Machine Learning in Genetics
- NOTOC**
Machine Learning in Genetics
Machine learning has become an increasingly important part of genetics and genomics, providing computational methods for analyzing biological datasets that are too large, complex, or high-dimensional for many traditional approaches. Applications range from interpreting DNA sequence and predicting gene regulation to identifying disease-associated variants, analyzing populations, classifying single cells, designing genome-editing experiments, and connecting genetic variation with biological traits.
The field includes supervised and unsupervised learning, random forests, gradient-boosted models, convolutional and recurrent neural networks, transformers, graph neural networks, generative models, transfer learning, and other approaches. More recent research increasingly uses large pretrained models that learn representations from DNA sequences, protein sequences, or single-cell expression data and then transfer those representations to multiple biological tasks.
Machine learning does not replace statistical genetics or experimental biology. Instead, it increasingly complements them by finding patterns, generating predictions, prioritizing hypotheses, and integrating many different kinds of genomic and biological information.
From Statistical Genetics to Deep Learning
Early machine-learning applications in genetics focused on classification, dimensionality reduction, feature selection, gene-gene interactions, and prediction from high-dimensional genomic data. These methods provided alternatives or supplements to conventional statistical approaches when relationships among genetic variables were nonlinear or involved large numbers of interacting features.
Deep learning expanded these possibilities by allowing models to learn useful representations directly from DNA sequences, molecular measurements, and other biological data. Convolutional neural networks became particularly important for recognizing sequence motifs and regulatory features, while recurrent networks and later transformer architectures made it possible to model dependencies over increasingly long genomic distances.
Modern genomic machine learning therefore spans a continuum from conventional predictive algorithms to large neural networks capable of learning complex relationships among sequence, regulation, molecular function, and phenotype.
At the same time, research has highlighted recurring methodological problems. Genomic datasets can contain strong population structure, batch effects, confounding variables, related individuals, technical artifacts, and extremely large numbers of features relative to sample size. Overfitting, data leakage, inappropriate cross-validation, and biased feature selection can produce apparently strong performance that does not generalize to independent populations.
Learning Biological Information from DNA Sequence
One of the most important uses of machine learning in genetics is learning relationships directly from DNA sequence.
Models such as DeepBind demonstrated that neural networks could learn sequence preferences associated with DNA- and RNA-binding proteins. DeepSEA extended sequence-based learning to predictions of chromatin features and the functional consequences of noncoding variants. Other systems, including Basset, DanQ, BPNet, Basenji, Sei, ExPecto, Enformer, and related models, have been developed to predict regulatory activity from DNA sequence.
These approaches attempt to learn elements of the regulatory code of the genome. Applications include identifying transcription-factor binding patterns, promoters, enhancers, chromatin-accessible regions, regulatory interactions, and sequence changes likely to alter gene expression.
Increasingly sophisticated models also attempt to account for long-range genomic interactions. Because regulatory elements can influence genes located considerable distances away along the genome, models that examine only short stretches of DNA may miss important relationships. Transformer-based and other long-context architectures are intended to capture more distant dependencies.
Machine learning has also been applied to three-dimensional genome organization. Models such as Akita, DeepC, and Orca predict aspects of chromatin folding and genomic contacts from DNA sequence, allowing researchers to investigate how sequence changes or structural variants might alter genome organization.
Genetic Variant Interpretation
Every human genome contains a very large number of genetic differences, and only a fraction have important functional or medical consequences. Machine learning has therefore become an important tool for prioritizing and interpreting genetic variants.
Methods such as CADD, DANN, REVEL, M-CAP, Eigen, LINSIGHT, FATHMM-MKL, and related systems combine genomic annotations, conservation information, population data, and biological measurements to estimate whether variants are likely to affect molecular function or contribute to disease.
Deep-learning approaches have extended this work. PrimateAI uses information from primate genetic variation to assist prediction of human missense-variant effects. EVE learns evolutionary constraints from protein sequences. Protein language models use patterns learned from large collections of sequences to predict the effects of amino-acid substitutions.
AlphaMissense represents another development in this area, applying an AI model related to protein-structure modeling to predict the likely pathogenicity of missense substitutions across the human proteome.
Machine-learning approaches are also being developed for noncoding variation. This is particularly important because many disease-associated variants lie outside protein-coding regions and may influence enhancers, promoters, chromatin accessibility, transcription-factor binding, splicing, or other forms of gene regulation.
Despite these advances, computational pathogenicity predictions are not equivalent to clinical diagnoses. Studies in the source material show continuing differences between model predictions and clinically curated classifications, reinforcing the importance of experimental evidence, phenotype information, inheritance patterns, and clinical interpretation.
Variant Calling and Genome Sequencing
Machine learning is also used earlier in the genomic-analysis process to determine which genetic variants are actually present in sequencing data.
DeepVariant demonstrated that neural networks could perform SNP and small insertion-and-deletion calling by learning patterns directly from sequencing-read evidence. Related approaches have been developed for long-read sequencing, family-based sequencing, structural variants, and increasingly complete human genome assemblies.
Clairvoyante and Clair3 illustrate the application of deep learning to long-read variant calling. DeepConsensus applies transformer methods to improve sequence accuracy, while DeepTrio incorporates inheritance information from parent-child trios.
These systems illustrate a broader transition in bioinformatics: instead of relying exclusively on manually constructed decision rules, algorithms can learn representations of sequencing evidence from large benchmark datasets.
Accurate benchmarking remains essential. Performance can vary among sequencing technologies, ancestry groups, genomic regions, read lengths, variant types, and reference representations. Long-read sequencing, complete genome assemblies, and pangenome references are creating new opportunities while also requiring new benchmarking methods.
Gene Expression, Splicing, and Regulatory Genomics
Machine learning increasingly links genetic sequence to downstream molecular consequences.
SpliceAI and Pangolin use neural networks to predict RNA splicing from DNA sequence and identify variants that may create, destroy, or modify splice sites. These approaches are particularly useful because disease-associated splicing changes may occur far from conventional splice-site boundaries.
Sequence-to-expression models attempt to predict transcriptional activity directly from genomic DNA. Enformer and related models integrate long-range sequence information to estimate regulatory activity and gene expression. More recent models attempt to predict RNA-sequencing coverage directly from sequence, connecting transcription, splicing, and expression within a common framework.
Machine learning is also used to reconstruct gene regulatory networks. Graph models, attention mechanisms, and deep-learning systems can search for relationships among genes, transcription factors, regulatory elements, and molecular pathways.
These approaches increasingly shift genomic machine learning from simple prediction toward attempts to identify biological mechanisms.
Genomic Foundation Models
A major recent development is the emergence of genomic foundation models.
These systems borrow ideas from large language models. Instead of words and sentences, they learn patterns in DNA, RNA, proteins, or cellular gene-expression profiles. Models can be pretrained on extremely large biological datasets and subsequently adapted to specialized tasks.
DNABERT applied BERT-style pretraining to DNA sequences. Nucleotide Transformer and other genomic language models have expanded this approach to much larger datasets and architectures. Researchers are benchmarking models such as DNABERT-2, Nucleotide Transformer, HyenaDNA, Caduceus, and GROVER across sequence classification, regulatory prediction, variant interpretation, and other tasks.
Evo 2 extends genomic language modeling across organisms and very long sequence contexts, with applications including sequence prediction, variant scoring, and generative genome modeling.
Other work attempts to incorporate naturally occurring genetic variation directly into model pretraining rather than training primarily on reference genomes. This reflects an important limitation of some sequence models: learning the reference genome does not necessarily mean learning how genetic differences among individuals affect biological traits.
Foundation models are also being developed for single-cell biology. Geneformer, scGPT, scBERT, and scFoundation learn representations from large collections of cellular gene-expression measurements that can be transferred to cell classification, perturbation prediction, gene-network analysis, and related tasks.
GWAS, Polygenic Scores, and Complex Traits
Genome-wide association studies identify statistical relationships between genetic variants and traits or diseases. Machine learning is increasingly being investigated as a complement to conventional GWAS methods.
Traditional GWAS often evaluates variants under relatively simple statistical assumptions. Machine-learning systems can instead examine nonlinear relationships, interactions among variants, combinations of polygenic scores, and relationships between genetic and molecular measurements.
Automated machine learning has been explored as a method for discovering informative combinations of genetic variants. Gradient-boosted models and neural networks have also been tested for predicting complex traits from SNPs and polygenic risk scores.
Polygenic scores combine information from large numbers of variants into estimates of genetic predisposition. Machine-learning methods have been investigated as ways to improve prediction or capture nonlinear genetic effects.
An important issue is ancestry transferability. Genetic prediction methods developed primarily using one population may perform differently in other populations because of differences in allele frequencies, linkage disequilibrium, environmental context, and representation in training datasets. Methods such as PRS-CSx and other cross-population approaches attempt to improve prediction across diverse ancestry groups.
The literature also cautions that complex neural networks do not automatically outperform conventional statistical approaches. Careful benchmarking is needed to determine when added model complexity produces meaningful improvements.
Population Genetics, Evolution, and Ancestry
Machine learning has become an important tool in population genetics, where researchers seek to reconstruct demographic history, identify natural selection, estimate population structure, and understand the geographic distribution of genetic variation.
Supervised models can be trained on simulated genomic datasets representing alternative evolutionary scenarios. The models then classify or estimate demographic processes from observed genomic data.
Deep-learning systems have been used to infer population size changes, migration, selective sweeps, geographic ancestry, and other features of population history. Convolutional neural networks can learn patterns directly from genomic alignments rather than depending entirely on manually designed summary statistics.
Systems such as Locator use neural networks to infer geographic location from genetic variation, while Neural ADMIXTURE adapts neural-network techniques to population clustering and ancestry estimation.
Cross-species genomic information can also improve human genetic interpretation. Evolutionary conservation, primate variation, and sequence comparisons among species provide information about which positions in the genome are constrained and therefore potentially functionally important.
These applications require careful interpretation because genetic ancestry, geographic origin, population structure, and socially defined identity are not interchangeable concepts.
Rare Disease and Clinical Genetics
Rare genetic diseases present a particularly difficult machine-learning problem. Individual disorders may affect very few people, limiting the amount of training data available. Patients may also have heterogeneous symptoms, and sequencing often identifies many candidate variants.
Machine-learning systems attempt to combine genomic variation with phenotypic information to prioritize likely disease-causing genes and variants.
