Artificial intelligence is no longer a peripheral tool in biological research; it sits at the center of advances spanning genome interpretation, protein engineering, drug discovery, and microbial ecology. Over the past few years, deep learning models trained on biological sequences and structures have produced results that would have seemed implausible a decade ago, from designing mRNA molecules with dramatically higher protein output to classifying viral lineages that had never been catalogued. The breadth of this shift is what makes it striking: AI is not solving one problem in biology but reshaping how researchers approach dozens of them simultaneously.
Reading the Genome With Language Models
One of the most intuitive applications of AI in biology borrows directly from how language models learn to process text. Instead of words and sentences, genomic language models train on DNA sequences, learning patterns that reveal which stretches of the genome matter and what happens when they change. A model called GPN demonstrated this approach in Arabidopsis, a plant species widely used in genetics research. Without any supervision or labeled data, GPN learned to score genetic variants across the entire genome, and the variants it flagged as most likely to be functional showed a 5.5-fold enrichment in rare variants, a strong signal that natural selection has been acting on those sites. GPN outperformed traditional conservation-based tools that rely on comparing sequences across related species.1PubMed Central. DNA language models are powerful predictors of genome-wide variant effects
Google DeepMind’s AlphaGenome pushed this concept further into non-coding DNA, the vast majority of the genome that does not encode proteins but plays critical regulatory roles. Described as the largest multimodal DNA sequence model for non-coding regions to date, AlphaGenome advanced the state of the art across nearly all prediction tasks and provides a unified tool for studying how genetic variation affects cellular regulation.2PubMed Central. The AlphaGenome deep learning model predicts effects of non-coding variants Understanding non-coding variants matters because they account for the majority of genetic associations with common diseases, yet they have been much harder to interpret than mutations that alter protein sequences directly.
A persistent bottleneck for these models is the sheer length of DNA they need to consider. The human genome is about three billion base pairs long, and important regulatory elements can influence genes from enormous distances. Standard transformer architectures, the same architecture behind ChatGPT, scale quadratically with input length, which means doubling the sequence you feed in roughly quadruples the computation required.3Nature Methods. Nucleotide Transformer: building and evaluating robust foundation models for human genomics Previous transformer-based genomic models were limited to context windows of 512 to 4,000 tokens, covering less than 0.001% of the human genome and severely limiting their ability to capture long-range interactions.4NeurIPS Proceedings. Bioinformatics AI: Driving Future Biological Breakthroughs New architectural approaches are actively being developed to work around this, but it remains one of the field’s hardest engineering challenges.
Designing Proteins and Molecules That Never Existed
If reading the genome is one half of the story, writing new biology is the other. Generative AI models can now design proteins and peptides from scratch, producing molecules with specific desired functions. One pipeline combined diffusion-based generative modeling with molecular simulation to design novel peptide inhibitors targeting a quorum-sensing receptor in Pseudomonas aeruginosa, a bacterium responsible for dangerous hospital-acquired infections. The AI-designed peptides showed strong structural stability and favorable binding properties, demonstrating how generative models can create functional molecules beyond what conventional trial-and-error approaches would find.5PubMed Central. Bridging Generative AI and Diffusion Models With Molecular Simulation to Design Anti-Quorum-Sensing De novo Peptides Targeting LasR of Pseudomonas aeruginosa
The broader drug discovery landscape has embraced several families of generative models. Variational autoencoders, generative adversarial networks, autoregressive transformers, and diffusion models are all being applied to navigate chemical space, designing small molecules optimized for properties like how well they are absorbed in the body, how toxic they might be, how easy they are to synthesize, and how tightly they bind their intended target.6Medicine in Drug Discovery. Generative AI for drug discovery and protein design: the next frontier in AI-driven molecular science Rather than screening millions of existing compounds to find a handful worth testing, researchers can now generate candidates tailored to specific criteria from the outset.
