpretable patternsHandles complex rate variation and codon modelsPhylogeneticLambert et al. [22]Comparable accuracy to likelihood methodsFaster by orders of magnitudePhylogenies with trait dataModerate interpretability; regression‐basedEffective for state‐dependent diversification modelsPhylogeneticBurgstaller‐Muehlbacher et al. [23]Comparable to likelihood‐based model selectionSignificantly faster than traditional methodsMultiple sequence alignmentsLimited interpretability; model selection focusRobust across diverse evolutionary modelsPhylogeneticChrysostomou [24]High classification accuracy for host speciesEfficient without sequence alignmentProtein sequences (neuraminidase)Moderate interpretability via classificationHandles viral recombination and evolutionLanguage/representation modelGuo et al. [25]Accurate prediction of subgenome dominance in ancient polyploidsModerate computational demandDNA sequences and methylation dataProvides mechanistic insights into dominanceLimited for neo/synthetic polyploid genomesLanguage/representation modelFang et al. [26]Superior mutation rate prediction performanceEfficient with transfer learningGenomic sequences onlyModerate interpretability; mutation rate profilesRobust across multiple species and mutation contextsLanguage/representation modelTayara & Chong [27]Improved functional quantification of noncoding DNAModerate computational costRaw sequences plus evolutionary informationEnhanced interpretability via motif captureHandles complex regulatory sequence patternsLanguage/representation modelGeorgakilas et al. [28]Outperforms stateof‐the‐art in small RNA loci predictionEfficient scanning of large genomic regionsSequence, conservation, and secondary structureHigh interpretability; iterative background selectionRobust across RNA classes and speciesLanguage/representation modelQin et al. [9]Successful identification of loci under selectionModerate computational requirementsGenotype data predicting traits and locationModerate interpretability; locus contribution analysisRobust to complex spatial selection patternsPopulation geneticsTripathi et al. [11]Outperforms existing methods in polygenic adaptation detectionHighly efficient with sliding window approachGenome‐wide SNP dataModerate interpretability; feature selection includedRobust to demographic history and selection gradientsPopulation geneticsXue et al. [13]High resolution in detecting partial selective sweepsEfficient with CNN and coalescent simulationsPopulation genomic dataLimited interpretability; classification focusedRobust to demographic scenarios and sweep typesPopulation geneticsHejase et al. [15]Accurate selection coefficient inferenceModerate computational demand with LSTMAncestral recombination graph dataModerate interpretability; allele frequency trajectoriesHandles partial soft sweeps and complex selectionPopulation geneticsNguembang et al. [16]∼90% accuracy in selection classificationModerate training on simulated dataGenomic windows from population samplesLimited interpretability; classification accuracy emphasizedRobust to neutral and selection processesPopulation geneticsLi et al. [29]Effective in discriminating conserved sequencesModerate computational costShort DNA segments one‐hot encodedHigh interpretability; motif visualizationCaptures functional conservation patternsLanguage/representation modelShim [30]Captures allele interactions in viral evolutionModerate computational demandTime‐sampled whole‐genome allele frequency dataModerate interpretability via vector clusteringAccounts for recombination and allele interactionsPopulation geneticsSanchez et al. [20]Flexible for demographic inference tasksUser‐friendly with moderate efficiencySNP data with customizable preprocessingModerate interpretability; task‐agnosticRobust across demographic scenariosPopulation geneticsYelmen & Jay [18]Enables data generation and dimensionality reductionComputationally intensive for generative tasksGenomic data modeled in latent spaceModerate interpretability; latent space analysisHandles complex genomic data structuresCross domainWang et al. [31]High performance in plant genome sequence analysisModerate to high computational costMultiomics data including sequence and epigeneticsModerate interpretability; motif mining and designRobust to diverse plant genomic featuresLanguage/representation modelBao et al. [32]Comprehensive coverage of deep learning modelsVaries by model; generally efficientDNA, RNA, and protein dataModerate interpretability; model‐specificAddresses livestock genomics challengesCross domainSun et al. [33]Benchmarking reveals model strengths and limitsEfficient with optimized training strategiesATAC‐seq and single‐cell dataHigh interpretability; sequence‐function modelsSpecialized for single‐cell and enhancer logicLanguage/representation modelHilten et al. [17]Quantitative analysis of interpretability designsModerate computational demandOmics data with diverse architecturesHigh focus on interpretability strategiesAddresses challenges in functional genomicsLanguage/representation modelYue et al. [34]Overview of model strengths for genomic tasksVaries; emphasizes model‐task fitDiverse genomic data typesModerate interpretability; practical considerationsDiscusses challenges and future directionsLanguage/representation modelRouthier and Mozziconacci [19]Reviews deep learning for genome annotationModerate computational requirementsGenomic sequencesModerate interpretability; functional annotationFocus on sequence determinants and synthetic designLanguage/representation modelChandrashekar et al. [35]Highlights AI integration in NGS pipelinesImproves computational time and automationDNA and RNA sequencing dataLimited interpretability focusEnhances large‐scale genomic data processingCross domainAlharbi and Rashid [36]Reviews deep learning in human genomicsVaries; generally efficientHuman genomic dataModerate interpretability; application overviewCovers under‐ and over‐charted genomic areasCross domainGupta et al. [37]Evaluates CNN and RNN for DNA classificationEfficient training and predictionDNA sequencesModerate interpretability; model comparisonAddresses classification accuracy and scalabilityLanguage/representation modelNayak et al. [38]Discusses DL for NGS biomedical researchVaries; supports large datasetsNGS data for disease and drug discoveryModerate interpretability; tool overviewSupports infectious disease and WGS analysisCross domainJunjun et al. [39]Reviews DL‐based variant‐calling methodsEfficient variant detectionSequencing dataModerate interpretability; variant classificationAddresses small and structural variantsCross domainDurge et al. [40]Compares genomic sequence processing modelsVaries; includes CNN, DBN, MLPMultiple genomic data typesModerate interpretability; performance metricsRecommends models for specific genomic tasksCross domainHe et al. [41]Reviews DL in DNA/RNA motif miningModels improve with data sizeSequence data with motif featuresModerate interpretability; motif discoveryHighlights challenges and future directionsCross domainZhang et al. [42]Benchmarks DL algorithms for motif findingEfficient with diverse datasetsChIP‐Seq, CLIP‐Seq, cancer, single‐cell dataHigh interpretability; motif accuracy and scalabilityAddresses noisy data and class imbalanceLanguage/representation modelLiu et al. [43]Reviews DL applications in genomicsModerate computational demandDNA, RNA, protein dataModerate interpretability; workflow focusDiscusses challenges and future perspectivesCross domainMontesinos‐López et al. [44]Reviews DL in genomic selectionVaries; requires large training dataGenomic prediction dataModerate interpretability; prediction accuracyCaptures nonlinear patterns in breeding dataCross domainMeher [45]Highlights DL in bioinformatics big dataEfficient for large datasetsOmics and sequence dataModerate interpretability; application overviewSupports genomic sequence and protein analysisCross domainAbd–Alhalem [46]Surveys DL for DNA sequence classificationEfficient with GPU accelerationDNA sequencesModerate interpretability; architecture overviewDiscusses limitations and suggestionsCross domainShen et al. [47]Brief review of DL in genomic studiesModerate computational costGenomic dataModerate interpretability; challenges notedCovers emerging DL techniquesCross domainAggarwal [48]Reviews ML in genomics data analysisEfficient for large datasetsGenomic dataModerate interpretability; potential and challengesHighlights future directionsCross domainAngermueller et al. [49]Early review of DL in computational biologyModerate computational demandRegulatory genomics and imaging dataModerate interpretability; pitfalls discussedFocus on large data sets and predictionCross domainMin et al. [50]Reviews DL in bioinformaticsVaries by architectureOmics and biomedical dataModerate interpretability; theoretical issuesSuggests future research directionsCross domainWashburn et al. [51]High accuracy in predicting transcript abundance directly from DNA sequences using deep learningModerate efficiency; deep learning models require significant training resources, but inference is relatively fastRaw