Xrare combines genetic evidence with patient phenotypes. PEDIA integrates exome data with facial-image analysis. DeepGestalt uses deep learning to recognize facial patterns associated with genetic syndromes, while GestaltMatcher uses learned facial representations to identify phenotypically similar patients.
Other systems combine rule-based genetic evidence, machine learning, and increasingly language-model techniques to assist variant prioritization.
Clinical applications highlight the difference between prediction and validated diagnosis. Rare-disease datasets are often small, external validation may be limited, and computational predictions may disagree with clinical-grade interpretation. Machine-learning outputs therefore function most safely as evidence to be considered alongside medical, genetic, laboratory, and phenotypic information.
Single-Cell Genomics and Multi-Omics
Single-cell technologies produce molecular measurements for thousands or millions of individual cells, creating datasets particularly suited to machine learning.
Methods such as scVI use deep generative modeling to represent single-cell gene expression, address batch effects, cluster cells, and perform differential analyses. DCA uses neural networks to denoise sparse expression measurements, while SAVER-X applies transfer learning to improve single-cell transcriptomic analysis.
Multi-omics models attempt to integrate different molecular measurements from the same biological system. totalVI jointly models RNA and protein measurements, while MultiVI integrates transcriptomic and chromatin-accessibility data.
Transfer-learning approaches such as scArches allow new datasets to be mapped onto existing cellular reference atlases. Foundation models such as scGPT, Geneformer, scBERT, and scFoundation extend this idea by pretraining on very large collections of cells.
Graph neural networks are also being used to represent biological relationships explicitly. GEARS predicts responses to combinations of genetic perturbations using information about relationships among genes, while other graph models investigate disease modules and gene-disease associations.
The growing integration of transcriptomics, epigenomics, proteomics, spatial measurements, perturbation data, and genetic information is moving machine learning toward increasingly multimodal descriptions of cells.
CRISPR and Genome Editing
Genome editing is another area in which machine learning has practical applications.
CRISPR experiments require selection of guide RNAs that efficiently edit intended genomic locations while minimizing unwanted edits elsewhere. Machine-learning models can be trained on experimental CRISPR datasets to predict both on-target activity and off-target effects.
DeepCRISPR combines neural networks with CRISPR datasets to predict guide efficiency and potential off-target activity. Other models have been developed for Cas9, Cpf1, and related genome-editing systems.
DeepSpCas9 and CRISPRon are examples of models designed to improve prediction of guide activity. Additional systems model mismatches, insertions, deletions, and other sequence configurations associated with off-target editing.
These applications illustrate how machine learning can move beyond genomic interpretation into experimental design, helping researchers select candidate interventions before laboratory testing.
DNA Methylation and Biological Aging
DNA methylation is an epigenetic modification that can be measured at large numbers of genomic positions. Machine-learning algorithms have been used extensively to identify methylation patterns associated with chronological and biological age.
Early epigenetic clocks used penalized statistical learning to estimate age from combinations of methylation measurements. Later systems such as AltumAge and DeepMAge introduced deep neural networks that can model nonlinear relationships among larger numbers of CpG sites.
More recent work applies interpretable machine learning, federated learning, longitudinal prediction, and causal approaches to methylation data.
These systems are being investigated not only for chronological-age prediction but also for biological aging, disease risk, and changes in epigenetic patterns over time.
The field demonstrates a broader characteristic of genomic machine learning: useful prediction can emerge from complex molecular patterns even when the biological meaning of all contributing features is not yet fully understood.
Genomic Prediction in Plants and Animals
Machine learning in genetics extends well beyond human genomics.
Plant and animal breeding programs use genomic prediction to estimate traits from large numbers of genetic markers. Traditional genomic-selection methods rely heavily on statistical and Bayesian models, while newer studies examine neural networks, support-vector machines, convolutional networks, graph models, and multi-omics approaches.
DeepGS uses convolutional neural networks to predict quantitative plant phenotypes from genotype information. Other research evaluates support-vector machines and locally connected networks for prediction in crop and livestock populations.
Modern plant-breeding studies increasingly combine genomic information with phenotypic measurements, environmental variables, multi-omics data, and large-scale datasets.
These applications show how machine learning can support selection of plants or animals with desirable combinations of traits while also providing a testing ground for algorithms intended to model complex genotype-to-phenotype relationships.
Cancer Genetics and Disease Prediction
Machine-learning methods are increasingly applied to cancer genetics and other disease-related genomic datasets.
Models can integrate germline variants, somatic mutations, gene expression, molecular networks, and clinical information to identify biomarkers or predict disease-related outcomes. Disease-specific models may perform differently from general pathogenicity predictors because genetic mechanisms vary among diseases and genes.
Machine-learning systems have also been developed specifically for cancer-predisposition genes and inherited cardiac conditions, illustrating a broader trend toward specialized models trained for particular biological or clinical contexts.
Graph-learning approaches add another dimension by representing genes and diseases as interconnected networks rather than independent variables. This can help models incorporate known molecular interactions while producing predictions about disease modules, candidate genes, or pathways.
Interpretability and Explainable AI
A recurring challenge throughout genomic machine learning is interpretability.
A model that predicts disease risk or regulatory activity accurately may still provide little explanation of why it produced a particular result. In basic research, however, understanding why a prediction occurs can be as important as the prediction itself.
Explainable artificial intelligence attempts to identify sequence motifs, variants, genes, pathways, or molecular features responsible for model predictions. Interpretable architectures may also incorporate known biological organization directly into neural networks.
Methods such as BPNet and ExplaiNN illustrate efforts to design genomic models whose learned sequence features can be examined directly. Other approaches use feature-attribution techniques to investigate complex pretrained models.
Interpretability is particularly important when machine-learning predictions are used to generate biological hypotheses or influence clinical interpretation.
Bias, Generalization, and Data Quality
The source material repeatedly emphasizes that model performance depends on the data used to build and evaluate a system.
Genomic datasets often contain far more variables than samples. They may also contain population structure, unequal ancestry representation, technical batch effects, missing measurements, related individuals, and differences among sequencing platforms.
If training and testing datasets are not separated carefully, information leakage can produce misleading performance estimates. Models may also learn technical artifacts or population-specific patterns rather than general biological relationships.
Ancestry bias is particularly important in human genetics. Models trained disproportionately on some populations may perform less accurately in underrepresented populations. Research on equitable machine learning and cross-population polygenic prediction attempts to address these disparities.
Independent replication and external validation therefore remain central requirements for evaluating genomic machine-learning systems.
Privacy and Federated Learning
Genomic data can contain highly sensitive information about individuals and biological relatives. Large machine-learning models often benefit from combining datasets across hospitals, research centers, and populations, creating tension between data access and privacy.
Federated learning offers one possible approach. Instead of transferring all patient data to one central location, participating institutions can train aspects of a shared model while retaining data locally.
The source material includes federated-learning applications for genetic variant pathogenicity and DNA-methylation age prediction.
Federated approaches do not eliminate every privacy or bias problem, but they illustrate how machine-learning infrastructure may be adapted to settings where centralizing genomic information is difficult or undesirable.
Emerging Directions
Several trends suggest where machine learning in genetics is moving.
Large genomic language models are increasing the amount of sequence context that algorithms can analyze. Multimodal systems are combining genetic information with transcriptomics, epigenomics, proteins, cellular phenotypes, clinical data, and environmental measurements.
Generative models are moving from analyzing existing biological sequences toward proposing or evaluating new sequences. Causal machine-learning approaches seek to distinguish biological relationships that may be causal from correlations in observational datasets.
Transfer learning allows information learned from large datasets to be applied to smaller specialized datasets. Graph neural networks provide ways to incorporate molecular networks and relationships among genes. Federated learning may allow institutions to collaborate without directly pooling sensitive datasets.
These developments are increasingly connecting several previously separate problems: sequence interpretation, gene regulation, variant pathogenicity, disease prediction, cellular modeling, population genetics, experimental design, and genome engineering.
Conclusion
Machine learning has developed from a specialized computational technique into a broad set of tools used throughout genetics and genomics. Its applications now span DNA sequence analysis, regulatory genomics, variant calling, pathogenicity prediction, RNA splicing, GWAS, polygenic scores, population genetics, rare-disease diagnosis, pharmacogenomics, single-cell biology, CRISPR design, epigenetic aging, genomic selection, and disease modeling.
Deep learning has been especially influential because it can learn complex representations directly from genomic and molecular data. Transformers, genomic language models, single-cell foundation models, graph neural networks, and generative systems are extending these capabilities further.
The growth of model complexity, however, makes rigorous evaluation increasingly important. Overfitting, confounding, data leakage, population bias, limited external validation, privacy concerns, and poor interpretability can all reduce the reliability or usefulness of genomic predictions.
The most important role of machine learning in genetics may therefore be neither simply replacing conventional statistical methods nor automating biological research. Its broader value lies in integrating increasingly large and diverse biological datasets, identifying patterns that can be experimentally or clinically investigated, and helping connect genetic sequence with molecular function, biological phenotype, evolution, and disease.
- TOC**
Machine Learning in Genetics
Foundations and General Reviews
1. Obtaining genetics insights from deep learning via explainable artificial intelligence [DOI:10.1038/s41576-022-00532-2 | Gherman Novakovsky et al. | Nature Reviews Genetics | 2023]
Reviews approaches for interpreting deep neural networks in genomics and discusses how explainable AI can turn predictive models into tools for identifying biological mechanisms.
2. Navigating the pitfalls of applying machine learning in genomics [DOI:10.1038/s41576-021-00434-9 | Sean Whalen et al. | Nature Reviews Genetics | 2022]
Explains common sources of bias, data leakage, confounding, inappropriate cross-validation, and misleading performance estimates when applying machine learning to genomic data.
3. Interpretable machine learning for genomics [DOI:10.1007/s00439-021-02387-9 | David S. Watson | Human Genetics | 2022]
Introduces interpretable machine-learning methods for genomic research and examines how explanations can help researchers understand predictions rather than treating models solely as black boxes.
4. Deep learning: new computational modelling techniques for genomics [DOI:10.1038/s41576-019-0122-6 | Gökcen Eraslan et al. | Nature Reviews Genetics | 2019]
Surveys deep-learning applications across genomics and explains how neural networks can model sequence, gene expression, epigenomic, and single-cell data while highlighting interpretability and data-quality challenges.
