AI agents in action: Uncovering our inner mutants

Behind the acronym VUS lies human tragedy. Every year, thousands of patients receive genetic results labelled “Variant of Unknown Significance”—a phrase that offers uncertainty instead of answers.

A single letter change in DNA can mean nothing … or everything. Without biological context, a VUS leaves clinicians and families trapped between hope and fear—and, at times, it becomes the difference between life and loss.

One such story, highlighted here in The Guardian, is that of a mother and her son, whose SCN1A mutation was initially classified as a VUS. When his symptoms worsened, diagnostic systems failed to recognize the variant’s true significance. Years later, his mother discovered that the same mutation had already been described in scientific literature and even documented in a patent as a disease-associated mutation. By the time this truth emerged, it was too late—the missed link between genomic data and existing scientific knowledge had cost her son his life.

Her experience exposes a critical challenge in modern genomics: while databases such as ClinVar and HGMD (Human Gene Mutation Database) curate known variants, millions of others remain unlinked, unannotated, or buried in research papers, case reports, and experimental data.

Early diagnosis saves lives—but only when every piece of evidence can be connected and understood.

Decoding genomic sequencing into a life-saving network of insight

Recent advances in high-throughput genome sequencing are reshaping how quickly we can read human DNA. In 2025, Roche Diagnostics announced a next-generation sequencing platform capable of producing a complete human genome—from raw sample to variant call file (VCF)—in as little as four hours. What once took days or weeks can now be done within a single clinical shift.

This acceleration opens extraordinary opportunities—but it also exposes a deeper challenge. As sequencing scales, millions of new variants are being detected across human and microbial genomes every year. Most of them fall into the category of VUS, meaning that their biological or clinical impact remains uncertain.

Fast sequencing tells us what is different, but not what it means. That’s where MutAgent comes in.

Inside the MutAgent framework: From sequencing to reasoning

The MutAgent framework was designed to make the full journey from genome to insight seamless and explainable. It orchestrates four specialized agents—VariantAgent, SemanticAgent,KnowledgeGraphAgent, and MutAgent—that operate in sequence to transform raw genomic data into traceable, evidence-backed knowledge.

Together, these agents operate from the point where modern sequencing technologies end—with the variant call format (VCF)—to:

• Standardize genomic variants using Human Genome Variation Society (HGVS) nomenclature and established bioinformatics and sequence-analysis tools – VariantAgent

• Semantically index variants and evidence from public literature and regulatory datasets into SciBite Search using ontology- and LLM-driven normalized entities and relationships – SemanticAgent

• Integrate semantic content into a custom EmBiology dataset, enriching licensed full-text–derived biological contexts and curated entities and relationships, with LLM-generated entities and relationships along with vectorized contextual evidence – KnowledgeGraphAgent

• Apply reasoning and retrieval through vector search, knowledge-graph exploration, and LLM-based evidence validation and synthesis to deliver explainable insights grounded in the underlying data – MutAgent

VariantAgent—Turning raw genomes into searchable knowledge

Every journey in MutAgent starts with the VCF—whether generated from fast clinical sequencing runs or large population resources like the 1000 Genomes Project.

VariantAgent operates at this point, using established bioinformatics and sequence-analysis tools such as bcftools and AlphaGenome to parse, normalize, and annotate genomic variants with reference-aware sequence context. It transforms any VCF into a HGVS-normalized, variant profile enriched with curated identifiers from dbSNP, ClinVar, Ensembl, and NCBI, producing structured, machine-readable representations that are consistent across reference builds, datasets, species, and both human and microbial genomes.

The result? A unified variant profile—comparable across patients, microbes, populations, and studies.

With this foundation in place, each variant is now ready to be semantically linked to literature, pathways, and biological context in the next stage of the MutAgent framework.

SemanticAgent—Linking variants to knowledge in one semantic space

With variants standardized by VariantAgent, SemanticAgent brings them into SciBite Search, where both variants and literature—and any custom structured datasets—are indexed in one unified semantic environment.

SciBite Search applies FAIR-compliant, ontology-driven annotations to variants and all indexed resources, while also ingesting LLM-generated entities and relationships that are validated by LLM-based judge agents and normalized to ontologies for semantic retrieval. Because everything is enriched with the same unified biological entities, SSQL queries can seamlessly connect VCF-derived mutations to relevant publications, datasets, and biological knowledge.