DNA sequences encoded (e.g., one‐hot encoding) as inputs to neural networks for transcriptome prediction.Limited interpretability; some attempts to link sequence motifs to regulatory features, but largely remains a “black box”Performs well across diverse genomic backgrounds, though robustness depends on training data diversityLanguage/representation modelSidi et al. [52]High accuracy in predicting gene sequences and capturing codon usage patterns with AI modelsModerate efficiency; training deep models on genomic sequence data requires resources, but inference is relatively efficientGene sequences encoded into machine‐readable formats (e.g., codon‐level embedding, sequence encoding) for trainingModerate interpretability; provides insights into codon usage bias, but deeper biological meaning still partially opaqueRobust in handling variability in codon usage across organisms, but sensitive to training dataset compositionLanguage/representation modelArnab et al. [53]High accuracy in detecting natural selection by leveraging spectral analysis of genomic summary statisticsEfficient once summary statistics are computed; spectral analysis reduces dimensionality and computational burdenUses spectral transforms of genomic summary statistics for compact and informative representationRelatively high interpretability; spectral components provide biologically meaningful signals of selectionPerforms well under diverse demographic histories and selection types, though dependent on quality of summary statisticsPopulation geneticsVoznica et al. [54]High accuracy in inferring epidemiological dynamics from phylogenetic treesModerate efficiency; training deep neural networks on large phylogenies can be resource intensivePhylogenetic trees encoded into numerical features suitable for deep learning architecturesModerate interpretability; some effort toward linking deep features to epidemiological meaning, but still partly “black box”Strong robustness; effective in capturing outbreak dynamics across complex transmission and evolutionary scenariosPhylogeneticEneli et al. [55]Promising accuracy in evolutionary inference of malaria vectors using generative modelsModerate to low efficiency; training generative models (e.g., GANs, VAEs) on genomic data can be computationally demandingGenomic sequences encoded for generative modeling, enabling the synthesis of realistic genomic variation patternsModerate interpretability; some insights into generated features, but biological meaning can be opaqueStrong potential robustness, as generative frameworks can capture diverse evolutionary processes, though validation is ongoingLanguage/representation model Descriptive summary of studies. As presented in Table 1, the studies reviewed in this systematic review fall into three primary domains: population genetics, phylogenetic, and language/representation models. The distribution of studies across these domains illustrates the growing breadth of deep learning applications in evolutionary genomics.1.Population genetics studies dominate the table, reflecting the widespread use of deep learning methods to analyze evolutionary processes such as selection, demographic history, and gene flow. These studies primarily focus on using genetic data to infer population structure and detect selection signals, with many utilizing simulation‐based models.2.Phylogenetic studies represent the second‐largest category, focusing on the inference of evolutionary relationships through the reconstruction of phylogenetic trees. These studies highlight how deep learning can improve the accuracy of tree construction and substitution model estimation, particularly when dealing with large and complex datasets.3.Language/representation models appear as a growing area in evolutionary genomics. These studies leverage deep learning models, particularly self‐supervised learning approaches, to learn representations from genomic sequences. While these models do not directly infer evolutionary relationships, they play a critical role in sequence‐to‐function tasks like motif discovery and protein structure prediction. 1.Population genetics studies dominate the table, reflecting the widespread use of deep learning methods to analyze evolutionary processes such as selection, demographic history, and gene flow. These studies primarily focus on using genetic data to infer population structure and detect selection signals, with many utilizing simulation‐based models. Population genetics studies dominate the table, reflecting the widespread use of deep learning methods to analyze evolutionary processes such as selection, demographic history, and gene flow. These studies primarily focus on using genetic data to infer population structure and detect selection signals, with many utilizing simulation‐based models. 2.Phylogenetic studies represent the second‐largest category, focusing on the inference of evolutionary relationships through the reconstruction of phylogenetic trees. These studies highlight how deep learning can improve the accuracy of tree construction and substitution model estimation, particularly when dealing with large and complex datasets. Phylogenetic studies represent the second‐largest category, focusing on the inference of evolutionary relationships through the reconstruction of phylogenetic trees. These studies highlight how deep learning can improve the accuracy of tree construction and substitution model estimation, particularly when dealing with large and complex datasets. 3.Language/representation models appear as a growing area in evolutionary genomics. These studies leverage deep learning models, particularly self‐supervised learning approaches, to learn representations from genomic sequences. While these models do not directly infer evolutionary relationships, they play a critical role in sequence‐to‐function tasks like motif discovery and protein structure prediction. Language/representation models appear as a growing area in evolutionary genomics. These studies leverage deep learning models, particularly self‐supervised learning approaches, to learn representations from genomic sequences. While these models do not directly infer evolutionary relationships, they play a critical role in sequence‐to‐function tasks like motif discovery and protein structure prediction. Overall, Table 1 highlights the diversity of applications in the field, as well as the different biological questions addressed by each domain. The following sections will delve deeper into the strengths, limitations, and trends within each of these domains. 3.2. Critical Analysis and Synthesis While deep learning approaches in evolutionary genomics share computational foundations, the reviewed studies address biologically distinct objectives. To avoid cross‐domain conflation, Table 2 synthesizes domain‐specific objectives, input representations, training paradigms, evaluation criteria, strengths, and limitations across four categories: population genetics, phylogenetic reconstruction, language/representation models, and cross‐domain studies. The analysis below expands directly on these structured dimensions. TABLE 2Critical analysis of reviewed papers.DomainPrimary objective and typical tasksTypical inputs/representationsTypical training paradigmEvaluation focus (common in studies)Key strengths (domain‐specific)Key limitations/challenges (domain‐specific)Population geneticsSelection detection; demographic inference; admixture/introgression; polygenic adaptation; population structure and historySNP windows; haplotypes; allele‐frequency trajectories; divergence matrices; ARG‐derived summaries; population‐genetic feature encodingsPredominantly simulation‐based training (coalescent or forward simulators) + supervised classification/regression; occasional empirical calibrationAccuracy for scenario classification/parameter inference; robustness across simulated scenarios; inference‐time efficiencyFlexible mapping of complex signals; supports diverse encodings; inference often fast once trainedSimulation‐to‐real domain shift; sensitivity to simulator parameters; limited interpretability (often post hoc); high training cost due to simulations and tuningPhylogeneticTree topology inference; substitution‐pattern inference; diversification/state‐dependent models; tree‐based epidemiological dynamicsMultiple sequence alignments (MSAs); alignment‐derived tensors; tree encodings; trait + tree featuresSupervised learning on simulated or benchmarked phylogenetic data; neural networks as approximators of classical inferenceTopology accuracy; robustness to substitution heterogeneity; inference speed comparison; confidence reliabilityPotential acceleration after training; robustness in heterogeneity regimesScalability constraints for large trees; sensitivity to alignment properties and taxon sampling; heterogeneous benchmarking; limited interpretabilityLanguage/representation modelsLearning sequence representations; motif discovery; functional element identification; conservation‐related analysis; sequence‐tofunction prediction; multiomics modelingRaw DNA/protein sequences; embeddings; conservation/structure