Artificial intelligence has moved from a speculative tool to a working partner across nearly every branch of biology, from predicting how proteins fold to flagging which patients might benefit from an experimental cancer drug. The shift is not theoretical: AI models now predict protein structures with near-experimental accuracy, design entirely new molecules for drug candidates, read tumor slides with the skill of experienced pathologists, and forecast how viruses might mutate next. What makes the current moment different from earlier waves of computational biology is the sheer range of problems AI can tackle simultaneously and the speed at which it operates once trained.
Predicting and Designing Proteins From Scratch
Proteins carry out most of the work inside cells, and their three-dimensional shape dictates what they do. For decades, determining that shape required painstaking lab techniques like X-ray crystallography or cryo-electron microscopy. AlphaFold, a deep-learning system developed by DeepMind, upended that timeline. Over the past few years it has provided near-experimental-accuracy models of protein structures directly from amino acid sequences, transforming structural biology in the process.1PubMed Central. Advantages and Limitations of AlphaFold in Structural Biology: Insights from Recent Studies Researchers who once waited months for a single structure can now generate a reliable prediction in hours.
But predicting existing proteins is only half the story. A newer wave of AI tools goes further by designing proteins that have never existed in nature. Generative architectures, including language models and diffusion processes, can produce novel proteins that fold correctly and carry out specified functions. Current state-of-the-art design protocols achieve experimental success rates approaching 20%, meaning roughly one in five AI-designed proteins actually works when synthesized in the lab.2PubMed. Generative artificial intelligence for de novo protein design That may not sound high, but it represents an enormous leap: random mutagenesis or rational design by hand might succeed far less often for truly novel folds. AI navigates the staggering complexity of protein sequence space with a precision and speed that was previously impossible, opening the door to custom-built enzymes, therapeutic antibodies, and biosensors tailored to specific tasks.3Nature Reviews Bioengineering. AI-driven protein design
Speeding Up Drug Discovery
Developing a new drug traditionally takes over a decade and costs well over a billion dollars, with most candidates failing in clinical trials. AI is reshaping the early stages of that pipeline. Machine-learning models now assist with identifying which biological targets to pursue, screening vast chemical libraries for promising molecules, optimizing those molecules for potency and safety, and predicting how a compound will be absorbed, distributed, metabolized, and excreted by the body.4PubMed Central. Artificial Intelligence in Small-Molecule Drug Discovery: A Critical Review of Methods, Applications, and Real-World Outcomes The goal is not to replace medicinal chemists but to hand them better starting points faster.
Virtual screening is one area where the gains are already tangible. AI-driven docking algorithms can evaluate millions of molecules against a protein target in a fraction of the time a traditional physics-based simulation would require. Tools like KarmaDock and DeepDock have been used for large-scale ligand screening, while machine-learning scoring functions improve predictions of how tightly a candidate drug binds its target.5Journal of Bio-X Research. Artificial Intelligence in Virtual Screening: Transforming Drug Research and Discovery—A Review Meanwhile, platforms like ADMET-AI provide fast, automated predictions of a compound’s safety and pharmacokinetic profile, letting researchers weed out problematic molecules before they ever enter a test tube.6Bioinformatics. ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries
A realistic assessment, though, is that AI augments rather than replaces the traditional drug-discovery workflow. Persistent challenges remain in translating computational predictions into real-world results: a molecule that looks perfect on screen may behave unpredictably in living tissue.7PubMed. Artificial intelligence in small molecule drug discovery from 2018 to 2023: Does it really work? The field is still learning which AI predictions hold up under experimental validation and which are artifacts of biased training data.
Reading the Genome With Foundation Models
Just as large language models learn patterns in text, genomic foundation models learn patterns in DNA sequences. The Nucleotide Transformer, for instance, was trained on over 3,000 human genomes and hundreds of genomes from other species, producing models with up to 2.5 billion parameters. These models generate context-specific representations of DNA sequences that allow accurate predictions even when labeled training data is scarce.8Nature Methods. Nucleotide Transformer: building and evaluating robust foundation models for human genomics Researchers have used embeddings from such models to predict gene expression from raw DNA, narrowing the gap between sequence-based prediction and approaches that use known genetic variants directly.9PubMed Central. Enhancing personalized gene expression prediction from DNA sequences using genomic foundation models
The results are promising but far from solved. Benchmarking studies have found that zero-shot predictions from DNA foundation models, without any fine-tuning, still achieve only modest accuracy when predicting individual-level gene expression.10Nature Communications. Benchmarking DNA foundation models for genomic and genetic tasks Fine-tuning helps considerably, but the field is still working out how much training data is needed, which architectures generalize best, and where the ceiling lies for predicting complex traits from sequence alone.