Proteins are not rigid objects; they flex and shift between different shapes, and those movements are often essential to their function. Machine learning models have been trained on simulation data to generate entire conformational ensembles of proteins, producing physically realistic snapshots of protein flexibility at negligible computational cost compared to traditional molecular dynamics simulations.7Nature Communications. Direct generation of protein conformational ensembles via machine learning This is significant because understanding how a protein moves can be just as important as knowing its average shape, especially for designing drugs that need to lock into a specific conformation.
Supercharging mRNA Therapeutics
The COVID-19 pandemic made mRNA vaccines a household concept, but designing an effective mRNA molecule is a delicate optimization problem. The sequence has to be translated efficiently by cells, remain stable long enough to produce its payload, and avoid triggering unwanted immune responses. Computational tools have increasingly been applied to optimize mRNA sequences according to these principles.8PubMed Central. mRNA vaccine sequence and structure design and optimization: Advances and challenges
GEMORNA, a generative model built on transformer architectures, was designed specifically for mRNA coding sequences and untranslated regions. In laboratory tests, GEMORNA-designed mRNAs produced up to 41 times more protein than an optimized benchmark sequence. When applied to therapeutic mRNAs, the model achieved up to a 15-fold boost in human erythropoietin expression and substantially improved antibody responses from a COVID vaccine construct in mice.9PubMed. Deep generative models design mRNA sequences with enhanced translational capacity and stability A 41-fold improvement in protein expression is not a marginal gain; it could translate into lower doses, simpler storage requirements, or stronger immune responses in future vaccines and protein-replacement therapies.
Precision Gene Editing
CRISPR gene editing depends heavily on the guide RNA that directs the molecular scissors to the right place in the genome. A poorly chosen guide can miss its target or, worse, cut in unintended locations. Machine learning and deep learning methods are now routinely used to predict both on-target editing efficiency and off-target risk for candidate guide RNAs.10PubMed Central. Using traditional machine learning and deep learning methods for on- and off-target prediction in CRISPR/Cas9: a review
DeepCRISPR was one of the first platforms to unify on-target and off-target prediction into a single deep learning framework, outperforming earlier computational tools.11PubMed Central. DeepCRISPR: optimized CRISPR guide RNA design by deep learning More recent tools like CRISPRon and CRISPRoff provide user-friendly web interfaces where researchers can input a target gene and receive ranked guide RNA candidates with predicted editing efficiency and off-target profiles, making the design process accessible to labs without deep computational expertise.12PubMed Central. CRISPRon/off: CRISPR/Cas9 on- and off-target gRNA design The practical upshot is fewer failed experiments and safer edits, which matters increasingly as CRISPR-based therapies move into clinical trials.
Understanding Cells One at a Time
Modern biology can now measure gene activity in individual cells rather than averaging across millions of them, and AI has become essential for making sense of the resulting data. Foundation models trained on single-cell gene expression data have achieved strong performance across a range of tasks, including predicting how individual cells respond to drugs, annotating cell types, enhancing noisy measurements, and inferring gene regulatory networks.13Nature Methods. Large-scale foundation model on single-cell transcriptomics The concept of a foundation model here is borrowed from the same approach behind large language models: train a single large model on vast amounts of data, then fine-tune it for many specific downstream tasks.
Single-cell data becomes even more powerful when paired with spatial information showing where each cell sits within a tissue. DeepTalk, a method based on graph attention networks, integrates single-cell RNA sequencing data with spatial transcriptomics data to infer how cells communicate with their neighbors at single-cell resolution.14Nature Communications. Deciphering cell–cell communication at single-cell resolution for spatial transcriptomics with subgraph-based graph attention network Mapping which cells are talking to which, and what signals they are sending, is central to understanding how tissues develop, how tumors evade the immune system, and why some treatments work in some patients but not others.
The integration challenge extends beyond pairing two data types. Researchers now generate transcriptomic, epigenomic, proteomic, and spatial imaging data from the same tissues, and multimodal integration approaches use techniques like tensor-based fusion to harmonize these layers into unified regulatory maps.15PubMed Central. Transformative advances in single-cell omics: a comprehensive review of foundation models, multimodal integration and computational ecosystems Without AI, combining these datasets would be impractical; the dimensionality is simply too high for manual analysis.