features; methylation/epigenetic inputs; single‐cell assay dataSelf‐supervised or transfer learning + downstream supervised tasks; large‐scale pretraining where availableTask‐specific predictive metrics; transferability; motif/feature visualizationStrong feature learning; reusable embeddings; useful for sequence‐tofunction pipelinesNot inherently evolutionary inference unless integrated; metrics not directly comparable across domains; interpretability challenges for large models; high training costCross‐domainReviews/benchmarks across genomics tasks; integrative methods; generative models for dimensionality reduction or data synthesisDiverse inputs (sequence, omics layers, summary features); latent representations; mixed‐input pipelinesHeterogeneous: surveys, benchmarks, generative modeling, integrative frameworksStructured comparison; breadth of coverage; interpretability emphasisMaps the methodological landscape; highlights transferable ideasRisk of over‐generalization; heterogeneous evaluation; generative validity and biological plausibility challenges Critical analysis of reviewed papers. Table 2 restructures the critical analysis by the three primary domains used in Table 1 (primary domain), separating population genetic inference, phylogenetic reconstruction, and language/representation learning models to avoid cross‐domain conflation and to clarify domain‐specific strengths, limitations, and evaluation practices. 3.2.1. Population Genetic Inference As summarized in Table 2, studies categorized under population genetics primarily focus on inference tasks such as selection detection, demographic history reconstruction, admixture and introgression detection, polygenic adaptation, and population structure analysis. These studies commonly operate on SNP windows, haplotypes, allele‐frequency trajectories, divergence matrices, ancestral recombination graph (ARG)–derived summaries, or other population‐genetic encodings. Such representations aim to capture structured signals produced by evolutionary processes at the population level. A defining methodological feature in this domain is simulation‐based training. Most models are trained using coalescent or forward‐time simulators to generate labeled evolutionary scenarios, followed by supervised classification or regression. In some cases, limited empirical calibration is incorporated. Evaluation typically emphasizes accuracy in scenario classification or parameter estimation, robustness across simulated evolutionary regimes, and—in some studies—inference‐time efficiency once models are trained. Deep learning offers flexibility in mapping heterogeneous population‐genetic representations to inferred evolutionary outputs. Once trained, models often provide rapid inference relative to repeated likelihood‐based procedures. The central challenge is generalizability. Model performance depends strongly on how well simulated training regimes approximate real evolutionary complexity. Domain shift between simulated and empirical data can reduce external validity. Interpretability is frequently limited to post hoc feature attribution. Moreover, computational cost in this domain is often dominated by simulation generation and hyperparameter tuning, making it important to distinguish between training cost and inference cost. Critical evaluation in this domain should prioritize transparency in simulator parameterization, assessment of robustness under mis‐specified scenarios, and explicit reporting of training versus inference computational demands. 3.2.2. Phylogenetic Reconstruction According to Table 2, the phylogenetic domain includes studies addressing tree topology inference, substitution‐pattern modeling, diversification processes, and tree‐based epidemiological dynamics. Inputs commonly include multiple sequence alignments (MSAs), alignment‐derived tensors, encoded phylogenetic trees, and trait‐plus‐tree features. Many approaches rely on supervised learning using simulated or benchmarked phylogenetic datasets. Neural networks frequently function as approximators or accelerators of classical likelihood‐based inference steps. Evaluation commonly centers on topology accuracy, robustness to substitution heterogeneity or bias‐prone regimes, inference speed comparisons, and the reliability of model‐derived confidence scores. Deep learning models can accelerate inference once trained and may demonstrate robustness in regimes that are challenging for some classical assumptions, such as heterogeneous substitution processes. Limitations are domain‐specific. Some architectures are constrained by fixed input structures (e.g., restricted taxon counts), and scalability to large phylogenies can be limited. Performance may be sensitive to alignment properties and taxon sampling. Additionally, benchmarking protocols vary across studies, complicating direct comparison of reported results. Meaningful evaluation in this domain requires standardized benchmarking conditions, explicit separation of training and inference costs, and careful interpretation of confidence metrics relative to classical phylogenetic uncertainty measures. 3.2.3. Language and Representation Learning Models The third domain in Table 2 includes language and representation models, which focus on learning sequence‐level embedding rather than directly estimating evolutionary parameters or reconstructing trees. Typical tasks include motif discovery, functional element identification, conservation‐related analysis, sequence‐to‐function prediction, and, in some cases, multiomics sequence modeling. Models operate on raw DNA or protein sequences, often encoded via one‐hot representations or embeddings, and may incorporate conservation, structural, epigenetic, or single‐cell assay features. Many approaches employ self‐supervised or transfer learning strategies, followed by downstream supervised tasks. Large‐scale pretraining is used where data availability permits. Evaluation typically relies on task‐specific predictive metrics (e.g., classification or regression accuracy), transferability across datasets or tasks, and interpretability analyses such as motif visualization. This domain benefits from strong feature learning from large sequence corpora and the development of reusable embeddings that support diverse downstream analyses. These models do not inherently constitute evolutionary inference frameworks unless explicitly integrated with evolutionary modeling assumptions. Their evaluation metrics are not directly comparable with those used in population genetics or phylogenetics. Interpretability can be challenging for large models, and pretraining can incur substantial computational costs. Representation models should be interpreted as complementary tools that enhance sequence‐level understanding and downstream analysis, rather than direct substitutes for population‐genetic or phylogenetic inference. 3.2.4. Cross‐Domain Studies and Integration Opportunities Table 2 also identifies cross‐domain studies, including reviews, benchmarks, integrative pipelines, and generative modeling frameworks that span multiple genomic subfields. These works often aim to provide structured comparisons, develop generalized pipelines, or propose methods intended to operate across heterogeneous genomic inputs. Methodologies vary widely, including benchmarking studies, integrative modeling approaches, and generative data augmentation frameworks. Evaluation typically emphasizes breadth of coverage and methodological comparison. Cross‐domain studies help map the methodological landscape and highlight transferable ideas, such as interpretability frameworks, representation learning strategies, and generative augmentation techniques. A key risk is over‐generalization: Heterogeneous evaluation settings and domain‐specific objectives make “one‐size‐fits‐all” conclusions unreliable. Generative approaches additionally require careful validation to ensure biological plausibility. While domain separation is necessary for analytical clarity, methodological transfer between domains—such as combining pretrained representations with simulation‐based frameworks—represents a promising direction for future integration without conflating biological objectives. 