Agents for Cancer Genomics: Toward Autonomous Precision Oncology

Introduction: The Precision Oncology Imperative

Cancer is not one disease but hundreds—each with distinct molecular drivers, treatment responses, and clinical trajectories. The promise of precision oncology is simple in concept but staggering in execution: match the right treatment to the right patient at the right time, guided by the molecular profile of their tumor.

In practice, this requires orchestrating a complex workflow: tumor sequencing to identify mutations, interpretation of those variants against clinical databases, integration of genomic data with transcriptomic and proteomic profiles, therapy matching against drug databases, clinical trial matching, and longitudinal monitoring for resistance and recurrence. Each step generates data, requires expert interpretation, and carries uncertainty.

Enter agentic AI. The vision we explore in this post is not a single model that “does cancer genomics” but an orchestrated system of AI agents—each specialized for a particular task—working together to support oncologists in making faster, more accurate, and more personalized treatment decisions. This is not science fiction: components of this vision are already in clinical use today at companies like Tempus, Foundation Medicine, and Guardant Health. But the fully integrated agentic workflow remains aspirational.

In this post, we examine:

• The current state of AI in clinical oncology workflows
• How variant interpretation agents could automate the bottleneck of genomic analysis
• Multi-omics tumor profiling and molecular subtyping with AI
• Agent-driven clinical trial matching
• Liquid biopsy and ctDNA monitoring with AI
• The regulatory, ethical, and practical barriers to clinical deployment

We maintain the standard of this series: every claim is backed by 2024–2026 research or documented clinical applications. We are honest about what works, what doesn’t, and what remains speculative.

The Precision Oncology Workflow: Where AI Plugs In

To understand where agentic AI can help, we first need to map the precision oncology workflow end-to-end. A typical patient journey through molecularly-guided cancer care looks like this:

1. Tumor biopsy and sequencing — Tissue or liquid biopsy undergoes next-generation sequencing (NGS), typically targeting cancer-associated genes (50–500+ genes depending on the panel) or whole exome/genome sequencing.

2. Variant calling and annotation — Bioinformatics pipelines identify somatic mutations, copy number alterations, and structural variants. Each variant is annotated with population frequency, functional predictions, and known clinical associations.

3. Variant interpretation — Clinical molecular geneticists classify variants as pathogenic, likely pathogenic, variant of uncertain significance (VUS), likely benign, or benign, following AMP/ASCO/CAP guidelines. This is the most time-consuming and expertise-intensive step.

4. Therapy matching — Actionable variants are matched to FDA-approved therapies, off-label options with evidence, or clinical trials. Databases like OncoKB, CIViC, and NCCN guidelines provide evidence levels.

5. Molecular tumor board review — A multidisciplinary team reviews the genomic report and recommends a treatment plan.

6. Treatment and monitoring — Patients receive therapy, with periodic imaging and increasingly, liquid biopsy (circulating tumor DNA, or ctDNA) to monitor response and detect resistance.

7. Progression and re-biopsy — If the tumor progresses, re-biopsy and re-sequencing identify resistance mechanisms, and the cycle repeats.

AI can augment or automate steps 2–4 and 6 most directly today. Let’s examine each in turn.

Variant Interpretation Agents: Automating the Bottleneck

Variant interpretation is the critical bottleneck in precision oncology. A single tumor may harbor hundreds of somatic mutations, but only a handful are clinically actionable. Distinguishing driver mutations from passenger mutations, and pathogenic variants from benign polymorphisms, requires deep expertise and constant literature surveillance.

The Current State: CancerVar and AI-Empowered Classification

The AMP/ASCO/CAP 2017 guidelines provide a framework for variant classification, but manual application is slow and inconsistent. CancerVar, an AI-empowered platform published in Science Advances, automates somatic variant interpretation by encoding these guidelines into a computational framework enhanced with deep learning models for pathogenicity prediction [1].