Mining Microbial Dark Matter
The vast majority of microbial life on Earth has never been cultured in a laboratory. Metagenomics, which sequences DNA directly from environmental samples, captures fragments of these organisms, but identifying what is there and what it does is a massive computational challenge. Machine learning has made it possible to discover novel functional genes more efficiently from this data. DeepARG, for instance, identified thousands of previously unannotated antibiotic resistance genes from metagenomic datasets, expanding our understanding of how resistance spreads in the environment.16PubMed Central. Microbial Dark Matter: from Discovery to Applications
Viruses present an even bigger mystery. DeepVirus, a hierarchical deep learning system, combines protein-level information with genome-aware representations to classify viruses across deep evolutionary hierarchies and, crucially, to detect candidate novel viral lineages that do not fit into known categories. Applied to large-scale metagenomic resources, it uncovered extensive viral diversity, including previously uncharacterized RNA-dependent RNA polymerases, expanding the known evolutionary space of RNA viruses.17bioRxiv. Illuminating the Virosphere’s Dark Matter using Hierarchical Deep Learning Given that emerging infectious diseases often originate from viral lineages we have not yet catalogued, this kind of computational prospecting could prove important for pandemic preparedness.
Metabolite Identification and Disease Signatures
Metabolomics, the study of small molecules produced by cellular processes, has long suffered from an identification bottleneck. Mass spectrometry can detect thousands of molecular signals in a biological sample, but matching those signals to specific chemical structures remains difficult. ChemEmbed, a deep learning framework for metabolite identification, was applied to a dataset of recurrently unidentified spectra and successfully confirmed 25 previously unknown compounds, demonstrating that AI can crack open previously intractable identification problems.18Briefings in Bioinformatics. ChemEmbed: a deep learning framework for metabolite identification using enhanced MS/MS data and multidimensional molecular embeddings
Beyond identification, end-to-end deep learning methods can analyze raw mass spectrometry data to find disease-specific metabolic profiles. In lung adenocarcinoma samples, one such approach matched 82 proteins and 121 metabolites, including 111 “hidden” metabolites discovered through correlation patterns rather than direct spectral matching, linking metabolic signals to disease states through protein-metabolite interaction networks.19Nature Communications. An end-to-end deep learning method for mass spectrometry data analysis to reveal disease-specific metabolic profiles This kind of analysis moves metabolomics from cataloguing what is present toward understanding what metabolic changes actually mean for a patient’s condition.
Precision Medicine and Imaging
AI’s role in personalized treatment extends across multiple data types simultaneously. In cancer immunotherapy, AI processes genomic and multi-omic data to identify biomarkers linked to treatment responses and disease prognosis. In radiomics, it extracts high-dimensional features from CT, MRI, and PET/CT images to discover imaging biomarkers tied to tumor heterogeneity and treatment response, enabling non-invasive monitoring. In pathomics, AI analyzes digital pathology images to uncover subtle changes in tissue microenvironments and cellular features that predict how patients will respond to immunotherapy.20Cancer Biology & Medicine. Advancing precision medicine: the transformative role of artificial intelligence in immunogenomics, radiomics, and pathomics for biomarker discovery and immunotherapy optimization The convergence of genomic, imaging, and pathology data through AI creates a more complete picture of each patient’s tumor than any single data type could provide alone.
Rebuilding Evolutionary Trees
Reconstructing evolutionary relationships among species is one of the oldest problems in biology, traditionally tackled with statistical methods that can be slow and computationally expensive. Phyloformer, a deep neural network approach, competes directly with these classical methods. Under standard models of protein evolution and with GPU acceleration, Phyloformer outpaced all other approaches and exceeded their accuracy in a metric that accounts for both tree shape and branch lengths. When a more complex model of evolution that includes dependencies between sites was used, Phyloformer outperformed all other methods across all metrics on alignments with fewer than 80 sequences. On nearly 4,000 empirical gene alignments from five datasets, it matched the topological accuracy of maximum likelihood methods.21Oxford Academic. Phyloformer: Fast, Accurate, and Versatile Phylogenetic Reconstruction with Deep Neural Networks Speed matters here because researchers routinely need to build thousands of gene trees in comparative genomics studies, and a tool that runs quickly on GPUs without sacrificing accuracy changes what analyses are feasible.