3.3. Thematic Review of the Literature While Section 3.2 analyzed the literature according to biological domains, this section reorganizes the same body of work thematically across domains. Therefore, themes discussed here may span more than one domain but are interpreted in light of the distinctions clarified above. Given the interdisciplinary nature of applying deep learning to genomic data analysis to discover evolutionary patterns, previous studies have approached this issue from a variety of approaches. The thematic review allows us to structure the existing literature, identify the main areas of deep learning use, and highlight conceptual trends emerging from previous research. The literature on the applications of deep learning to discover evolutionary patterns in genomic data reveals several dominant themes. The main focus is on the development and evaluation of diverse deep learning architectures—such as convolutional, recurrent, and residual neural networks—designed to infer evolutionary relationships and detect selection signals from genomic sequences. Another key theme involves the integration of simulation‐based training with empirical genomic data to enhance model robustness and interpretability. In addition, the investigation of genomic data representation and preprocessing techniques to improve predictive accuracy appears to be crucial. Challenges related to model interpretability, scalability, and handling of complex evolutionary scenarios such as hybridization and polyploidy are also often discussed, along with the evolving role of generative models and large‐scale data integration. Table 3 provides a thematic review of selected studies, showing that research in this area can be grouped into several main categories. As can be seen, the majority of studies have focused on the development and application of deep learning architectures for evolutionary inference (e.g., phylogenetic tree construction, selection detection, and genomic sequence classification). This suggests that researchers are primarily looking to replace traditional methods with more accurate and scalable architectures. TABLE 3Thematic review of studies.ThemeAppears inTheme descriptionDeep learning architectures for evolutionary inference28/50 papersSeveral studies have developed and tested deep learning models, including CNNs, RNNs, residual networks, and hybrids, for inferring phylogenetic trees, detecting selection, and classifying genomic sequences. These architectures have shown improved accuracy and robustness over traditional approaches, especially in handling nonlinear evolutionary processes and heterogeneous substitution patterns [6–8, 12, 41]. Residual neural networks show particular promise in quadruple tree inference and handling substitution heterogeneity [8]Simulation‐based training and empirical data integration25/50 papersThe use of convolutional and forward simulations to generate training datasets under complex evolutionary scenarios is widespread and facilitates model training for population inference, selection detection, and hybridization analysis. Integration with real genomic data increases the applicability of the model and allows the detection of signals such as recent balancing selection, inbreeding, and multigene adaptation [6, 9, 10, 13, 15]. Such approaches help to overcome the limitations of parametric models and improve generalizabilityGenomic data representation and preprocessing strategies21/50 papersEffective encoding of genomic data, including one‐hot encoding of sequences, matrix representation of SNP data, and inclusion of evolutionary conservation and secondary structure, is a recurring theme. Preprocessing methods significantly affect model performance in predicting regulatory elements, mutation rates, and gene expression from genomic sequences [6, 27–29, 41]. New pipelines automate feature extraction and background selection to optimize training [28]Interpretability and robustness of deep learning models19/50 papersResearch emphasizes the need to clarify how models extract evolutionary signals and ensure robustness to uncertain or noisy training data. Techniques such as network kernel inspection, model weight inspection, and feature importance analysis are used to improve biological interpretability. Challenges include overcoming model opacity and avoiding overfitting, especially in complex evolutionary contexts [6, 8, 17, 41, 42]Handling complex evolutionary phenomena15/50 papersStudies explore the applications of deep learning in complex patterns such as hybridization, introgression, polyploidy, and spatially variable selection. CNN‐based methods have been developed to infer hybrid speciation, while models have been applied to infer subgenome dominance and detect multigene adaptation across populations. This highlights the ability of deep learning to solve lattice and nonlinear evolutionary processes [9–12, 25]Generative models and dimensionality reduction in genomics10/50 papersDeep generative models, including autoencoders and generative adversarial networks (GANs), are emerging as powerful tools for dimensionality reduction, data synthesis, and feature learning in evolutionary genomics. These models help simulate realistic genomic data and extract latent representations that capture evolutionary structure, complementing discriminative approaches [6, 18, 28]Large‐scale, multiomics and cross‐species applications9/50 papersThe application of deep learning extends to the integration of multiomics and cross‐species analyses, such as gene expression landscape prediction across species and cross‐species identification of small RNAs. These approaches enhance the understanding of conserved regulatory elements and evolutionary patterns beyond single‐species contexts [28, 31]Scalability and computational efficiency8/50 papersGiven the computational requirements, several studies emphasize the increased efficiency of GPU‐accelerated deep learning and automated pipelines. Models such as RAISING demonstrate large‐scale genome‐wide implementation with reduced computational time while maintaining high accuracy in detecting adaptive signals [8, 11, 20]Comparative evaluation of discriminative vs. generative models7/50 papersSome works compare discriminative deep learning methods with generative approaches and evaluate their relative strengths in evolutionary inference tasks. While discriminative models excel in classification and prediction, generative models play a role in flexible simulation and hidden feature extraction, suggesting complementary roles [6, 18, 34]Challenges and future directions in deep learning for evolutionary genomics6/50 papersReviews and perspectives underscore challenges such as model interpretability, data imbalance, hyperparameter tuning, and the need for more biologically informed architectures. They advocate for integrating evolutionary knowledge, improving generalizability, and expanding applications to diverse genomic contexts and data types [17, 36, 43, 47] Thematic review of studies. On the other hand, the use of simulated data alongside experimental data is seen as a major recurring theme in studies. This approach allows for the investigation of complex evolutionary scenarios but at the same time presents challenges in terms of generalizability to real data. Another important topic reviewed in the table is data representation and preprocessing methods, which play a key role in improving model performance. The use of innovative representations such as single‐letter encoding of sequences, SNP matrices, and even structural and conservation information has increased the accuracy of the models. The section on interpretability and robustness of models also shows that although some efforts have been made to clarify the decisions of networks, this area remains one of the biggest challenges for deep learning in biology. Other studies have focused on complex evolutionary phenomena (such as hybridization, introgression, and polygenic traits) and have shown that deep learning has the ability to model these nonlinear patterns better than classical statistical methods. Looking ahead, the areas of generative models and dimensionality reduction as well as multiomics data integration represent emerging opportunities that can provide a deeper understanding of evolutionary processes, although they are still in their early stages of development. Overall, Table 3 shows that although a variety of research paths have been pursued in this area, three main axes—developing architectures, simulating and representing data, and improving interpretability—have played the most important role in advancing deep learning in evolutionary genomics. Figure 2 shows the thematic distribution of the selected studies. As is clear, the most research focus has been on the development of deep learning architectures (23.1%), which indicates the importance of innovation in model design for the analysis of evolutionary genomic data. After that, data simulation (18.5%) and model interpretability (16.7%) have the largest contributions; this shows that, in addition to model accuracy, the ability to represent complex evolutionary scenarios and clarify the results are among the main priorities of researchers. Thematic distribution of reviewed papers. Topics such as data representation (13.9%) and complex evolutionary phenomena (11.1%) also constitute an important part of the efforts and have helped to increase the accuracy and applicability of models in real conditions. In contrast, emerging topics such as generative models (9.3%) and multiomics data integration (7.4%) have had a smaller share, but given recent trends, they are expected to receive more attention in the coming years. Overall, this distribution shows that research has gradually moved toward newer and more multilayered topics such as multiomics and generative models, in addition to focusing on improving architectures and data simulation. 