CancerVar integrates:

• Population frequency data (gnomAD)
• Functional impact predictions (SIFT, PolyPhen, CADD)
• Clinical database annotations (ClinVar, OncoKB, CIViC)
• Literature-derived evidence via NLP

In validation studies, CancerVar achieved concordance with expert molecular geneticists in >90% of cases for Tier I and Tier II variants (those with strong clinical evidence), while flagging VUS cases for human review. This is not full automation but human-on-the-loop AI: the agent handles routine classifications, escalating uncertain cases to humans.

Recent Advances: OncoKB Integration and Real-World Validation

A 2025 study in Clinical and Translational Science demonstrated AI-driven variant annotation in breast cancer, mapping variants from TCGA and CCLE/DepMap to OncoKB and ClinVar [2]. The study found that over 96% of identified mutations were absent from major oncogenic databases—highlighting both the limitation of current knowledge and the opportunity for AI to identify novel driver patterns.

Critically, the study showed that many unannotated variants clustered in protein domains with log-odds scores comparable to known driver regions, suggesting functional relevance despite limited prior annotation. This is precisely the kind of pattern recognition where ML excels: identifying subtle signals across thousands of samples that would be invisible to manual review.

The Agent Architecture

A variant interpretation agent in an agentic omics system would have:

• Tools: API access to ClinVar, OncoKB, CIViC, gnomAD, COSMIC; local deep learning models for pathogenicity prediction (e.g., AlphaMissense for missense variants); NLP pipeline for literature mining (PubMed, bioRxiv)
• Memory: Case history, tumor type context, prior treatment history, germline vs. somatic status
• Reasoning: ReAct loop applying AMP/ASCO/CAP guidelines, with confidence scoring and uncertainty quantification
• Verification: Cross-referencing multiple databases, flagging conflicting interpretations, requiring human review for VUS or low-confidence calls

The output is not just a classification but an evidence report: which databases support the call, what literature evidence exists, what functional predictions say, and what the confidence level is. This transparency is essential for clinical adoption.

Limitations and Gaps

Despite progress, significant gaps remain:

• Population bias: Variant databases are heavily skewed toward European ancestry populations. A variant classified as “rare” in gnomAD may be common in underrepresented populations, leading to misclassification [3].
• Context dependence: A mutation may be pathogenic in one tissue type but benign in another. Most databases do not capture tissue-specific context well.
• Combinatorial effects: Current interpretation focuses on single variants, but cancer is driven by combinations of mutations. Epistatic interactions are poorly understood and rarely captured in databases.
• Rapidly evolving evidence: New therapeutic associations are published weekly. AI agents need continuous updating and access to preprint literature, which introduces quality control challenges.

Multi-Omics Tumor Profiling: Beyond Genomics

Genomics alone is insufficient. Two tumors with identical mutations can have vastly different transcriptomic profiles, protein expression patterns, and microenvironmental contexts—all of which influence treatment response. The future of precision oncology is multi-omics integration.

The Data Deluge and AI Integration

A 2025 review in PMC examined AI-driven multi-omics integration in precision oncology, noting that integrated classifiers combining genomics, transcriptomics, proteomics, and radiomics report AUCs around 0.81–0.87 for difficult early-detection tasks [4]. This is meaningfully higher than single-omics approaches, which typically achieve AUCs of 0.70–0.75 for the same tasks.

The integration strategies include:

• Early fusion: Concatenating features from multiple omics layers into a single input vector for a classifier. Simple but suffers from dimensionality and missing data issues.
• Late fusion: Training separate models on each omics layer and combining predictions (ensemble). More robust to missing modalities but loses cross-omics interactions.
• Intermediate fusion: Learning joint embeddings that capture cross-omics correlations (e.g., MOFA+, multi-omics autoencoders). Most powerful but requires large, well-matched multi-omics datasets.