When the Models Cheat
The enthusiasm around AI in biology comes with a serious credibility problem that the field is still wrestling with: data leakage. When training and test sets share similar sequences, a model can appear to generalize well while actually just memorizing patterns it has already seen. Research on protein interaction benchmarks found that commonly used splitting strategies based on sequence or metadata similarity introduce major data leakage, potentially producing overoptimistic evaluations that measure a model’s ability to overfit rather than its practical utility.22arXiv. Revealing data leakage in protein interaction benchmarks
The same issue affects genomic models. In models that predict human gene expression from DNA sequence, performance on test sequences varies with their similarity to training sequences, consistent with homology-based data leakage. Because a sequence and its function are linked, even a maximally overfit model with no real understanding of gene regulation can predict the expression of sequences that resemble its training data.23bioRxiv. Detecting and avoiding homology-based data leakage in genome-trained sequence models This does not mean every impressive benchmark result is fake, but it does mean the community needs more rigorous evaluation practices. A model’s real-world value depends on how well it performs on truly novel sequences, not on how well it recognizes slightly scrambled versions of what it trained on.
Making Black Boxes Transparent
Even when a model works well, researchers often cannot explain why it makes specific predictions, which limits trust and biological insight. ExplaiNN addresses this by combining the predictive power of convolutional neural networks with the interpretability of linear models. It can predict transcription factor binding and chromatin accessibility while providing transparent explanations at both the global level (what drives predictions across an entire cell state) and the local level (what sequence features matter for a specific prediction).24PubMed Central. ExplaiNN: interpretable and transparent neural networks for genomics Interpretability is not just an academic nicety. When a model identifies a genomic variant as functional, a biologist needs to understand the reasoning before designing an expensive experiment around it. Transparent models bridge the gap between computational prediction and wet-lab follow-up.
Biosecurity Risks of Generative Biology
The same generative capabilities that allow researchers to design beneficial proteins and molecules carry inherent dual-use risks. AI-driven tools could be misused to engineer pathogens, toxins, or destabilizing biomolecules, and automated AI science agents could amplify these risks by handling experimental design steps that currently require human expertise.25PubMed Central. A call for built-in biosecurity safeguards for generative AI tools Current safety screening by DNA synthesis providers relies heavily on matching ordered sequences against databases of known dangerous organisms. AI-generated proteins could be functionally equivalent to known toxins while sharing little sequence similarity, rendering those homology-based screening methods essentially blind to such designs. The wide availability of open-source generative tools further lowers the barrier to misuse.26PubMed Central. Protein design, generative AI and biological security
This is not a hypothetical concern for the distant future. The tools already exist and are improving rapidly. The biosecurity community is calling for built-in safeguards within generative AI tools themselves, rather than relying solely on downstream screening that was designed for an era when novel sequence generation was not so easy.
A Regulatory Landscape Still Taking Shape
As AI-designed molecules move toward clinical use, regulators face the challenge of evaluating products created through processes that are fundamentally different from traditional drug development. The U.S. Food and Drug Administration and the European Medicines Agency have both begun issuing guidance on AI applications in human therapeutics, but these frameworks differ substantially in scope, terminology, and how they apply in practice.27PubMed Central. Reimagining drug regulation in the age of AI: a framework for the AI-enabled Ecosystem for Therapeutics A drug candidate designed by a generative model and validated through computational simulations raises questions that existing regulatory pipelines were not built to answer: How do you validate the training data? What counts as sufficient evidence that a model’s prediction is reliable? How do you audit a design process that involved millions of generated candidates filtered by an algorithm?
These questions are likely to become more urgent as the first wave of AI-designed therapeutics, including mRNA constructs, novel peptides, and optimized antibodies, enters clinical testing. The gap between what AI can now design and what regulatory frameworks can evaluate is one of the more consequential bottlenecks in translating bioinformatics AI from research tool to medical reality.