3.4. Literature Review Based on Time Although the chronological evolution of the literature reflects a continuous expansion of deep learning in evolutionary genomics, the dominance of specific domains has shifted over time. In early years, foundational sequence‐based and representation‐focused studies were more prominent, whereas later periods saw rapid expansion in population genetic inference and phylogenetic reconstruction. More recently, language and generative models have gained visibility, particularly in multiomics and large‐scale genomic applications. Table 4 shows a description of review papers based on their publication years. TABLE 4Chronological review of literature.Year rangeResearch directionGeneral descriptionDescription by domain2016–2017Foundational deep learning applications in genomicsInitial studies explored the potential of deep learning models such as convolutional neural networks to interpret genomic sequences, identify functional elements, and predict conservation. Emphasis was placed on leveraging sequence motifs and regulatory elements, establishing deep learning as a promising tool for genomics beyond traditional methods. Early frameworks also adapted natural language processing techniques to model viral genome evolutionDuring this foundational phase, most contributions were aligned with what is now categorized as language/representation models, focusing on sequence interpretation and regulatory genomics rather than explicit evolutionary parameter inference2018–2019Deep learning for phylogenetic inference and population geneticsThis period marked the application of deep residual networks and CNNs to infer phylogenetic tree topologies with improved accuracy under complex evolutionary models. Deep learning methods were increasingly used for detecting selective sweeps, mutation rate prediction, and elucidating sequence functions, often incorporating simulation‐based training. Tools began to address challenges such as substitution heterogeneity, incomplete lineage sorting, and partial sweeps with notable advances in model robustness and scalabilityThis period marked a clear rise in population genetic and phylogenetic applications, with simulation‐based inference and topology prediction becoming central research themes2020–2021Expanding deep learning architectures and interpretabilityResearch diversified into integrating recurrent and residual networks for complex evolutionary pattern recognition, including admixture and introgression inference. There was a growing focus on model interpretability, benchmarking of motif‐finding algorithms, and coupling stochastic simulations with deep learning to estimate molecular evolution parameters. Additionally, frameworks for genomic selection and variant calling began to harness deep learning to capture nonlinear relationships in large datasetsPopulation genetic inference became the dominant domain in terms of application diversity, while phylogenetic deep learning matured in methodological sophistication2022–2023Integration of simulation, generative models, and multimodal dataRecent works advanced the coupling of deep learning with stochastic simulations and ancestral recombination graphs to improve selection inference and polygenic adaptation detection. Generative models gained prominence for data augmentation and dimensionality reduction in evolutionary genomics. Cross‐species predictive models and multibranch architectures that integrate sequence, structure, and conservation information emerged, enhancing precision in functional element identification and evolutionary event characterizationThe literature diversified substantially during this period, with population genetics still dominant numerically, but language/representation models gaining momentum through self‐supervised and generative approaches2024–2025Large‐scale models and automation in genomic evolutionary analysesThe latest research emphasizes automation, scalability, and the application of large language models and deep generative models for interpreting complex genomic and epigenomic data. Advances include improved prediction of subgenome dominance, codon usage patterns, and rapid model selection without explicit likelihood calculations. Benchmarking frameworks for interpretability and sequence‐function prediction have been developed to guide future model design and deployment in evolutionary genomicsIn the most recent phase, language/representation models and multiomics integration appear increasingly prominent, reflecting a shift toward large‐scale representation learning and automated genomic analysis frameworks Chronological review of literature. Table 4 provides a chronological review of the research literature, showing how the use of deep learning in evolutionary genomics has evolved over the past decade. In the early years (2016–2017), the focus was on testing the initial capacity of neural networks (especially CNNs) to identify sequence motifs and conserved elements, and this period was more foundational. In 2018–2019, the scope of applications expanded, and deep networks were used for phylogenetic inference and population genetic analysis. At this stage, the use of simulated data to train models and detect natural selection or predict mutation rates became a focus. In 2020–2021, research focused on diversifying architectures (including RNNs and ResNets) as well as increasing the interpretability of models. The use of stochastic simulations combined with deep learning to estimate evolutionary parameters and apply them to new tasks such as selection genomics and variant calling also grew during this period. From 2022 onwards, the integration of experimental and simulated data with generative and multibranch models expanded. Generative models were used in particular to reduce dimensionality and generate new data. At the same time, attention to multiomics and cross‐species prediction increased, indicating a move toward more multilayered and realistic applications. Finally, the last years (2024–2025) are characterized by a focus on scalability, automation, and the use of large language models and advanced generative models. This course shows that the field of deep learning in evolutionary genomics has reached a stage of relative maturity, and its future path will be toward broader applications, more efficient models, and more interpretable frameworks. Figure 3 shows the growth trend in the number of peer‐reviewed articles from 2016 to 2025. As can be seen, in the early years (2016 and 2017), the number of studies was very limited, with only one or two articles published. Since 2018, the trend of research publication has accelerated, and especially since 2019, we have seen a significant increase in the number of articles. Distribution of publication years. The years 2022 and 2023 are considered the peak of research activity in this field, with the highest number of articles published in this period. This growth can be attributed to the advancement of computational tools, the increased access to large‐scale biological data, and the increasing attention to the applications of deep learning in evolutionary genomics. In recent years (2024 and 2025), although the number of articles has decreased slightly compared to the peak in 2023, the level of scientific activity remains high. This indicates the consolidation of deep learning’s position in this field and the gradual movement of research from the quantitative growth stage toward qualitative deepening and more practical studies. Figure 4 shows the trend of the topic distribution of reviewed articles during the years 2016–2025. As can be seen, in the early years (2016–2018) the focus of studies was very limited and mainly focused on model architectures and to some extent data representation. From 2019 onwards, newer areas such as data simulation and interpretability also entered the research literature. Temporal evolution of research themes across the reviewed studies (heatmap). The turning point occurred in the years 2020–2022, when, in addition to an increase in the number of studies, the scope of topics also expanded, and areas such as complex evolutionary phenomena and generative models were also considered. The years 2022 and 2023 show the greatest diversity and volume of studies; in particular, the simultaneous attention to multiomics integration has indicated a movement of research toward multilayered analyses and combined data. In recent years (2024–2025), although the topic diversity has been maintained, the number of studies has slightly decreased. This likely reflects a shift from a quantitative growth phase to a more qualitative focus on emerging topics such as generative models and multiomics. Overall, this heatmap illustrates the evolution of research from a narrow focus on architectures in the early years to a broadening of the spectrum of topics in recent years, a trend that reflects the maturity and sophistication of the use of deep learning in evolutionary genomics. 