5. A primer on deep learning in genomics [DOI:10.1038/s41588-018-0295-5 | James Zou et al. | Nature Genetics | 2019]
Provides an accessible introduction to deep learning for genome analysis, with examples involving regulatory genomics, pathogenicity scoring, and variant calling.
6. Avoiding common pitfalls in machine learning omic data science [DOI:10.1038/s41563-018-0241-z | Andrew E. Teschendorff | Nature Materials | 2019]
Discusses overfitting, dimensionality, confounding, biased feature selection, validation design, and other issues that can undermine machine-learning studies using high-dimensional omics data.
7. Machine learning technology in the application of genome analysis: A systematic review [DOI:10.1016/j.gene.2019.04.062 | Jie Wu and Yiqiang Zhao | Gene | 2019]
Reviews machine-learning algorithms used in genome analysis and discusses model selection, data characteristics, and practical considerations for genomic applications.
8. Opportunities and obstacles for deep learning in biology and medicine [PMCID:PMC5938574 | Travers Ching et al. | Journal of the Royal Society Interface | 2018]
Examines how deep learning can exploit expanding genomic and biomedical datasets while addressing limited labeled data, interpretability, privacy, and reproducibility.
9. Deep learning for computational biology [DOI:10.15252/msb.20156651 | Christof Angermueller et al. | Molecular Systems Biology | 2016]
Reviews early applications of deep learning to molecular biology, including regulatory genomics, sequence analysis, gene expression, and biomedical data integration.
10. Machine learning applications in genetics and genomics [DOI:10.1038/nrg3920 | Maxwell W. Libbrecht and William Stafford Noble | Nature Reviews Genetics | 2015]
Reviews major applications of supervised, semi-supervised, unsupervised, generative, and discriminative machine-learning methods to genetic and genomic datasets, including sequence annotation, molecular profiling, and biological network inference.
Genomic Foundation Models and Emerging Methods
11. Genome modelling and design across all domains of life with Evo 2 [DOI:10.1038/s41586-026-10176-5 | Authors listed in article | Nature | 2026]
Presents Evo 2, a large genomic language model trained across diverse organisms for long-context sequence modeling, variant scoring, biological prediction, and generative genome design.
12. Pre-training genomic language model with variants for better modeling functional genomics [DOI:10.1038/s44387-026-00103-4 | Tianyu Liu et al. | npj Artificial Intelligence | 2026]
Investigates whether explicitly incorporating genetic variation during language-model pretraining improves sequence-to-function predictions and functional-genomics modeling.
13. Beyond QTL and GWAS: how deep learning, graph models, and multi-omics are reshaping plant genomic prediction analysis [DOI:10.3389/fgene.2026.1783939 | Tong Wang et al. | Frontiers in Genetics | 2026]
Reviews emerging deep learning, graph neural networks, transfer learning, foundation models, and multi-omics strategies that extend genomic prediction beyond conventional QTL and GWAS approaches.
14. Nucleotide Transformer: building and evaluating robust foundation models for human genomics [DOI:10.1038/s41592-024-02523-z | Authors listed in article | Nature Methods | 2025]
Develops large transformer models pretrained on thousands of human and nonhuman genomes and evaluates their ability to transfer learned sequence representations to multiple genomic tasks.
15. Benchmarking DNA foundation models for genomic and genetic tasks [DOI:10.1038/s41467-025-65823-8 | Haonan Feng et al. | Nature Communications | 2025]
Compares DNABERT-2, Nucleotide Transformer, HyenaDNA, Caduceus, and GROVER on sequence classification, gene-expression prediction, variant interpretation, and genomic-structure tasks.
16. Annotating the genome at single-nucleotide resolution with DNA foundation models [DOI:10.1038/s41592-025-02881-2 | Authors listed in article | Nature Methods | 2025]
Uses pretrained DNA foundation models for multilabel genome annotation at single-nucleotide resolution across multiple regulatory and gene-related element classes.
17. Fast-forwarding plant breeding with deep learning-based genomic prediction [PMCID:PMC12225013 | Authors listed in article | 2025]
Examines the promise and limitations of deep-learning genomic prediction for plant breeding, including multi-omics integration, data augmentation, environmental information, and benchmarking.
18. Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation [DOI:10.1038/s41588-024-02053-6 | Johannes Linder et al. | Nature Genetics | 2025]
Develops a sequence-based model that predicts RNA-sequencing coverage directly from DNA, providing a framework for connecting genomic sequence with transcription, splicing, and gene-expression regulation.
19. Transformers and genome language models [DOI:10.1038/s42256-025-01007-9 | Micaela E. Consens et al. | Nature Machine Intelligence | 2025]
Reviews transformer architectures and genomic language models, examining tokenization, pretraining, long-range sequence modeling, biological interpretation, and applications across functional genomics.
20. Predicting gene expression from DNA sequence using deep learning models [DOI:10.1038/s41576-025-00841-2 | Lucía Barbadilla-Martínez et al. | Nature Reviews Genetics | 2025]
Reviews sequence-to-expression neural networks and discusses what these models reveal about gene regulation, regulatory variants, tissue specificity, and genotype-to-phenotype relationships.
21. MutBERT: probabilistic genome representation improves genomics foundation models [DOI:10.1093/bioinformatics/btaf229 | Weicai Long et al. | Bioinformatics | 2025]
Introduces a masked genomic language model that represents population allele frequencies probabilistically, emphasizing variable genomic positions rather than repeatedly learning nearly identical human genome sequences.
22. A disease-specific language model for variant pathogenicity in cardiac and regulatory genomomics [DOI:10.1038/s42256-025-01016-8 | Huixin Zhan et al. | Nature Machine Intelligence | 2025]
Develops disease-focused language modeling for interpreting genetic variation, illustrating how specialized AI models can incorporate genomic context to improve variant pathogenicity prediction.
23. Multimodal learning for mapping genotype–phenotype dynamics [DOI:10.1038/s43588-024-00765-7 | Farhan Khodaee et al. | Nature Computational Science | 2025]
Explores multimodal machine-learning strategies for connecting genetic information with complex phenotypic measurements and modeling dynamic genotype-to-phenotype relationships.
24. Causal machine learning for single-cell genomics [DOI:10.1038/s41588-025-02124-2 | Alejandro Tejada-Lapuerta et al. | Nature Genetics | 2025]
Reviews how causal inference and machine learning can be combined with single-cell data to distinguish correlations from potential causal biological relationships.
25. Genetic association studies using disease liabilities from deep neural networks [DOI:10.1016/j.ajhg.2025.01.019 | Lu Yang et al. | American Journal of Human Genetics | 2025]
Uses deep neural networks to derive quantitative disease-liability phenotypes that can increase information available for genetic association analyses compared with binary disease classifications.
26. Generative modeling of genomes [DOI:10.1038/s41592-024-02589-9 | Lin Tang | Nature Methods | 2025]
Discusses generative genomic modeling and the ability of large sequence models to learn biological sequence patterns and generate or evaluate genomic DNA.
27. Expanding drug targets for 112 chronic diseases using a machine learning-assisted genetic priority score [DOI:10.1038/s41467-024-53333-y | Robert Chen et al. | Nature Communications | 2024]
Integrates human genetic evidence with machine learning to prioritize genes with potential therapeutic relevance across more than one hundred chronic diseases.
28. A review of machine learning models applied to genomic prediction in animal breeding [DOI:10.3389/fgene.2023.1150596 | Authors listed in article | Frontiers in Genetics | 2023]
Reviews machine-learning approaches for estimating breeding values, genotype imputation, feature selection, and modeling nonlinear genetic effects in livestock.
29. DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome [DOI:10.1093/bioinformatics/btab083 | Yanrong Ji et al. | Bioinformatics | 2021]
Adapts BERT-style transformer pretraining to genomic DNA and demonstrates transfer learning for promoter, splice-site, transcription-factor-binding, and functional-variant prediction.
30. A review of deep learning applications for genomic selection [DOI:10.1186/s12864-020-07319-x | Authors listed in article | BMC Genomics | 2021]
Reviews neural-network approaches to genomic prediction and selection in plant breeding, comparing them with Bayesian and conventional statistical genetics methods.
Regulatory Genomics, Epigenomics and Genome Organization
31. Deep learning for regulatory genomics: a survey of models, challenges, and applications [PMID:41768279 | Author listed in PubMed record | Briefings in Bioinformatics | 2025]
Reviews neural networks, transformers, graph models, chromatin accessibility prediction, transcription-factor binding, RNA regulation, and three-dimensional genomic modeling.
32. Nucleotide-resolution DNA foundation models of prokaryotic genomes [DOI:10.1038/s41588-024-02062-5 | Michael Fletcher | Nature Genetics | 2025]
Discusses large-scale genomic foundation models capable of learning nucleotide-level organization across microbial genomes and transferring that information to genomic prediction tasks.
33. Deep Learning in Gene Regulatory Network Inference: A Survey [DOI:10.1109/TCBB.2024.3442536 | Jiayi Dong, Jiahao Li, and Fei Wang | IEEE/ACM Transactions on Computational Biology and Bioinformatics | 2024]
Surveys deep-learning approaches for reconstructing gene regulatory networks from biological data and compares their scalability, assumptions, and applicability.
34. Improving the performance of supervised deep learning for regulatory genomics using phylogenetic augmentation [DOI:10.1093/bioinformatics/btae190 | Andrew G. Duncan et al. | Bioinformatics | 2024]
Investigates evolutionary sequence augmentation as a way to increase training data and improve supervised deep-learning models of genomic regulation.
35. Enhancing personalized gene expression prediction from DNA sequences using genomic foundation models [DOI:10.1016/j.xhgg.2024.100347 | Pratik Ramprasad et al. | Human Genetics and Genomics Advances | 2024]
Uses pretrained genomic foundation-model embeddings to improve prediction of person-to-person differences in gene expression from individual DNA sequences.
36. Tiberius: end-to-end deep learning with an HMM for gene prediction [DOI:10.1093/bioinformatics/btae685 | Lars Gabriel et al. | Bioinformatics | 2024]
Combines deep neural networks with a hidden Markov model to identify genes and exon-intron structures directly from genomic DNA sequence.