Mapping Cells in Space and Function
Modern biology increasingly studies individual cells rather than blended tissue samples, and AI has become essential for making sense of the resulting data floods. Single-cell RNA sequencing generates expression profiles for thousands of genes in each of millions of cells. Tools like ProjectSVR use machine learning to map new single-cell datasets onto well-curated reference atlases, letting researchers identify cell types and states without starting the annotation process from scratch every time.11Briefings in Bioinformatics. ProjectSVR: mapping single-cell RNA-seq data to reference atlases by supported vector regression
Spatial transcriptomics, which preserves information about where in a tissue each cell sits, adds another dimension. Deep-learning frameworks now integrate histology images with spatial gene-expression data to reveal how cells organize themselves and interact within tumors or developing organs.12PubMed Central. Deep Learning-Enabled Integration of Histology and Transcriptomics for Tissue Spatial Profile Analysis In breast cancer, for example, spatial heterogeneity analysis using AI has uncovered complex interactions between tumor cells and their surrounding microenvironment, pointing toward potential therapeutic targets.13PubMed. SpatioFreq: A Deep Learning Framework for Decoding Cellular and Tissue Landscapes Across Organisms Using Spatial Transcriptomics Another framework, ROICellTrack, combines cellular imaging with transcriptomic profiling in bladder cancer tissues, identifying distinct cancer-immune cell mixtures and the receptor-ligand interactions linked to immune infiltration.14Bioinformatics. ROICellTrack: a deep learning framework for integrating cellular imaging modalities in subcellular spatial transcriptomic profiling of tumor tissues
Making CRISPR Gene Editing Safer
CRISPR gene editing relies on guide RNAs to direct a molecular scissor to a precise spot in the genome. The catch is that guides sometimes cut in the wrong place, producing off-target edits that could disable important genes or trigger harmful mutations. Machine learning and deep learning are now routinely used to predict both how effectively a guide will cut its intended target and how likely it is to cause off-target damage.15PubMed Central. Using traditional machine learning and deep learning methods for on- and off-target prediction in CRISPR/Cas9: a review
A newer tool called crispAI goes a step further by quantifying the uncertainty in its off-target predictions, rather than simply giving a single risk score. It models the noise in experimental cleavage data and outputs calibrated confidence intervals, so researchers know not just where off-target cuts might happen but how confident the model is about each prediction. The tool also provides a genome-wide guide-efficiency score that lets scientists rank candidate guides more informatively than previous aggregate scoring methods.16Nucleic Acids Research. Learning to quantify uncertainty in off-target activity for CRISPR guide RNAs For clinical gene therapy, where even a single dangerous off-target event could harm a patient, this kind of probabilistic risk assessment is a meaningful step forward.