3.5. Studies Analysis Based on Agreement and Divergence Across the reviewed studies, there is broad consensus that deep learning models, particularly convolutional and residual neural networks, show great potential in uncovering complex evolutionary patterns from genomic data and often outperform traditional methods in terms of accuracy and scalability. Most studies agree on the advantages of simulation‐based training data for capturing evolutionary complexity and the need for interpretable models to provide biological insights. However, there is disagreement about the best architectures for specific tasks, the trade‐offs between discriminative and generative approaches, and how to handle heterogeneous evolutionary processes such as induction and selection. These disagreements often stem from differences in study focus, data types, and evolutionary scenarios modeled, as well as different priorities in computational efficiency versus interpretability. The reviewed literature demonstrates both substantial agreement and notable divergence, but these patterns are domain‐dependent. Therefore, agreement and divergence are examined at two levels: (i) within‐domain consensus and (ii) cross‐domain methodological differences. Table 5 shows the common and different views of research on the application of deep learning in evolutionary genomics. TABLE 5Agreement and divergence across studies.Analytical dimensionPopulation geneticsPhylogeneticLanguage/representationCross‐domain observationCore objectiveParameter and scenario inferenceTree topology and substitution modelingSequence representation and functional predictionObjectives are not directly interchangeableTraining paradigmMostly simulation‐based supervised learningSimulated or benchmarked supervised learningSelf‐supervised + downstream supervised tasksTraining philosophies differ substantiallyEvaluation metricsClassification accuracy, parameter errorTopology accuracy, robustness testsTask‐specific predictive metricsMetrics not directly comparableInterpretabilityPost hoc feature attribution; locus‐level insightsConfidence estimation; limited mechanistic insightMotif/feature visualization; embedding analysisInterpretability definitions differComputational profileHeavy simulation + training; fast inferenceTraining cost varies; potential inference accelerationHigh pretraining cost (large models)Cost structures are domain‐specificMajor limitationDomain shift from simulationScalability to large treesLack of explicit evolutionary groundingNo unified benchmarking framework Agreement and divergence across studies. Across population genetic studies, there is broad agreement that deep learning models—particularly CNN‐based and hybrid architectures—are effective in detecting selection signatures, demographic patterns, and admixture scenarios under complex evolutionary simulations. Most studies rely on simulation‐based training and acknowledge its necessity for modeling nonlinear evolutionary processes. There is also general consensus that inference time is typically efficient once models are trained. However, many studies recognize the risk of domain shift between simulated and empirical data and the limited interpretability of trained networks. Within phylogenetic applications, there is agreement that deep learning can approximate or accelerate topology inference and substitution modeling under certain regimes. Studies converge on the importance of robustness to substitution heterogeneity and biased parameter zones. However, scalability to larger phylogenies and standardized benchmarking remain recurring concerns. Studies in the language modeling domain agree that deep neural architectures are highly effective at learning sequence‐level representations useful for motif discovery, regulatory prediction, and sequence‐to‐function tasks. There is also agreement that transfer learning and large‐scale pretraining enhance generalization across tasks. However, many acknowledge limited direct interpretability and the absence of explicit evolutionary modeling in most architectures. Despite shared computational foundations, significant divergence exists across domains:•Objective divergence: Population genetics seeks parameter inference; phylogenetics seeks tree reconstruction; representation models focus on feature learning.•Evaluation divergence: Metrics are not directly comparable across domains.•Training paradigm divergence: Simulation‐based supervision dominates population genetics; self‐supervised learning is more common in representation models.•Interpretability divergence: Biological interpretability has different meanings across domains.•Computational trade‐offs: Training cost structures differ significantly. •Objective divergence: Population genetics seeks parameter inference; phylogenetics seeks tree reconstruction; representation models focus on feature learning. Objective divergence: Population genetics seeks parameter inference; phylogenetics seeks tree reconstruction; representation models focus on feature learning. •Evaluation divergence: Metrics are not directly comparable across domains. Evaluation divergence: Metrics are not directly comparable across domains. •Training paradigm divergence: Simulation‐based supervision dominates population genetics; self‐supervised learning is more common in representation models. Training paradigm divergence: Simulation‐based supervision dominates population genetics; self‐supervised learning is more common in representation models. •Interpretability divergence: Biological interpretability has different meanings across domains. Interpretability divergence: Biological interpretability has different meanings across domains. •Computational trade‐offs: Training cost structures differ significantly. Computational trade‐offs: Training cost structures differ significantly. These divergences suggest that deep learning in evolutionary genomics should not be treated as a single unified methodological field, but rather as a set of computationally related yet biologically distinct research programs. 4. Theoretical and Practical Implications The integration of deep learning into evolutionary genomics does not represent a single methodological shift, but rather a set of domain‐specific adaptations across population genetics, phylogenetic reconstruction, and sequence representation learning. As structured in Table 2 and discussed in Section 3, these domains differ in biological objectives, input representations, training paradigms, and evaluation criteria. Consequently, their theoretical and practical implications must be interpreted within domain‐appropriate contexts. 4.1. Theoretical Implications 4.1.1. Population Genetic Inference In population genetics, deep learning is predominantly embedded within simulation‐based inference frameworks. By learning mappings between simulated evolutionary scenarios and inference targets (e.g., selection classes or demographic parameters), neural models provide flexible approximations in settings where explicit likelihoods are difficult to specify or computationally expensive [6, 15]. The theoretical implication is not the replacement of population‐genetic modeling but a shift toward amortized inference strategies: Models learn scenario‐to‐parameter mappings during training and subsequently apply them rapidly. However, this paradigm also shifts epistemic responsibility toward the realism and coverage of simulation regimes. If training simulations inadequately represent evolutionary heterogeneity, the learned mapping may lack external validity. Thus, the theoretical contribution of deep learning in this domain depends critically on how simulation assumptions are constructed, validated, and stress‐tested. 4.1.2. Phylogenetic Reconstruction In phylogenetics, deep learning methods are often proposed as accelerators of classical likelihood‐based procedures [8, 14]. Neural networks can learn to predict tree topologies or model characteristics under specific simulated regimes, potentially reducing repeated likelihood evaluations once trained. The theoretical implication here is more computational than epistemological: Deep learning may function as a surrogate for parts of the inference pipeline. However, uncertainty interpretation remains domain‐specific. Confidence scores produced by neural networks are not automatically equivalent to bootstrap or Bayesian posterior support and must be evaluated accordingly. Thus, theoretical progress in this domain hinges on clarifying the relationship between neural confidence and phylogenetic uncertainty. 4.1.3. Language and Representation Learning Models In contrast to direct inference tasks, representation models focus on learning sequence‐level embedding from DNA or protein data [17, 25]. These models often employ self‐supervised or transfer learning strategies and are evaluated on downstream predictive tasks. The theoretical implication is a shift toward representation‐first genomics, where biologically meaningful structure is captured implicitly within embedding. However, such representations do not inherently encode evolutionary mechanisms unless explicitly integrated with evolutionary modeling assumptions. Therefore, theoretical interpretation requires careful distinction between predictive utility and mechanistic inference. 