37. Coding genomes with gapped pattern graph convolutional network [DOI:10.1093/bioinformatics/btae188 | Ruo Han Wang et al. | Bioinformatics | 2024]
Represents genomic sequences as gapped-pattern graphs and applies graph convolutional networks to generate useful embeddings for diverse genome-analysis tasks.
38. Deep learning in regulatory genomics: from identification to design [DOI:10.1016/j.copbio.2022.102887 | Xuehai Hu, Alisdair R. Fernie, and Jianbing Yan | Current Opinion in Biotechnology | 2023]
Reviews long-range regulatory sequence models, model interpretation, and emerging uses of deep learning to design synthetic regulatory DNA.
39. ExplaiNN: interpretable and transparent neural networks for genomics [DOI:10.1186/s13059-023-02985-y | Gherman Novakovsky et al. | Genome Biology | 2023]
Develops an interpretable neural architecture that predicts transcription-factor binding and chromatin accessibility while exposing sequence motifs responsible for its predictions.
40. Applications of deep learning in understanding gene regulation [DOI:10.1016/j.crmeth.2022.100384 | Zhongxiao Li et al. | Cell Reports Methods | 2023]
Reviews deep-learning systems for studying transcription, chromatin, regulatory sequences, gene expression, and other components of gene-regulatory biology.
41. A sequence-based global map of regulatory activity for deciphering human genetics [DOI:10.1038/s41588-022-01102-2 | Kathleen M. Chen et al. | Nature Genetics | 2022]
Describes Sei, a deep-learning framework predicting thousands of regulatory profiles and organizing DNA sequences into regulatory classes useful for interpreting disease variants.
42. DeepSTARR predicts enhancer activity from DNA sequence and enables the de novo design of synthetic enhancers [DOI:10.1038/s41588-022-01048-5 | Bernardo P. de Almeida et al. | Nature Genetics | 2022]
Uses deep learning and massively parallel enhancer assays to predict regulatory activity and design synthetic enhancers with desired properties.
43. Sequence-based modeling of three-dimensional genome architecture from kilobase to chromosome scale [DOI:10.1038/s41588-022-01065-4 | Jian Zhou | Nature Genetics | 2022]
Introduces Orca, a sequence-based neural network predicting three-dimensional genome architecture across multiple scales and estimating effects of structural variation.
44. Effective gene expression prediction from sequence by integrating long-range interactions [DOI:10.1038/s41592-021-01252-x | Žiga Avsec et al. | Nature Methods | 2021]
Presents Enformer, a transformer-based genomic model capable of using long-range sequence information to predict gene expression and regulatory effects of variants.
45. Base-resolution models of transcription-factor binding reveal soft motif syntax [DOI:10.1038/s41588-021-00782-6 | Žiga Avsec et al. | Nature Genetics | 2021]
Introduces BPNet and demonstrates how interpretable neural networks can uncover transcription-factor motifs and rules governing cooperative regulatory interactions.
46. A self-attention model for inferring cooperativity between regulatory features [DOI:10.1093/nar/gkab349 | Fahad Ullah and Asa Ben-Hur | Nucleic Acids Research | 2021]
Introduces SATORI, which combines convolutional networks and self-attention to detect interactions among transcription-factor binding sites and other regulatory features.
47. Predicting 3D genome folding from DNA sequence with Akita [DOI:10.1038/s41592-020-0958-x | Geoff Fudenberg, David R. Kelley, and Katherine S. Pollard | Nature Methods | 2020]
Introduces Akita, a convolutional neural network capable of predicting locus-specific three-dimensional genome contact patterns directly from DNA sequence.
48. DeepC: predicting 3D genome folding using megabase-scale transfer learning [DOI:10.1038/s41592-020-0960-3 | Ron Schwessinger et al. | Nature Methods | 2020]
Uses transfer learning to predict large-scale chromatin architecture and estimate how structural and single-base variants may alter genome folding.
49. Deep learning for inferring transcription factor binding sites [PMID:32905524 | Peter K. Koo and colleagues | Current Opinion in Systems Biology | 2020]
Reviews neural-network methods for predicting transcription-factor binding and emphasizes model interpretation as a route to discovering sequence-function relationships.
50. DeePromoter: Robust Promoter Predictor Using Deep Learning [DOI:10.3389/fgene.2019.00286 | Mhaned Oubounyt et al. | Frontiers in Genetics | 2019]
Applies deep neural networks to recognition of promoter regions in human and mouse DNA sequences and illustrates machine learning for functional genome annotation.
51. DeepTACT: predicting 3D chromatin contacts via bootstrapping deep learning [DOI:10.1093/nar/gkz167 | Wenran Li, Wing Hung Wong, and Rui Jiang | Nucleic Acids Research | 2019]
Integrates DNA sequence and chromatin accessibility data to predict promoter-enhancer and promoter-promoter contacts at regulatory-element resolution.
52. Deep learning sequence-based ab initio prediction of variant effects on expression and disease risk [DOI:10.1038/s41588-018-0160-6 | Jian Zhou et al. | Nature Genetics | 2018]
Introduces ExPecto, which predicts tissue-specific gene-expression effects of mutations directly from surrounding DNA sequence.
53. Sequential regulatory activity prediction across chromosomes with convolutional neural networks [DOI:10.1101/gr.227819.117 | David R. Kelley et al. | Genome Research | 2018]
Introduces Basenji, a convolutional neural-network system for predicting regulatory activity across long genomic sequences from DNA alone.
54. Predicting enhancers with deep convolutional neural networks [PMID:29219068 | Bin Liu et al. | BMC Genomics | 2017]
Introduces DeepEnhancer, which predicts enhancer sequences directly from DNA and uses transfer learning to specialize predictions for individual cell types.
55. DeepCpG: accurate prediction of single-cell DNA methylation states using deep learning [DOI:10.1186/s13059-017-1189-z | Christof Angermueller et al. | Genome Biology | 2017]
Uses neural networks to impute missing single-cell DNA-methylation measurements and identify sequence motifs associated with methylation variability.
56. Basset: learning the regulatory code of the accessible genome with deep convolutional neural networks [DOI:10.1101/gr.200535.115 | David R. Kelley, Jasper Snoek, and John L. Rinn | Genome Research | 2016]
Uses convolutional neural networks to learn chromatin-accessibility patterns across cell types and predict how sequence variants affect regulatory activity.
57. DanQ: a hybrid convolutional and recurrent deep neural network for quantifying the function of DNA sequences [DOI:10.1093/nar/gkw226 | Daniel Quang and Xiaohui Xie | Nucleic Acids Research | 2016]
Combines convolutional and recurrent neural networks to learn both sequence motifs and longer-range relationships among regulatory elements in DNA.
58. DeepChrome: deep-learning for predicting gene expression from histone modifications [DOI:10.1093/bioinformatics/btw427 | Ritambhara Singh et al. | Bioinformatics | 2016]
Uses convolutional neural networks to predict gene-expression states from combinations of histone modifications and investigates interactions among epigenetic marks.
59. Predicting the sequence specificities of DNA- and RNA-binding proteins by deep learning [DOI:10.1038/nbt.3300 | Babak Alipanahi et al. | Nature Biotechnology | 2015]
Introduces DeepBind, an influential neural-network system that learns sequence preferences of DNA- and RNA-binding proteins and can assess how sequence variants alter binding.
60. Predicting effects of noncoding variants with deep learning-based sequence model [DOI:10.1038/nmeth.3547 | Jian Zhou and Olga G. Troyanskaya | Nature Methods | 2015]
Introduces DeepSEA, which learns chromatin features directly from DNA sequence and predicts the functional effects of noncoding genetic variants.
Variant Pathogenicity and Functional Interpretation
61. Deep learning approaches for non-coding genetic variant effect prediction: current progress and future prospects [DOI:10.1093/bib/bbae446 | Xiaoyu Wang et al. | Briefings in Bioinformatics | 2024]
Reviews deep-learning methods that predict functional consequences of noncoding variants using epigenomic, bulk sequencing, and single-cell datasets.
62. Accurate proteome-wide missense variant effect prediction with AlphaMissense [DOI:10.1126/science.adg7492 | Jun Cheng et al. | Science | 2023]
Presents AlphaMissense, an AI model related to AlphaFold that predicts likely pathogenicity for missense substitutions throughout the human proteome.
63. Genome-wide prediction of disease variant effects with a deep protein language model [DOI:10.1038/s41588-023-01465-0 | Nadav Brandes et al. | Nature Genetics | 2023]
Uses the ESM1b protein language model to predict functional effects of human missense variants without conventional disease-specific supervised training.
64. Large model predicts variant effects [DOI:10.1038/s41592-023-02050-3 | Lin Tang | Nature Methods | 2023]
Discusses the use of large protein language models to predict effects of human genetic variants and the growing role of self-supervised models in variant interpretation.
65. Predicting variant pathogenicity with AlphaMissense [DOI:10.1038/s41576-023-00668-9 | Kirsty Minton | Nature Reviews Genetics | 2023]
Summarizes AlphaMissense and its potential role in classifying the large number of human missense variants whose clinical significance remains uncertain.
66. StrVCTVRE: a supervised learning method to predict the pathogenicity of human genome structural variants [DOI:10.1016/j.ajhg.2021.12.007 | Adam G. Sharo et al. | American Journal of Human Genetics | 2022]
Applies supervised machine learning to structural-variant features to distinguish potentially pathogenic deletions and duplications from benign variation.
67. Disease variant prediction with deep generative models of evolutionary data [DOI:10.1038/s41586-021-04043-8 | Jonathan Frazer et al. | Nature | 2021]
Introduces EVE, a deep generative model that learns evolutionary constraints from protein sequences and predicts whether human genetic variants are likely pathogenic.
68. MVP predicts the pathogenicity of missense variants by deep learning [DOI:10.1038/s41467-020-20847-0 | Hongjian Qi et al. | Nature Communications | 2021]
Develops a deep residual neural-network framework integrating sequence, evolutionary, and functional annotations to classify missense variants.
69. Improved pathogenicity prediction for rare human missense variants [DOI:10.1016/j.ajhg.2021.08.012 | Yingzhou Wu et al. | American Journal of Human Genetics | 2021]
Introduces VARITY, a gradient-boosted machine-learning framework designed specifically to improve prediction of pathogenicity among rare human missense variants.