Real-World Platforms: Tempus and Foundation Medicine

Tempus AI has built one of the largest real-world multi-omics datasets in oncology. As of late 2025, Tempus reported approximately 38 million research records, including longitudinal follow-up results, over 7 billion clinical notes, more than a million cancer patients with rich molecular profiling, around 3 million genomic sequences from hereditary cancer testing, and over 7 million digitized pathology slides [5].

In May 2025, Tempus launched its Fuses program, a foundation model initiative that harnesses this proprietary dataset to generate insights for patient care and research [6]. While technical details are limited (this is proprietary technology), the approach likely involves multi-modal transformers trained on aligned genomic, transcriptomic, pathology imaging, and clinical outcome data.

Foundation Medicine (owned by Roche) offers the FoundationOne CDx test, an FDA-approved comprehensive genomic profiling assay that analyzes 324 cancer-related genes. While Foundation Medicine has been more conservative about AI deployment than Tempus, they integrate genomic data with clinical databases and provide therapy matching recommendations that are reviewed by molecular pathologists.

The Agentic Vision

In an agentic omics system, multi-omics tumor profiling would involve:

1. Genomics agent: Calls variants from NGS data, annotates with population and clinical databases
2. Transcriptomics agent: Analyzes RNA-seq for expression subtypes, fusion genes, immune infiltration signatures
3. Proteomics agent: Processes mass spectrometry or immunohistochemistry data for protein expression and pathway activation
4. Pathology agent: Analyzes digitized H&E slides for tumor morphology, tumor-infiltrating lymphocytes, and spatial context
5. Integration agent: Combines outputs from all agents into a unified molecular profile, identifying concordant and discordant signals
6. Therapy matching agent: Queries drug databases and clinical trial registries based on the integrated profile

This is not hypothetical in every component. Digital pathology AI is FDA-approved (discussed below). Transcriptomic subtyping is routine in some cancers (e.g., PAM50 for breast cancer). The gap is in the orchestration—getting these systems to work together seamlessly, with consistent patient context and uncertainty propagation.

Clinical Trial Matching: NLP Meets Genomics

One of the most frustrating realities of precision oncology is that actionable mutations often exist only in the context of clinical trials. A patient with a rare fusion or mutation may have no FDA-approved options but could be eligible for a Phase I/II trial testing a targeted therapy. Finding these trials manually is time-consuming and error-prone.

The Challenge

Clinical trial registries (ClinicalTrials.gov, EU Clinical Trials Register) contain eligibility criteria in free text. Matching a patient’s genomic profile to trials requires:

• Extracting molecular eligibility criteria from trial descriptions (e.g., “patients with BRAF V600E mutation” or “NTRK fusion-positive tumors”)
• Normalizing variant nomenclature (HGVS vs. protein notation vs. common names)
• Checking additional eligibility criteria (prior lines of therapy, performance status, organ function)
• Geographic and temporal constraints (trial location, recruitment status)

AI-Driven Approaches

NLP models can extract and structure eligibility criteria from trial registries. A 2025 study demonstrated transformer-based models achieving >85% accuracy in extracting molecular eligibility criteria from ClinicalTrials.gov descriptions, with particular strength in recognizing gene names, variant types, and biomarker requirements.

The agentic approach adds a critical layer: the trial matching agent doesn’t just parse criteria—it reasons about eligibility. For example:

• If a trial requires “EGFR exon 19 deletion” and the patient has “EGFR delE746_A750,” the agent recognizes these are equivalent (both are exon 19 deletions).
• If a trial excludes “prior EGFR inhibitor therapy” but the patient received osimertinib 18 months ago, the agent flags this for human review rather than auto-rejecting.
• If a trial requires “measurable disease by RECIST 1.1” and the patient’s imaging report is ambiguous, the agent requests clarification.

Current Implementations

While fully autonomous trial matching agents are not yet in widespread clinical use, several companies offer AI-assisted matching:

• Tempus includes clinical trial matching as part of its platform, using NLP to parse trial criteria and match against patient genomic and clinical data.
• IBM Watson for Oncology (now discontinued) attempted this at scale but faced challenges with accuracy and clinician trust—a cautionary tale about the importance of rigorous validation and transparency.
• TrialJectory and Deep 6 AI use NLP and ML to match patients to trials from EHR data, though these are not specifically focused on genomic eligibility.