Engineering Microbes With Machine Learning
Synthetic biology often involves rewiring microorganisms to produce useful chemicals, from biofuels to pharmaceutical precursors. The challenge is that cellular metabolism is a web of interconnected pathways: tweaking one enzyme can have cascading effects on the whole system. Machine learning helps by predicting how pathway dynamics will respond to genetic changes, sometimes outperforming traditional kinetic models that require detailed knowledge of each enzymatic reaction.17PubMed Central. A machine learning approach to predict metabolic pathway dynamics from time-series multiomics data
One practical example is optimizing how much protein each gene in an engineered pathway produces. By training machine-learning models on data from combinatorial libraries of ribosome binding sites in E. coli, researchers accurately predicted which sequence combinations would maximize output. In one case, this approach boosted production of a monoterpenoid compound by over 60% while screening under 3% of the possible library, a massive saving in lab time and reagents.18ACS Synthetic Biology. Machine Learning of Designed Translational Control Allows Predictive Pathway Optimization in Escherichia coli Tools like the Automated Recommendation Tool (ART) take this concept further by combining machine learning with automated experimental design, enabling productive strain engineering cycles even when starting data is limited.19PubMed Central. Machine Learning and Deep Learning in Synthetic Biology: Key Architectures, Applications, and Challenges
AI at the Microscope and in the Clinic
Cancer diagnosis still depends heavily on pathologists examining tissue slides under a microscope, a skilled but subjective process. Deep learning models trained on whole-slide histopathology images have shown they can classify tumors with performance comparable to experienced pathologists. In non-small cell lung cancer, a convolutional neural network achieved an average area under the curve of 0.97 for distinguishing between cancer subtypes and normal tissue. The same model could also predict specific gene mutations directly from the tissue images, with six commonly mutated genes in lung adenocarcinoma predicted at performance levels ranging from about 0.73 to 0.86.20PubMed Central. Classification and mutation prediction from non–small cell lung cancer histopathology images using deep learning Similar results have been demonstrated in liver cancer, where a deep-learning model reached roughly 96% accuracy for distinguishing benign from malignant tumors and could predict four common mutations with external validation scores between 0.71 and 0.89.21npj precision oncology. Classification and mutation prediction based on histopathology H&E images in liver cancer using deep learning
Beyond pathology, AI is being used to integrate multiple types of biological data for cancer diagnosis and prognosis. One hybrid model, OmicsFusionNet, combines genomic, transcriptomic, and epigenomic data to classify tumors across 23 cancer types, reaching about 80% accuracy overall and up to roughly 99.8% when combining RNA sequencing with methylation data for specific tasks.22Journal of Genetic Engineering and Biotechnology. AI based multiomics integration for cancer diagnosis and prognosis At the clinical level, AI has shown strong results in molecular subtyping, survival prediction, and forecasting which patients are likely to respond to immunotherapy.23Advanced Intelligent Discovery. AI‐Driven Cancer Multi‐Omics: A Review From the Data Pipeline Perspective
Tracking Viral Evolution and Predicting Mutations
The COVID-19 pandemic made painfully clear how quickly viruses mutate and how much those mutations matter for public health. AI models are now being trained to anticipate viral evolution rather than simply react to it. One approach borrows directly from natural language processing: researchers treated viral protein sequences as sentences and mutations as word changes, looking for alterations that preserve the virus’s ability to infect cells while changing how the immune system recognizes it. This language-model approach accurately predicted structural escape patterns in influenza, HIV, and SARS-CoV-2 using sequence data alone.24PubMed. Learning the language of viral evolution and escape
Other groups have taken a more direct classification approach, training models like XGBoost on mutation and clade information from infectious viruses. When clade information was included alongside raw mutation data, prediction accuracy reached as high as 0.999 in some configurations, though the practical usefulness of such models depends on whether they generalize to genuinely novel variants that look nothing like the training data.25PubMed Central. A prediction of mutations in infectious viruses using artificial intelligence The ultimate goal is a surveillance system that can flag dangerous new mutations before they spread widely, giving vaccine designers a head start.