70. Disease-specific variant pathogenicity prediction significantly improves variant interpretation in inherited cardiac conditions [DOI:10.1038/s41436-020-00972-3 | Xiaolei Zhang et al. | Genetics in Medicine | 2021]
Introduces CardioBoost, which incorporates disease-specific genetic mechanisms into machine-learning models for interpreting variants associated with inherited cardiomyopathies and arrhythmias.
71. CAPICE: a computational method for Consequence-Agnostic Pathogenicity Interpretation of Clinical Exome variations [DOI:10.1186/s13073-020-00775-w | Shuang Li et al. | Genome Medicine | 2020]
Uses machine learning to score coding and noncoding variants without requiring separate models for different molecular consequence classes.
72. Predicting the clinical impact of human mutation with deep neural networks [DOI:10.1038/s41588-018-0167-z | Laksshman Sundaram et al. | Nature Genetics | 2018]
Introduces PrimateAI, which uses neural networks and variation observed in nonhuman primates to distinguish pathogenic human missense variants from benign changes.
73. ClinPred: Prediction Tool to Identify Disease-Relevant Nonsynonymous Single-Nucleotide Variants [DOI:10.1016/j.ajhg.2018.08.005 | Najmeh Alirezaie et al. | American Journal of Human Genetics | 2018]
Combines machine-learning algorithms, pathogenicity scores, and population allele frequencies to identify disease-relevant nonsynonymous variants.
74. Fast, scalable prediction of deleterious noncoding variants from functional and population genomic data [DOI:10.1038/ng.3810 | Yi-Fei Huang, Brad Gulko, and Adam Siepel | Nature Genetics | 2017]
Introduces LINSIGHT, combining functional genomic data with evolutionary information to estimate the probability that individual noncoding variants are deleterious.
75. REVEL: An Ensemble Method for Predicting the Pathogenicity of Rare Missense Variants [DOI:10.1016/j.ajhg.2016.08.016 | Nilah M. Ioannidis et al. | American Journal of Human Genetics | 2016]
Uses a random-forest ensemble of multiple conservation and functional scores to improve pathogenicity prediction for rare missense variants.
76. M-CAP eliminates a majority of variants of uncertain significance in clinical exomes at high sensitivity [DOI:10.1038/ng.3703 | Karthik A. Jagadeesh et al. | Nature Genetics | 2016]
Develops a machine-learning pathogenicity classifier designed to reduce the number of uncertain missense variants encountered during clinical exome interpretation.
77. A spectral approach integrating functional genomic annotations for coding and noncoding variants [DOI:10.1038/ng.3477 | Iuliana Ionita-Laza et al. | Nature Genetics | 2016]
Introduces Eigen, an unsupervised approach integrating multiple functional annotations to score coding and noncoding genetic variants.
78. DANN: a deep learning approach for annotating the pathogenicity of genetic variants [DOI:10.1093/bioinformatics/btu703 | Daniel Quang, Yifei Chen, and Xiaohui Xie | Bioinformatics | 2015]
Replaces the linear model used by CADD with a deep neural network capable of learning nonlinear relationships among genomic annotations.
79. An integrative approach to predicting the functional effects of non-coding and coding sequence variation [PMID:25583119 | Hashem A. Shihab et al. | Bioinformatics | 2015]
Presents FATHMM-MKL, which uses multiple-kernel learning to integrate functional genomic annotations for predicting effects of coding and noncoding variants.
80. A general framework for estimating the relative pathogenicity of human genetic variants [DOI:10.1038/ng.2892 | Martin Kircher et al. | Nature Genetics | 2014]
Introduces CADD, which combines many genomic annotations through machine learning to produce a general score estimating the deleteriousness of coding and noncoding variants.
Variant Calling and Genome Sequencing
81. Artificial intelligence in variant calling: a review [PMID:40337525 | Authors listed in PubMed record | 2025]
Reviews AI-based callers including DeepVariant, DeepTrio, Clair, Clairvoyante, DNAscope, Medaka, and related approaches across different sequencing technologies.
82. A comprehensive review of deep learning-based variant calling methods [DOI:10.1093/bfgp/elae003 | Junjun Ren et al. | Briefings in Functional Genomics | 2024]
Compares neural-network approaches for calling SNPs, indels, and structural variants and discusses data encoding, model architecture, and training strategies.
83. Deep-learning based representation and recognition for genome variants—from SNVs to structural variants [DOI:10.1093/nsr/nwae335 | Authors listed in article | National Science Review | 2024]
Examines deep-learning variant callers through the paired problems of representing sequencing information and recognizing different classes of genomic variation.
84. Symphonizing pileup and full-alignment for deep learning-based long-read variant calling [DOI:10.1038/s43588-022-00387-x | Ruibang Luo et al. | Nature Computational Science | 2023]
Introduces Clair3, which combines rapid pileup-based calling with neural-network analysis of full read alignments for accurate long-read variant detection.
85. DeepConsensus improves the accuracy of sequences with a gap-aware sequence transformer [DOI:10.1038/s41587-022-01435-7 | Pi-Chuan Chang et al. | Nature Biotechnology | 2023]
Uses a transformer architecture to improve PacBio HiFi consensus sequence accuracy and thereby improve downstream genome assembly and variant calling.
86. Variant calling and benchmarking in an era of complete human genome sequences [DOI:10.1038/s41576-023-00590-0 | Nathan D. Olson et al. | Nature Reviews Genetics | 2023]
Reviews developments in variant calling involving long reads, pangenomes, complete genome assemblies, deep learning, and modern benchmarking standards.
87. DeepTrio: Variant Calling in Families Using Deep Learning [DeepTrio | Alexey Kolesnikov et al. | bioRxiv / Google Research | 2021]
Extends neural-network variant calling to parent-child trios so inheritance information can be learned directly from family sequencing data.
88. Accurate, scalable cohort variant calls using DeepVariant and GLnexus [DOI:10.1093/bioinformatics/btaa1081 | Taedong Yun et al. | Bioinformatics | 2021]
Demonstrates a scalable workflow combining DeepVariant and GLnexus for producing accurate joint variant calls across large sequencing cohorts.
89. A multi-task convolutional deep neural network for variant calling in single molecule sequencing [DOI:10.1038/s41467-019-09025-z | Zilong Luo et al. | Nature Communications | 2019]
Presents Clairvoyante, an early deep-learning system designed to call genomic variants from noisy long-read single-molecule sequencing data.
90. A universal SNP and small-indel variant caller using deep neural networks [DOI:10.1038/nbt.4235 | Ryan Poplin et al. | Nature Biotechnology | 2018]
Introduces DeepVariant, which converts sequencing-read evidence into a representation learned by a neural network to call SNPs and small insertions and deletions.
Single-Cell Genomics, Splicing, Multi-Omics and Gene Networks
91. Gene spatial integration: enhancing spatial transcriptomics analysis via deep learning and batch effect mitigation [DOI:10.1093/bioinformatics/btaf350 | Rian Pratama et al. | Bioinformatics | 2025]
Uses deep learning to integrate spatial-transcriptomic datasets while reducing technical batch effects and preserving biologically meaningful spatial gene-expression patterns.
92. InterVelo: a mutually enhancing model for estimating pseudotime and RNA velocity in multi-omic single-cell data [DOI:10.1093/bioinformatics/btaf500 | Yurou Wang et al. | Bioinformatics | 2025]
Integrates multiple single-cell molecular modalities to jointly estimate cellular developmental trajectories and RNA velocity.
93. scGPT: toward building a foundation model for single-cell multi-omics using generative AI [DOI:10.1038/s41592-024-02201-0 | Haotian Cui et al. | Nature Methods | 2024]
Introduces a generative transformer pretrained on millions of cells that can transfer to cell annotation, perturbation prediction, multi-omics integration, and gene-network inference.
94. Large-scale foundation model on single-cell transcriptomics [DOI:10.1038/s41592-024-02305-7 | Minsheng Hao et al. | Nature Methods | 2024]
Presents scFoundation, a large pretrained model designed to learn transferable representations from massive collections of single-cell gene-expression profiles.
95. Predicting transcriptional outcomes of novel multigene perturbations with GEARS [DOI:10.1038/s41587-023-01905-6 | Yusuf Roohani, Kexin Huang, and Jure Leskovec | Nature Biotechnology | 2024]
Uses graph neural networks and gene relationships to predict transcriptional responses to genetic perturbation combinations not observed during training.
96. GNN4DM: a graph neural network-based method to identify overlapping functional disease modules [DOI:10.1093/bioinformatics/btae573 | András Gézsi and Péter Antal | Bioinformatics | 2024]
Uses graph neural networks to combine genomic information with molecular-interaction networks and discover overlapping gene modules associated with disease.
97. MultiVI: deep generative model for the integration of multimodal data [DOI:10.1038/s41592-023-01909-9 | Tal Ashuach et al. | Nature Methods | 2023]
Integrates transcriptomic and chromatin-accessibility measurements in a shared latent representation, including cells for which one modality is missing.
98. Transfer learning enables predictions in network biology [DOI:10.1038/s41586-023-06139-9 | Christina V. Theodoris et al. | Nature | 2023]
Introduces Geneformer, a transformer pretrained on single-cell transcriptomes that learns gene-network relationships and can predict effects of genetic perturbations.
99. XGDAG: explainable gene–disease associations via graph neural networks [DOI:10.1093/bioinformatics/btad482 | Andrea Mastropietro et al. | Bioinformatics | 2023]
Applies graph neural networks and explainability methods to predict gene-disease associations while identifying biological relationships supporting individual predictions.
100. Predicting RNA splicing from DNA sequence using Pangolin [DOI:10.1186/s13059-022-02664-4 | Tony Zeng and Yang I. Li | Genome Biology | 2022]
Uses deep learning to predict tissue-specific splice-site strength and assess the impact of common, rare, and lineage-specific genetic variants on RNA splicing.
101. Mapping single-cell data to reference atlases by transfer learning [DOI:10.1038/s41587-021-01001-7 | Mohammad Lotfollahi et al. | Nature Biotechnology | 2022]
Introduces scArches, which applies transfer learning to map new single-cell datasets onto large reference atlases without retraining complete models or sharing raw reference data.
102. scBERT as a large-scale pretrained deep language model for cell type annotation of single-cell RNA-seq data [DOI:10.1038/s42256-022-00534-z | Fan Yang et al. | Nature Machine Intelligence | 2022]
Treats gene-expression profiles as a language-modeling problem and applies transformer pretraining to automated cell-type annotation.