4.2. Practical Implications The practical value of deep learning in evolutionary genomics depends on how models are trained, evaluated, and reported. These considerations vary substantially across domains. 4.2.1. Distinguishing Training Cost From Inference Cost Across domains, deep learning models may provide rapid inference once trained. However, training can be computationally demanding—particularly in population genetics, where large simulation corpora must be generated [6, 9], and in representation learning, where large‐scale pretraining may be required [25]. Practical reporting should therefore distinguish the following:1.Data generation or simulation time2.Model training time3.Inference time4.Compute environment and hardware requirements 1.Data generation or simulation time Data generation or simulation time 2.Model training time Model training time 3.Inference time Inference time 4.Compute environment and hardware requirements Compute environment and hardware requirements Failure to separate these components can produce misleading impressions of scalability or efficiency. 4.2.2. Validation Under Distribution Shift and Evolutionary Complexity Robustness under distribution shift is a domain‐dependent concern. In population genetics, simulation‐to‐reality gaps pose a major challenge [6, 15]. Evaluation should include stress testing under demographic misspecification, recombination variation, and selection heterogeneity. In phylogenetics, benchmarking should span diverse alignment regimes, taxon sampling schemes, and substitution heterogeneity conditions [8, 14]. In representation learning, practical deployment depends on transferability across organisms and tasks, as well as the biological interpretability of learned features [17, 33]. Robust validation frameworks and transparent benchmarking are therefore essential for domain‐appropriate credibility. 4.2.3. Balanced Interpretation and Structured Comparison Credibility improves when studies avoid qualitative claims of “overall superiority” and instead report domain‐specific objectives, input structures, evaluation metrics, and limitations explicitly. Because evaluation metrics differ fundamentally across the three domains—parameter estimation accuracy in population genetics, topology accuracy in phylogenetics, and task‐specific prediction metrics in representation learning—cross‐domain performance comparisons can be misleading. Structured domain‐aligned synthesis, such as that provided in Tables 2, 6, and 7, facilitates more precise interpretation. TABLE 6Domain‐specific limitations of deep learning in evolutionary genomics.DomainLimitation categoryDescriptionRepresentative studiesPopulation geneticsSimulation‐to‐reality gapDependence on simulated training data may reduce generalizability to empirical genomic datasets[6, 9, 15]Population geneticsComputational burdenLarge simulation corpora and hyperparameter tuning increase training cost[11, 20]PhylogeneticFixed input constraintsQuartet‐based or fixed‐topology architectures limit scalability to large trees[7, 8]PhylogeneticBenchmark heterogeneityVariation in alignment properties and evaluation criteria reduces comparability[14]Representation/language modelsLack of evolutionary groundingOptimization focuses on predictive tasks without explicit evolutionary constraints[25, 33]Representation/language modelsHigh pretraining costLarge‐scale models require substantial computational resources[17, 51]Cross‐domainLack of standardized benchmarksThe absence of unified benchmarks limits reproducibility and fair comparison[32, 33] Domain‐specific limitations of deep learning in evolutionary genomics. TABLE 7Domain‐specific gaps and future research directions.DomainResearch gapFuture directionRepresentative referencesPopulation geneticsSimulation‐to‐reality gapHybrid empirical–simulation training and ARG‐based representations[6, 9, 15]PhylogeneticScalability to large treesPermutation‐invariant and graph‐based neural architectures[8, 14]Representation/language modelsLack of evolutionary groundingEvolution‐aware pretraining and mechanistic embedding validation[17, 25, 33]Cross‐domainAbsence of standardized benchmarksUnified benchmarking, statistical testing, transparent reporting[32, 33] Domain‐specific gaps and future research directions. 4.2.4. Adoption in Applied Contexts Applied areas such as conservation genomics, agriculture, and pathogen surveillance may benefit from deep learning when pipelines are reproducible, computational requirements are transparent, and outputs are interpretable for biological decision‐making [9, 20]. In practice, the most sustainable adoption scenarios appear to be those where deep learning complements existing evolutionary modeling workflows—accelerating repeated analyses, enabling screening across large datasets, or providing enriched sequence representations—while final conclusions remain anchored in domain‐appropriate validation and uncertainty assessment. 5. Limitation, Gaps, and Future Research Directions This section outlines key limitations of the current deep learning approaches in evolutionary genomics and suggests promising directions for future research. Based on the findings in Section 3 and the insights from Tables 6 and 7, we also discuss the methodological gaps that hinder the widespread application of these methods. 5.1. Limitations Despite the substantial methodological advances reviewed in the preceding sections, the limitations identified in the literature are not uniform across domains. Rather, they reflect fundamental differences in biological objectives, input representations, training paradigms, and evaluation strategies across population genetic inference, phylogenetic reconstruction, and representation learning frameworks. Accordingly, the following analysis restructures the identified limitations in a domain‐specific manner to clarify methodological boundaries, avoid cross‐domain conflation, and provide a more balanced and analytically grounded assessment of current deep learning applications in evolutionary genomics. 5.1.1. Population Genetic Inference In population genetic applications, deep learning models are predominantly embedded within simulation‐based inference pipelines. A primary limitation is the simulation‐to‐reality gap, where models trained on coalescent or forward simulations may fail to generalize to empirical genomic data if demographic heterogeneity, recombination structure, or selection regimes are insufficiently represented [6, 9, 15]. Model generalizability therefore depends critically on the realism and coverage of simulation regimes. In addition, parameter misspecification and limited robustness to demographic shifts may lead to biased inference. While inference time is often efficient once trained, the generation of large simulated training corpora and hyperparameter optimization impose substantial computational cost [6, 11, 20]. Interpretability also remains limited, as many architectures provide post hoc feature attribution rather than mechanistic evolutionary explanations [17]. 5.1.2. Phylogenetic Reconstruction Within phylogenetic reconstruction, deep learning methods are often designed to approximate or accelerate classical likelihood‐based inference. However, several domain‐specific constraints persist. Some neural architectures rely on fixed input structures (e.g., quartet‐based models), limiting scalability to larger taxon sets [7, 8]. Scalability to large phylogenies and heterogeneous substitution regimes remains a structural challenge. Furthermore, performance sensitivity to alignment properties and taxon sampling complicates benchmarking across studies [14]. Although inference speed may improve once trained, training cost and cross‐study comparability remain important limitations. Interpretability in this domain is frequently limited to confidence estimation rather than biologically grounded mechanistic insight. 5.1.3. Representation and Language Models Representation learning and language modeling frameworks introduce a distinct set of limitations. These models focus on sequence embedding and feature learning rather than direct evolutionary parameter inference. Consequently, a central limitation is the lack of explicit evolutionary grounding, as many architectures optimize predictive performance without incorporating population‐genetic or phylogenetic constraints [25, 33]. Large‐scale pretraining requires substantial computational resources, limiting accessibility and reproducibility. Interpretability remains challenging, particularly for large transformer‐based models, where the biological meaning of learned embeddings is not always transparent [17, 51]. Moreover, evaluation metrics are often task‐specific and not directly comparable with inference‐focused domains. 5.1.4. Cross‐Domain Structural Limitations Across all domains, several structural limitations persist. These include the absence of standardized benchmarking frameworks, heterogeneous evaluation protocols, and inconsistent reporting of computational costs [32, 33]. In addition, data imbalance and limited taxonomic diversity reduce external validity [25, 31]. Without transparent reporting of training conditions, validation partitions, and resource consumption, meaningful cross‐study comparison remains difficult. 