Seeing Molecules Move
Proteins are not static sculptures. They flex, twist, and shift between conformations, and those movements are often essential to their function. Simulating these motions traditionally requires enormous computational resources: molecular dynamics simulations can take weeks on supercomputers to capture even microseconds of molecular time. Machine learning offers a shortcut. Generative models trained on simulation data can produce physically realistic conformational ensembles at negligible computational cost, effectively learning the rules of protein motion from examples and then generating new plausible snapshots without running the full simulation.26Nature Communications. Direct generation of protein conformational ensembles via machine learning
This matters especially for intrinsically disordered proteins, which lack a fixed shape and are notoriously difficult to study with traditional structural methods. Deep-learning models have been developed to sample the conformational landscape of highly dynamic molecules like the amyloid-beta peptide involved in Alzheimer’s disease, generating diverse and physically plausible structures that help researchers understand how these proteins misfold and aggregate.27PubMed Central. Sampling Conformational Ensembles of Highly Dynamic Proteins via Generative Deep Learning On the imaging side, neural-field networks have improved cryo-electron microscopy reconstructions, pushing 3D density resolution beyond what conventional algorithms achieve and resolving structural elements that were previously invisible, such as flexible regions of the SARS-CoV-2 spike protein.28Nature Machine Intelligence. High-resolution real-space reconstruction of cryo-EM structures using a neural field network
Matching Patients to Clinical Trials
One of the quieter bottlenecks in medicine is matching patients to clinical trials they might benefit from. Eligibility criteria can run pages long, and checking a patient’s medical record against those criteria is tedious work that often does not get done thoroughly. AI systems powered by large language models are starting to automate this process. TrialMatchAI, for instance, uses a dedicated language-model module to extract and standardize biomarker, genomic, and mutation data from both patient records and trial criteria, then recommends trials where a patient is likely eligible. The system has shown particular promise in oncology, where matching requires integrating complex molecular and genetic data.29Nature Communications. TrialMatchAI: an end-to-end AI-powered clinical trial recommendation system to streamline patient-to-trial matching A similar system, PRISM, has been validated against real-world electronic health records for large-scale trial matching.30PubMed Central. PRISM: Patient Records Interpretation for Semantic clinical trial Matching system using large language models If these tools reach routine clinical use, they could significantly increase the number of patients who find their way into appropriate trials, which in turn would accelerate the pace of research itself.
Reconstructing Evolutionary History
AI is also reshaping how biologists study evolution. Ancestral sequence reconstruction, the practice of inferring what ancient proteins looked like before they diverged into today’s family of variants, has traditionally relied on probabilistic models that assume each position in a protein evolves independently. That assumption is often wrong: mutations at one site can change the effect of mutations at other sites, a phenomenon called epistasis. Generative AI models can now account for these interactions, producing more diverse and less biased reconstructions of ancestral sequences than previous state-of-the-art methods.31Molecular Biology and Evolution. Reconstruction of Ancestral Protein Sequences Using Autoregressive Generative Models
At a broader scale, machine learning has been applied to predict how entire metabolic systems evolve across bacterial lineages. A framework called Evodictor, trained on roughly 3,000 bacterial genomes, successfully predicted which genes would be gained or lost along branches of the evolutionary tree. The fact that the model works across diverse bacteria suggests that the evolutionary pressures shaping metabolic systems are more universal than previously appreciated.32PubMed Central. Machine learning enables prediction of metabolic system evolution in bacteria
The Black Box Problem and Biosecurity Risks
For all its promise, AI in biology faces serious unresolved challenges. The most frequently cited is interpretability. Many of the best-performing models, especially deep neural networks, function as black boxes: they produce accurate predictions but offer little insight into why. In healthcare, where a wrong prediction could harm a patient, this opacity is not just inconvenient but potentially dangerous. Explainability and accountability are increasingly seen as legal requirements for AI systems that affect human lives, and fairness is a growing concern, since models can inadvertently discriminate based on the demographics overrepresented or underrepresented in their training data.33Briefings in Bioinformatics. Explainable AI for Bioinformatics: Methods, Tools and Applications Data bias, limited interpretability, and the gap between computational predictions and wet-lab reality remain persistent obstacles across drug discovery and other applications.34Annual Reviews. Precision Drug Discovery in the Era of Artificial Intelligence: A Critical Review
Regulation is another open question. AI-driven technologies are anticipated to play an unprecedented role in transforming drug development, and regulators around the world are grappling with how to evaluate AI-generated evidence, validate computational predictions, and ensure safety standards keep pace with the technology.35PubMed Central. Regulating the AI-enabled ecosystem for human therapeutics
Then there is the dual-use problem. The same generative AI that designs beneficial proteins could, in theory, design harmful ones. Deep generative models can produce novel biological molecules that do not resemble anything in existing genome databases, which means they could potentially bypass the sequence-matching safety screens that DNA synthesis companies currently use to catch dangerous orders. AI science agents that automate experimental design could further amplify these risks. A growing number of researchers and policymakers are calling for built-in biosecurity safeguards to be integrated directly into generative AI tools for biology, rather than relying on after-the-fact screening.36PubMed Central. A call for built-in biosecurity safeguards for generative AI tools The tension between openness, which accelerates scientific progress, and restriction, which limits misuse, is one the field has not yet resolved and probably will not anytime soon.