103. Joint probabilistic modeling of single-cell multi-omic data with totalVI [DOI:10.1038/s41592-020-01050-x | Adam Gayoso et al. | Nature Methods | 2021]
Introduces totalVI for jointly modeling RNA and protein measurements at single-cell resolution using deep probabilistic methods.
104. Predicting Splicing from Primary Sequence with Deep Learning [DOI:10.1016/j.cell.2018.12.015 | Kishore Jaganathan et al. | Cell | 2019]
Introduces the model underlying SpliceAI, using deep neural networks to predict splice junctions and identify coding and noncoding variants that alter RNA splicing.
105. Learning the language of splicing [DOI:10.1038/s41576-019-0097-3 | Dorothy Clyde | Nature Reviews Genetics | 2019]
Discusses how SpliceAI demonstrated that deep neural networks can identify cryptic splice variants contributing to genetic disorders.
106. scGen predicts single-cell perturbation responses [DOI:10.1038/s41592-019-0494-8 | Mohammad Lotfollahi, F. Alexander Wolf, and Fabian J. Theis | Nature Methods | 2019]
Uses a variational autoencoder to predict how single cells respond to perturbations, including conditions and cell types not directly represented during model training.
107. Data denoising with transfer learning in single-cell transcriptomics [DOI:10.1038/s41592-019-0537-1 | Jingshu Wang et al. | Nature Methods | 2019]
Presents SAVER-X, which combines deep autoencoders and Bayesian modeling to transfer gene-gene relationships between datasets and denoise sparse single-cell expression profiles.
108. Single-cell RNA-seq denoising using a deep count autoencoder [DOI:10.1038/s41467-018-07931-2 | Gökcen Eraslan et al. | Nature Communications | 2019]
Presents DCA, a neural-network approach that models count distributions and nonlinear gene relationships to denoise large single-cell RNA-sequencing datasets.
109. scPred: accurate supervised method for cell-type classification from single-cell RNA-seq data [DOI:10.1186/s13059-019-1862-5 | Jose Alquicira-Hernandez et al. | Genome Biology | 2019]
Uses supervised machine learning to construct predictive models from single-cell transcriptomic data for assigning cells to biologically defined types.
110. Deep generative modeling for single-cell transcriptomics [DOI:10.1038/s41592-018-0229-2 | Romain Lopez et al. | Nature Methods | 2018]
Introduces scVI, a probabilistic deep generative framework for modeling single-cell gene expression, correcting batch effects, clustering cells, and performing differential expression analysis.
GWAS, Polygenic Scores and Complex Traits
111. Exploring the use of machine and deep learning in genome-wide association studies: a comprehensive review [PMID:42477691 | Authors listed in PubMed record | 2026]
Reviews dozens of AI-based approaches designed to move GWAS beyond simple additive linear assumptions while addressing dimensionality and interpretability.
112. Clinical use of polygenic risk scores: current status, barriers and future directions [DOI:10.1038/s41576-025-00900-8 | Iftikhar J. Kullo | Nature Reviews Genetics | 2026]
Reviews the movement of polygenic risk scores toward clinical practice while examining predictive limitations, ancestry transferability, implementation, and interpretation.
113. Performance of deep-learning-based approaches to improve polygenic scores [DOI:10.1038/s41467-025-60056-1 | Martin Kelemen et al. | Nature Communications | 2025]
Benchmarks deep-learning approaches for polygenic prediction and examines whether neural networks can capture genetic effects beyond conventional additive polygenic-score methods.
114. Rethinking GWAS: how lessons from genetic screens and artificial intelligence could reveal biological mechanisms [DOI:10.1093/bioinformatics/btaf153 | Dennis J. Hazelett | Bioinformatics | 2025]
Discusses how AI and richer functional models could help move GWAS interpretation from statistical association toward mechanisms linking variants, genes, and phenotypes.
115. BTS: a scalable Bayesian Tissue Score for prioritizing GWAS variants and their functional contexts across >1000s of omics datasets [DOI:10.1093/bioinformatics/btaf509 | Pavel P. Kuksa et al. | Bioinformatics | 2025]
Integrates GWAS statistics with large collections of functional genomic annotations to identify disease-relevant tissues, regulatory elements, and potentially causal variants.
116. Debiased machine learning for ultra-high dimensional mediation analysis [DOI:10.1093/bioinformatics/btaf282 | Kecheng Wei et al. | Bioinformatics | 2025]
Combines machine learning, cross-fitting, and statistical correction to identify molecular mediators in extremely high-dimensional omics datasets.
117. High-dimensional biomarker identification for interpretable disease prediction via machine learning models [DOI:10.1093/bioinformatics/btaf266 | Yifan Dai et al. | Bioinformatics | 2025]
Introduces HiFIT, combining feature screening and machine-learning importance testing to discover informative biomarkers from high-dimensional molecular datasets.
118. Machine Learning to Advance Human Genome-Wide Association Studies [PMID:38254924 | Authors listed in PubMed record | 2024]
Reviews machine-learning approaches to GWAS, locus discovery, polygenic prediction, epistasis, and emerging genomic foundation models.
119. Preeclampsia prediction using machine learning and polygenic risk scores from clinical and genetic risk factors in early and late pregnancies [DOI:10.1161/HYPERTENSIONAHA.123.21053 | Vesela P. Kovacheva et al. | Hypertension | 2024]
Integrates polygenic risk scores with clinical variables using machine learning to assess prediction of preeclampsia during pregnancy.
120. Automated machine learning for genome wide association studies [PMID:37672022 | Authors listed in PubMed record | 2023]
Develops an AutoML framework for GWAS that searches predictive models, identifies informative variants, estimates out-of-sample performance, and produces interpretable genomic biosignatures.
121. Current approaches to genomic deep learning struggle to fully capture human genetic variation [DOI:10.1038/s41588-023-01517-5 | Ziqi Tang, Shushan Toneyan, and Peter K. Koo | Nature Genetics | 2023]
Discusses evidence that powerful sequence-to-expression models may predict reference-genome regulatory activity better than individual-specific expression differences caused by human variation.
122. Non-linear machine learning models incorporating SNPs and PRS improve polygenic prediction in diverse human populations [DOI:10.1038/s42003-022-03812-z | Authors listed in article | Communications Biology | 2022]
Uses gradient-boosted trees to model nonlinear and interaction effects among SNPs and polygenic risk scores across several complex human traits.
123. Improving polygenic prediction in ancestrally diverse populations [DOI:10.1038/s41588-022-01054-7 | Yunfeng Ruan et al. | Nature Genetics | 2022]
Introduces PRS-CSx, which integrates GWAS summary statistics across ancestry groups to improve cross-population polygenic prediction.
124. Decoding disease: from genomes to networks to phenotypes [DOI:10.1038/s41576-021-00389-x | Authors listed in article | Nature Reviews Genetics | 2021]
Reviews computational and machine-learning strategies connecting disease-associated genetic variants to molecular networks, cellular pathways, and phenotypes.
125. GenNet framework: interpretable deep learning for predicting phenotypes from genetic data [DOI:10.1038/s42003-021-02622-z | Arno van Hilten et al. | Communications Biology | 2021]
Develops biologically informed neural networks that restrict connections according to known genomic structure, enabling phenotype prediction with greater interpretability.
126. Opening the Black Box: Interpretable Machine Learning for Geneticists [DOI:10.1016/j.tig.2020.03.005 | Authors listed in article | Trends in Genetics | 2020]
Reviews methods for understanding which genomic features drive complex machine-learning predictions and how interpretable models can produce biological hypotheses.
127. Polygenic scores via penalized regression on summary statistics [DOI:10.1002/gepi.22050 | Timothy Shin Heng Mak et al. | Genetic Epidemiology | 2017]
Introduces lassosum, which applies penalized regression to GWAS summary statistics and linkage-disequilibrium reference data to construct polygenic scores.
128. Machine learning in genome-wide association studies [DOI:10.1002/gepi.20473 | Silke Szymczak et al. | Genetic Epidemiology | 2009]
Explores penalized regression, ensemble models, and network approaches as complements to conventional single-SNP GWAS methods for identifying complex genetic interactions.
129. Machine learning for detecting gene-gene interactions: a review [DOI:10.2165/00822942-200605020-00002 | Brett A. McKinney et al. | Applied Bioinformatics | 2006]
Reviews neural networks, random forests, multifactor dimensionality reduction, and related approaches for detecting epistatic interactions contributing to complex disease.
Population Genetics, Selection, Evolution and Ancestry
130. Artificial Intelligence in Human Genetics [DOI:10.1146/annurev-genom-111124-032550 | Nadav Brandes | Annual Review of Genomics and Human Genetics | 2026]
Reviews AI approaches for genotype-to-phenotype mapping, causal-gene identification, mutation interpretation, disease-risk estimation, and genomic foundation models.
131. Assessing simulation-based supervised machine learning for demographic parameter inference from genomic data [DOI:10.1038/s41437-025-00773-x | Arnaud Quelin et al. | Heredity | 2025]
Examines the accuracy and robustness of supervised machine-learning approaches trained on simulated genomes for estimating demographic parameters.
132. Integrated genetic and geographic ancestry prediction via large-scale genomic data and machine learning [DOI:10.1186/s40246-025-00837-3 | Jing Chen et al. | Human Genomics | 2025]
Uses large genomic datasets and machine-learning classifiers to infer fine-scale genetic ancestry and geographic origin in East and Southeast Asian populations.
133. AncestryGeni: a novel genetic ancestry classification pipeline for small and noisy sequence data [DOI:10.1093/bioinformatics/btaf391 | Eran Elhaik et al. | Bioinformatics | 2025]
Applies machine-learning classification to ancestry inference when genomic datasets contain limited, sparse, or noisy sequence information.
134. Equitable machine learning counteracts ancestral bias in precision medicine [DOI:10.1038/s41467-025-57216-8 | Leslie A. Smith et al. | Nature Communications | 2025]
Introduces PhyloFrame, an approach designed to reduce ancestry-related predictive bias when machine-learning models are applied to genomic precision-medicine datasets.
135. Harnessing deep learning for population genetic inference [DOI:10.1038/s41576-023-00636-3 | Xin Huang et al. | Nature Reviews Genetics | 2024]
Surveys deep-learning architectures used to infer population structure, demographic events, and natural selection while emphasizing robustness and interpretability.