5.2. Gaps and Future Research Directions Building upon the domain‐specific limitations outlined in Section 5.1, future research directions must be structured in a manner that reflects the biological and methodological distinctions across population genetic inference, phylogenetic reconstruction, and representation learning frameworks. Rather than treating deep learning in evolutionary genomics as a unified methodological field, future progress depends on domain‐aware innovation, transparent benchmarking, interpretability‐driven validation, and principled integration with evolutionary theory. The following synthesis identifies research gaps and articulates structured, domain‐aligned future directions that directly address current methodological constraints. 5.2.1. Clarifying the Novelty and Conceptual Contribution A central challenge highlighted in the review process concerns the clear articulation of novelty. Future research should move beyond incremental performance gains and instead focus on conceptual innovation. This includes (i) integrating evolutionary theory explicitly into neural architectures; (ii) developing hybrid probabilistic–neural frameworks that combine mechanistic modeling with representation learning; and (iii) designing domain‐adaptive inference strategies that explicitly address simulation‐to‐reality gaps. Such contributions would differentiate future studies from purely predictive deep learning applications and ground them within evolutionary biology. 5.2.2. Population Genetic Inference Future research in population genetics should prioritize robustness under demographic heterogeneity and distribution shift. Hybrid training paradigms that combine empirical genomic datasets with simulated data may reduce the simulation‐to‐reality gap [6, 15]. Integration of ARG‐based representations and hierarchical neural models could improve sensitivity to complex selection scenarios [9, 13]. Additionally, combining classical machine learning feature engineering with deep representation learning may enhance interpretability and reduce overfitting, directly addressing reviewer concerns regarding ML–DL integration. 5.2.3. Phylogenetic Reconstruction In phylogenetic, scalable architectures capable of handling large and variable taxon sets remain a priority. Permutation‐invariant models, graph neural networks operating on tree space, and divide‐and‐conquer inference strategies represent promising directions [8, 14]. Future work should also establish standardized topology benchmarking datasets with frozen training/validation/test splits to ensure comparability and statistical evaluation across methods. Explicit reporting of computational complexity and time efficiency must accompany accuracy metrics. 5.2.4. Representation and Language Models Representation learning frameworks should incorporate explicit evolutionary constraints into pretraining objectives to avoid purely task‐specific optimization [25, 33]. Mechanistic validation of embeddings—through motif enrichment, conservation analysis, and cross‐species transfer testing—will be essential for biological credibility [17]. Efficiency‐oriented transformer compression and resource‐aware training strategies can also improve scalability and reproducibility. 5.2.5. Cross‐Domain Integration Opportunities A major future opportunity lies in cross‐domain methodological transfer. Pretrained sequence embeddings may enhance population genetic inference pipelines, while generative models developed in representation learning could augment phylogenetic simulations [18]. Hybrid probabilistic–deep learning systems combining summary statistics, ARG structures, and learned embedding represent a promising frontier. Such integration directly responds to reviewer requests for clearer articulation of how subfields can inform one another. 5.2.6. Threats to Validity and Benchmark Standardization Future research must also address threats to validity. These include selection bias in training datasets, heterogeneous evaluation metrics across domains, limited taxonomic scope, and insufficient reporting of hyperparameter tuning and computational resources [32, 33]. Standardized benchmarking frameworks incorporating known‐reality simulations, empirical validation datasets, statistical significance testing, and transparent model cards are essential to ensure reproducibility and fair comparison. Calibration metrics and uncertainty estimation should accompany predictive performance. 5.2.7. Analytical Synthesis Across domains, three structural priorities emerge: (i) integration of biological theory into neural architectures; (ii) transparent and statistically grounded benchmarking; and (iii) interpretability‐driven validation. Future work that aligns methodological innovation with evolutionary mechanisms—rather than solely improving predictive accuracy—will likely determine whether deep learning advances from pattern recognition to mechanistic insight. 6. Conclusion This systematic review screened and synthesized 50 studies published between 2016 and 2025 that apply deep learning to the analysis of genomic data for evolutionary questions. To reduce cross‐domain conflation and improve interpretability of the literature landscape, the evidence was organized into three partially overlapping but conceptually distinct domains: (i) population genetic inference; (ii) phylogenetic reconstruction; and (iii) sequence representation learning via DNA/protein language models. This domain‐aligned structure clarifies differences in biological objectives, modeling assumptions, and evaluation strategies that are often treated homogenously in cross‐domain discussions of deep learning in genomics. Across these domains, deep learning operates under distinct inferential logics. In population genetic applications, performance is closely tied to the realism of training distributions and robustness to demographic misspecification. In phylogenetic settings, neural components are frequently integrated into established inferential pipelines, with domain‐specific constraints governing scalability and comparability. In representation learning, sequence embeddings extend analytical flexibility but require explicit evolutionary contextualization before being interpreted as evolutionary inference. These distinctions highlight that methodological advances cannot be evaluated independently of biological framing and validation design. Beyond synthesizing prior results, this review contributes a structural clarification of the field by explicitly aligning methodological categories with biological inference objectives. It consolidates domain‐specific strengths and limitations into a coherent analytical framework and emphasizes several cross‐cutting priorities: the distinction between training and inference scalability, the centrality of distribution shifts in simulation‐driven pipelines, and the need for standardized, statistically grounded benchmarking practices. By foregrounding these structural considerations, the review moves the discussion beyond performance reporting toward principled evaluation. The synthesis further underscores that methodological progress must be assessed in light of reproducibility, transparent reporting, and clearly articulated threats to validity. Without consistent benchmarking protocols and calibration‐aware evaluation, cross‐study comparisons remain difficult and risk overstating generalizability. Taken together, the reviewed evidence supports a balanced interpretation: Deep learning provides flexible modeling capacity that can complement established evolutionary workflows, but it does not replace classical statistical, likelihood‐based, or Bayesian approaches. Its long‐term impact will depend on biologically informed validation, improved interpretability, and reproducible evaluation standards that align predictive performance with evolutionary mechanisms and uncertainty. With these safeguards in place, deep learning is best viewed as a complementary methodological layer that enhances evolutionary genomics across diverse data types and inference tasks [56]. Author Contributions Raha Hassanpour Faramoushjani: visualization, writing–original draft, investigation, conceptualization, methodology, data curation. Sanam Ansari: writing–review and editing, supervision, software, formal analysis. Funding No funding was received for this manuscript. Disclosure A preprint of this article has previously been published (https://doi.org/10.22541/au.174584549.90714325/v1 and https://doi.org/10.22541/au.175795181.17000149/v1). Conflicts of Interest The authors declare no conflicts of interest. Supporting Information Additional supporting information can be found online in the Supporting Information section. Supporting information Supporting Information Appendix A contains the risk‐of‐bias assessment results for the included studies. The supporting file includes the ROBINS‐I assessment and a summary figure showing the distribution of low, moderate, and high risk of bias across the seven ROBINS‐I domains.