In this shared semantic space, every variant becomes a linked and queryable unit of evidence—ready for deeper integration into MutAgent’s final hub—the EmBiology knowledge graph.

KnowledgeGraphAgent—The EmBiology hub turning knowledge to insight

All annotated entities flowing from SciBite Search are integrated into EmBiology, Elsevier’s curated biological knowledge graph. Here, KnowledgeGraphAgent extends EmBiology’s nodes and relationships from full-text journal articles with ontology-driven custom datasets—including patient genomic variants, public literature and regulatory resources, and proprietary pharma data—and enriches them with vector representations of contextualized evidence to support semantic retrieval and downstream reasoning.

This unified graph links each variant to its wider biological network: pathways, regulators, phenotypes, mechanisms, therapeutic targets, and supporting evidence.

A single SCN1A mutation, for example, becomes connected to neuronal signaling pathways, epileptic phenotypes, and candidate therapeutic strategies surfaced across curated and custom datasets.

By transforming previously siloed data into an interconnected biological web, KnowledgeGraphAgent provides the foundation for MutAgent’s “life-saving network of insight.”

MutAgent—The reasoning agent making sense of the hub

At the top of the framework sits MutAgent, the reasoning agent that makes sense of every connection flowing through the EmBiology knowledge-graph hub. MutAgent iterates across all variants, datasets, relationships, and evidence—transforming complex biological networks into precise, life-saving insights.

Through a conversational AI interface, clinicians, scientists, and regulators can literally talk to their data.

They can ask natural-language questions such as:

• “Diseases linked to SCN1?”

• “Patients with SCN1 mutations?”

• “SCN1 mutations to target with gene therapy and which gene therapy approach to use?”

Behind the scenes, MutAgent’s agentic AI models reason across the indexed variants, literature, and knowledge-graph relationships—using vector embeddings, ontology mappings, and LLM-based validation to retrieve, interpret, and explain results, always with traceable provenance to the underlying evidence.

Back to the future: A unified environment for precision insight

If MutAgent had been available when the mother from The Guardian story mentioned above had first received her son’s genetic report, the journey from genome to understanding could have followed a very different path. Instead of stopping at a Variant of Unknown Significance, sequencing data, existing literature, and biological knowledge could have been evaluated together within a single, connected environment—early enough to inform action.

By unifying sequencing data, scientific literature, and biological knowledge in one system, MutAgent transforms disconnected information into precision insight. It enables traceable reasoning from patient variants to publication evidence, cross-domain linking between human and microbial mutations, and interactive, explainable answers that support earlier diagnosis and more informed clinical decisions.

Crucially, this unified view allows variant mechanisms to be connected to therapeutic possibilities—from pathway reactivation strategies such as exa-cel, to gene correction approaches like pz-cel, and Adeno-Associated Virus (AAV)–based gene supplementation as used in Zolgensma. Rather than ending with variant detection, MutAgent helps surface what is already known—and what may be actionable—at the moment it matters most.

By uniting SciBite Search, EmBiology, and agentic AI into a single reasoning framework, MutAgent turns fragmented genomic and textual evidence into a living system of connected intelligence. In doing so, it moves precision medicine closer to a future where critical knowledge is not missed; evidence is not siloed, and disease-associated mutations are no longer left disconnected from existing evidence.

Beyond humans—Bringing MutAgent to antimicrobial resistance

MutAgent applies the same reasoning framework to antimicrobial resistance (AMR), where microbial genomics lack centralized mutation repositories.

Using annotated microbial variants from the EMERALD project—which adds environmental context, resistance states, and microbial interactions to metagenomic sequences—together with resources such as CARD, NCBI pathogen and AMR datasets, and EBI’s AMR mutation database, MutAgent builds a structured view of microbial resistance.

Through this unified pipeline, MutAgent uncovers bacterial genomic mutations that drive antimicrobial resistance, including self-resistance genes within biosynthetic gene clusters (BGCs), their regulators, and horizontal gene-transfer events. Once indexed into SciBite Search, these microbial variants can be semantically linked to published resistance mechanisms, compensatory pathways, and potential therapeutic strategies.

In doing so, MutAgent unites human and microbial mutation reasoning within a single explainable AI environment for precision medicine and global AMR discovery.

Learn more

Discover how Elsevier’s professional services team and SciBite support pharma and healthcare organizations.