136. Deep Learning in Population Genetics [DOI:10.1093/gbe/evad008 | Authors listed in article | Genome Biology and Evolution | 2023]
Reviews neural-network applications to demographic history, selection scans, geographic assignment, and interpretation of large population-genomic datasets.
137. Ancestry inference using machine learning [DOI:10.1038/s41592-023-02009-4 | Lin Tang | Nature Methods | 2023]
Describes Neural ADMIXTURE, which adapts interpretable machine learning to ancestry estimation in biobank-scale genomic datasets.
138. Improved pathogenicity prediction using primate genomics [DOI:10.1038/s41588-023-01455-2 | Kyle Vogan | Nature Genetics | 2023]
Discusses PrimateAI-3D and the use of genetic variation across hundreds of primate species to improve interpretation of protein-altering human variants.
139. Neural ADMIXTURE for rapid genomic clustering [DOI:10.1038/s43588-023-00482-7 | Albert Dominguez Mantes et al. | Nature Computational Science | 2023]
Recasts ancestry estimation as an interpretable neural-network problem, providing rapid population-structure inference for biobank-scale genomic datasets.
140. Current progress and open challenges for applying deep learning across the biosciences [DOI:10.1038/s41467-022-29268-7 | Authors listed in article | Nature Communications | 2022]
Reviews deep learning across genome engineering, phylogenetics, biological networks, protein science, and multi-omics while highlighting data and interpretability limitations.
141. Automatic inference of demographic parameters using generative adversarial networks [DOI:10.1111/1755-0998.13386 | Zhanpeng Wang et al. | Molecular Ecology Resources | 2021]
Introduces pg-gan, which uses generative adversarial networks to tune population-genetic simulations and infer demographic parameters from genomic data.
142. Cross-species regulatory sequence activity prediction [PMCID:PMC7392335 | David R. Kelley | PLOS Computational Biology | 2020]
Examines neural-network training across species to improve prediction of functional genomic signals and explore the evolutionary conservation of regulatory sequence rules.
143. Predicting geographic location from genetic variation with deep neural networks [DOI:10.7554/eLife.54507 | C. J. Battey, Peter L. Ralph, and Andrew D. Kern | eLife | 2020]
Introduces Locator, a deep neural network that predicts geographic ancestry from genotype data without requiring predefined population categories.
144. Deep learning for genomics [DOI:10.1038/s41588-018-0328-0 | Nature Genetics Editors | Nature Genetics | 2019]
Describes the potential of deep learning for population genetics, functional genomics, mutation interpretation, crop genetics, evolutionary studies, and clinical genomics.
145. The unreasonable effectiveness of convolutional neural networks in population genetic inference [DOI:10.1093/molbev/msy224 | Lex Flagel, Yaniv Brandvain, and Daniel R. Schrider | Molecular Biology and Evolution | 2019]
Demonstrates that convolutional neural networks can learn directly from population-genomic alignments and perform demographic and selection inference with little hand-crafted feature engineering.
146. Supervised Machine Learning for Population Genetics: A New Paradigm [DOI:10.1016/j.tig.2017.12.005 | Daniel R. Schrider and Andrew D. Kern | Trends in Genetics | 2018]
Reviews supervised learning in evolutionary genomics and explains how simulated genomic datasets can train classifiers for selection and demographic inference.
147. Localization of adaptive variants in human genomes using averaged one-dependence estimation [DOI:10.1038/s41467-018-03100-7 | Lauren Alpert Sugden et al. | Nature Communications | 2018]
Introduces SWIF(r), a supervised machine-learning classifier integrating multiple population-genetic statistics to identify variants affected by natural selection.
148. diploS/HIC: an updated approach to classifying selective sweeps [DOI:10.1534/g3.118.200262 | Daniel R. Schrider and Andrew D. Kern | G3: Genes, Genomes, Genetics | 2018]
Uses supervised machine learning and spatial patterns in population-genetic summary statistics to distinguish different forms of selective sweeps.
149. Deep Learning for Population Genetic Inference [DOI:10.1371/journal.pcbi.1004845 | Sara Sheehan and Yun S. Song | PLOS Computational Biology | 2016]
Demonstrates likelihood-free deep-learning inference of population-genetic parameters, including demographic history and natural selection.
Rare Disease, Clinical Genetics and Pharmacogenomics
150. Disease- and gene-specific deep learning for pathogenicity prediction of rare missense variants in cancer predisposition genes [DOI:10.1186/s13040-026-00533-5 | Da-Bin Lee, Hyun-Uk Kang, and Kyu-Baek Hwang | BioData Mining | 2026]
Develops gene- and disease-specific neural models for classifying uncertain missense variants in hereditary cancer-predisposition genes.
151. AI in Genomics: From Variant Calling to Multi-Omics Integration [DOI:10.1002/bies.70160 | Hina Sultana et al. | BioEssays | 2026]
Reviews AI applications to variant calling, gene expression, single-cell transcriptomics, CRISPR optimization, and integration of multiple genomic and molecular data types.
152. aiDIVA – hybrid AI for rare disease diagnostics using evidence-based, machine learning and language models [DOI:10.1038/s41525-026-00611-x | Dominic Boceck et al. | npj Genomic Medicine | 2026]
Combines evidence-based scoring, machine learning, and language-model techniques in a hybrid system designed to prioritize variants in rare-disease diagnostics.
153. Advancing genome-based precision medicine: a review on machine learning applications for rare genetic disorders [PMID:40668553 | Authors listed in PubMed record | 2025]
Reviews machine-learning approaches to disease classification, biomarker discovery, therapeutic optimization, and risk assessment using genomic information from rare disorders.
154. Machine learning models for pharmacogenomic variant effect predictions – recent developments and future frontiers [DOI:10.1080/14622416.2025.2504863 | Roman Tremmel et al. | Pharmacogenomics | 2025]
Reviews machine-learning prediction of functional effects for pharmacogenomic variants, including rare variants, protein models, evolutionary information, and variant interactions.
155. Machine learning models for pharmacogenomic variant effect predictions [PMID:40401639 | Roman Tremmel et al. | Pharmacogenomics | 2025]
Provides a PubMed-indexed overview of modern ML approaches for translating drug-related genetic variation into functional predictions relevant to precision prescribing.
156. Discordance between a deep learning model and clinical-grade variant pathogenicity classification in a rare disease cohort [DOI:10.1038/s41525-025-00480-w | Sek Won Kong et al. | npj Genomic Medicine | 2025]
Compares deep-learning pathogenicity predictions with clinically curated variant classifications and highlights circumstances in which computational and clinical interpretations diverge.
157. Federated learning for the pathogenicity annotation of genetic variants in multi-site clinical settings [DOI:10.1093/bioinformatics/btaf523 | Nigreisy Montalvo et al. | Bioinformatics | 2025]
Applies federated machine learning so multiple clinical institutions can collaborate on variant pathogenicity prediction without centralizing sensitive genomic datasets.
158. Explicable prioritization of genetic variants by integration of rule-based and machine learning algorithms for diagnosis of rare Mendelian disorders [DOI:10.1186/s40246-024-00595-8 | Ho Heon Kim et al. | Human Genomics | 2024]
Integrates machine-learning predictions with interpretable rule-based evidence to prioritize variants for Mendelian disease diagnosis.
159. Phenotypic evaluation of deep learning models for classifying germline variant pathogenicity [DOI:10.1038/s41698-024-00710-x | Ryan D. Chow et al. | npj Precision Oncology | 2024]
Evaluates deep-learning pathogenicity scores by asking whether their predictions correspond to expected disease phenotypes in individuals carrying germline variants.
160. Machine learning in rare disease [DOI:10.1038/s41592-023-01886-z | Jineta Banerjee et al. | Nature Methods | 2023]
Discusses opportunities and difficulties created by small sample sizes, heterogeneous phenotypes, genomic data, and high-dimensional molecular profiling in rare-disease machine learning.
161. A systematic review on machine learning approaches in the diagnosis and prognosis of rare genetic diseases [DOI:10.1016/j.jbi.2023.104429 | P. Roman-Naranjo et al. | Journal of Biomedical Informatics | 2023]
Reviews sequencing-based machine-learning approaches for diagnosis, variant prioritization, inheritance analysis, and patient stratification in rare genetic disease.
162. Artificial Intelligence in the Genetic Diagnosis of Rare Disease [DOI:10.1016/j.cll.2022.09.023 | Kiely N. James et al. | Clinics in Laboratory Medicine | 2023]
Reviews how AI can assist phenotype recognition, genomic variant prioritization, candidate-gene identification, and interpretation of rare inherited disorders.
163. GestaltMatcher facilitates rare disease matching using facial phenotype descriptors [DOI:10.1038/s41588-021-01010-x | Tzung-Chien Hsieh et al. | Nature Genetics | 2022]
Uses deep facial embeddings to match patients with similar dysmorphic features, including disorders represented by very small numbers of diagnosed individuals.
164. Phenotype-aware prioritisation of rare Mendelian disease variants [DOI:10.1016/j.tig.2022.07.002 | Catherine Kelly et al. | Trends in Genetics | 2022]
Reviews computational tools that combine patient phenotypes with genomic variants to prioritize likely molecular diagnoses in rare inherited disease.
165. Machine Learning: An Overview and Applications in Pharmacogenetics [DOI:10.3390/genes12101511 | Giovanna Cilluffo et al. | Genes | 2021]
Reviews supervised and unsupervised machine learning used to study genetic determinants of drug response, including antidepressant, cancer, and warfarin pharmacogenetics.
166. The use of machine learning in rare diseases: a scoping review [DOI:10.1186/s13023-020-01424-6 | Julia Schaefer et al. | Orphanet Journal of Rare Diseases | 2020]
Reviews hundreds of rare-disease machine-learning studies involving diagnosis, prognosis, imaging, clinical data, and omics and identifies limitations in external validation.
167. Xrare: a machine learning method jointly modeling phenotypes and genetic evidence for rare disease diagnosis [DOI:10.1038/s41436-019-0439-8 | Qigang Li et al. | Genetics in Medicine | 2019]
Combines patient phenotype information with variant-level genetic evidence to prioritize disease-causing variants during rare-disease diagnosis.
168. PEDIA: prioritization of exome data by image analysis [DOI:10.1038/s41436-019-0566-2 | Tzung-Chien Hsieh et al. | Genetics in Medicine | 2019]
Combines deep-learning analysis of facial phenotypes with exome variant information to improve prioritization of candidate disease genes.
169. Identifying facial phenotypes of genetic disorders using deep learning [DOI:10.1038/s41591-018-0279-0 | Yaron Gurovich et al. | Nature Medicine | 2019]
Introduces DeepGestalt, a facial-analysis neural network trained to recognize phenotypic patterns associated with hundreds of genetic syndromes.
CRISPR and Machine Learning for Genome Editing
170. Prediction of CRISPR-Cas9 on-target activity based on a hybrid neural network [DOI:10.1016/j.csbj.2025.05.001 | Chuxuan Li et al. | Computational and Structural Biotechnology Journal | 2025]
Introduces CRISPR_HNN, combining convolution, attention, and recurrent neural-network components to model local and long-range features affecting guide activity.
171. Advancing CRISPR with deep learning: a comprehensive review of models and databases [DOI:10.1016/j.omtn.2025.102691 | Roghayyeh Alipanahi et al. | Molecular Therapy – Nucleic Acids | 2025]
Reviews neural-network models for CRISPR on-target and off-target prediction together with datasets and databases supporting machine-learning genome-editing research.
172. Generating, modeling and evaluating a large-scale set of CRISPR/Cas9 off-target sites with bulges [DOI:10.1093/nar/gkae428 | Ofir Yaish and Yaron Orenstein | Nucleic Acids Research | 2024]
Expands training data for CRISPR off-target sites containing insertions and deletions and evaluates machine-learning approaches for these difficult sequence configurations.
173. Enhancing CRISPR-Cas9 gRNA efficiency prediction by data integration and deep learning [DOI:10.1038/s41467-021-23576-0 | Xi Xiang et al. | Nature Communications | 2021]
Introduces CRISPRon, trained on integrated experimental datasets to improve prediction of SpCas9 guide-RNA on-target efficiency.
174. SpCas9 activity prediction by DeepSpCas9, a deep learning-based model with high generalization performance [DOI:10.1126/sciadv.aax9249 | Hui Kwon Kim et al. | Science Advances | 2019]
Develops a deep-learning predictor of SpCas9 editing efficiency designed to generalize across cell types and experimental datasets.
175. DeepCRISPR: optimized CRISPR guide RNA design by deep learning [DOI:10.1186/s13059-018-1459-4 | Guohui Chuai et al. | Genome Biology | 2018]
Combines deep learning with large CRISPR datasets to predict both guide-RNA efficiency and potential off-target activity.
176. Deep learning improves prediction of CRISPR–Cpf1 guide RNA activity [DOI:10.1038/nbt.4061 | Hui Kwon Kim et al. | Nature Biotechnology | 2018]
Uses deep learning to derive sequence rules governing Cpf1 guide-RNA activity and improve selection of efficient genome-editing targets.
177. Prediction of off-target activities for the end-to-end design of CRISPR guide RNAs [DOI:10.1038/s41551-017-0178-6 | Jennifer Listgarten et al. | Nature Biomedical Engineering | 2018]
Develops machine-learning models for CRISPR off-target prediction and integrates them with on-target activity estimates for guide-RNA design.
178. Off-target predictions in CRISPR-Cas9 gene editing using deep learning [DOI:10.1093/bioinformatics/bty554 | Jiecong Lin and Ka-Chun Wong | Bioinformatics | 2018]
Applies convolutional and feed-forward neural networks to guide-RNA and DNA sequence pairs for predicting CRISPR-Cas9 off-target mutations.
179. Optimized sgRNA design to maximize activity and minimize off-target effects of CRISPR-Cas9 [DOI:10.1038/nbt.3437 | John G. Doench et al. | Nature Biotechnology | 2016]
Uses large experimental datasets and predictive modeling to develop guide-RNA design rules balancing editing efficiency and specificity.
DNA Methylation, Epigenetic Prediction and Biological Aging
180. A robust computational framework for methylation age and disease-risk prediction based on pairwise learning [DOI:10.1038/s43588-025-00939-x | Yu Zhang et al. | Nature Computational Science | 2026]
Introduces MAPLE, using pairwise machine learning to improve robustness of methylation-based age and disease-risk predictions across technical platforms and cohorts.
181. Privacy-preserving federated unsupervised domain adaptation with application to age prediction from DNA methylation data [DOI:10.1093/bioinformatics/btaf465 | Cem Ata Baykara et al. | Bioinformatics | 2025]
Combines federated learning and domain adaptation to build methylation-based age predictors across datasets without pooling sensitive individual-level molecular data.
182. Longitudinal prediction of DNA methylation to forecast epigenetic outcomes [DOI:10.1016/j.ebiom.2025.105709 | Arthur Leroy et al. | eBioMedicine | 2025]
Uses probabilistic machine learning to forecast future methylation profiles from earlier measurements and derive future epigenetic-age estimates.
183. Exploring the correlation between DNA methylation and biological age using an interpretable machine learning framework [DOI:10.1038/s41598-024-75586-9 | Sheng Zhou et al. | Scientific Reports | 2024]
Applies interpretable machine learning to DNA methylation data to identify CpG patterns associated with biological and chronological aging.
184. Causality-enriched epigenetic age uncouples damage and adaptation [DOI:10.1038/s43587-023-00557-0 | Kejun Ying et al. | Nature Aging | 2024]
Combines large-scale genetic evidence with machine-learning epigenetic clocks to distinguish methylation signals associated with potentially damaging and adaptive aging processes.
185. A pan-tissue DNA-methylation epigenetic clock based on deep learning [DOI:10.1038/s41514-022-00085-y | Lucas Paulo de Lima Camillo et al. | npj Aging | 2022]
Introduces AltumAge, a neural-network epigenetic clock trained on methylation measurements from multiple tissues and a large number of CpG sites.
186. DeepMAge: A Methylation Aging Clock Developed with Deep Learning [DOI:10.14336/AD.2020.1202 | Fedor Galkin et al. | Aging and Disease | 2021]
Uses deep neural networks to estimate biological age from DNA methylation while examining nonlinear relationships among age-associated CpG sites.
187. Chronological age prediction based on DNA methylation: massive parallel sequencing and random forest regression [DOI:10.1016/j.fsigen.2017.07.015 | Jana Naue et al. | Forensic Science International: Genetics | 2017]
Uses random-forest regression and targeted methylation sequencing to predict chronological age from a selected panel of CpG markers.
188. DNA methylation age of human tissues and cell types [DOI:10.1186/gb-2013-14-10-r115 | Steve Horvath | Genome Biology | 2013]
Develops the widely used pan-tissue epigenetic clock using penalized machine learning to predict chronological age from DNA-methylation patterns.
189. Genome-wide methylation profiles reveal quantitative views of human aging rates [DOI:10.1016/j.molcel.2012.10.016 | Gregory Hannum et al. | Molecular Cell | 2013]
Uses genome-wide DNA methylation and predictive modeling to build an age estimator and investigate inter-individual variation in epigenetic aging.
Genomic Selection, Cancer Genetics and Graph Learning
190. GPS: Harnessing data fusion strategies to improve the accuracy of machine learning-based genomic and phenotypic selection [DOI:10.1016/j.xplc.2025.101416 | Hongshan Wu et al. | Plant Communications | 2025]
Combines genomic and phenotypic information through several data-fusion strategies and compares statistical, machine-learning, and deep-learning approaches for plant selection.
191. Expanding genomic prediction in plant breeding: harnessing big data, machine learning, and advanced software [DOI:10.1016/j.tplants.2024.12.009 | José Crossa et al. | Trends in Plant Science | 2025]
Reviews how machine learning, environmental information, high-throughput phenotyping, and increasingly large genomic datasets are reshaping genomic prediction in plant breeding.
192. Disease-Specific Prediction of Missense Variant Pathogenicity with DNA Language Models and Graph Neural Networks [DOI:10.3390/bioengineering12101098 | Mohamed Ghadie, Sameer Sardaar, and Yannis Trakadis | Bioengineering | 2025]
Combines DNA language-model embeddings, biomedical knowledge graphs, and graph convolutional networks to predict variant pathogenicity in a disease-specific context.
193. A Systematic Review of the Application of Graph Neural Networks to Extract Candidate Genes and Biological Associations [DOI:10.1002/ajmg.b.33031 | Ankita Saxena et al. | American Journal of Medical Genetics Part B: Neuropsychiatric Genetics | 2025]
Reviews graph neural-network approaches for discovering candidate genes and biological associations from genomic, molecular-network, and disease data.
194. Deep-GenMut: Automated genetic mutation classification in oncology: a deep learning comparative study [DOI:10.1016/j.heliyon.2024.e32279 | Emad A. Elsamahy et al. | Heliyon | 2024]
Compares deep-learning and biomedical-language-model approaches for automatically classifying genetic mutation information relevant to cancer genomics.
195. Applications of support vector machine in genomic prediction in pig and maize populations [DOI:10.3389/fgene.2020.598318 | Wei Zhao et al. | Frontiers in Genetics | 2020]
Evaluates support-vector-machine approaches for genomic prediction of complex quantitative traits in livestock and crop breeding populations.
196. Using local convolutional neural networks for genomic prediction [DOI:10.3389/fgene.2020.561497 | Torsten Pook et al. | Frontiers in Genetics | 2020]
Develops locally connected convolutional architectures intended to capture regional marker effects and interactions for genomic prediction.
197. CHASMplus reveals the scope of somatic missense mutations driving human cancers [DOI:10.1016/j.cels.2019.05.005 | Collin Tokheim and Rachel Karchin | Cell Systems | 2019]
Extends machine-learning driver prediction to cancer-type-specific models and systematically identifies rare driver mutations across thousands of tumors.
198. A deep convolutional neural network approach for predicting phenotypes from genotypes [DOI:10.1007/s00425-018-2976-9 | Wenlong Ma et al. | Planta | 2018]
Introduces DeepGS, a convolutional neural network designed to predict quantitative phenotypes directly from genome-wide marker genotypes for genomic selection.
199. Cancer-specific high-throughput annotation of somatic mutations: computational prediction of driver missense mutations [PMID:19654296 | Hannah Carter et al. | Cancer Research | 2009]
Introduces CHASM, a random-forest classifier trained to distinguish cancer-driving missense mutations from likely passenger mutations.