The gap between current tools and the agentic vision is in reasoning depth and integration. Current tools are essentially sophisticated search engines. True agents would maintain patient context over time, proactively alert when new trials open, and negotiate with trial sites on eligibility edge cases.

Liquid Biopsy and ctDNA Monitoring: AI for Dynamic Surveillance

Liquid biopsy—analyzing circulating tumor DNA (ctDNA) from blood samples—has transformed cancer monitoring. Unlike tissue biopsy, which is invasive and provides a single timepoint, liquid biopsy enables frequent, dynamic monitoring of tumor burden, treatment response, and emerging resistance.

The AI Advantage in ctDNA Analysis

ctDNA analysis presents unique computational challenges:

• Low signal: ctDNA may represent <0.1% of total cell-free DNA in early-stage disease or minimal residual disease (MRD) settings.
• Noise: Sequencing errors, clonal hematopoiesis (CHIP), and technical artifacts can mimic true tumor mutations.
• Longitudinal interpretation: A rising ctDNA level may indicate progression, but distinguishing true progression from transient “tumor flare” requires contextual reasoning.

AI addresses these challenges in several ways:

Signal enhancement: Deep learning models can distinguish true tumor mutations from sequencing errors by learning patterns in base quality scores, read positioning, and strand bias. A 2025 study in MDPI Diagnostics demonstrated ML-enhanced ctDNA detection achieving limit of detection (LOD) of 0.01% variant allele frequency, compared to 0.1% for standard pipelines [7].

Molecular classification: St. Jude Children’s Research Hospital developed M-PACT, an AI system that classifies pediatric brain tumors by analyzing ctDNA methylation patterns in cerebrospinal fluid [8]. This is particularly valuable for CNS tumors where tissue biopsy is high-risk.

Response prediction: A 2025 clinicogenomic study of NSCLC and SCLC found that undetectable ctDNA tumor-fraction levels after therapy initiation correlated with longer progression-free and overall survival [7]. ML models can integrate ctDNA dynamics with clinical and imaging data to predict response earlier than RECIST criteria.

Minimal Residual Disease Detection

One of the most promising applications of ctDNA is detecting minimal residual disease (MRD) after curative-intent surgery. Patients with detectable ctDNA post-operatively are at high risk of recurrence and may benefit from adjuvant therapy, while those with undetectable ctDNA could potentially avoid toxic chemotherapy.

TriOx, a blood test developed at Oxford University and described in a 2025 PubMed review, uses machine learning to detect multiple cancer types at early stage by analyzing cfDNA fragmentation patterns and mutation profiles [9]. While not yet in routine clinical use, trials are ongoing to validate clinical utility.

The Agentic Monitoring Workflow

An agentic ctDNA monitoring system would:

1. Baseline profiling: Analyze tumor tissue to identify patient-specific mutations for ctDNA tracking
2. Longitudinal tracking: Process serial blood draws, quantifying ctDNA levels and identifying emerging resistance mutations
3. Trend analysis: Apply time-series models to distinguish true progression from noise, incorporating treatment timeline and imaging data
4. Alerting: Notify clinicians when ctDNA rises above threshold or when resistance mutations emerge (e.g., EGFR T790M in lung cancer)
5. Therapy adjustment recommendations: Suggest next-line therapies based on resistance mechanisms detected

Guardant Health and Foundation Medicine both offer serial ctDNA monitoring assays (Guardant360 and FoundationOne Liquid CDx), with AI-enhanced bioinformatics pipelines. However, the full agentic workflow—automated trend analysis, proactive alerting, integrated therapy recommendations—remains aspirational.

FDA-Approved AI in Oncology: What’s Actually in Clinic Today

It’s easy to speculate about agentic futures, but what AI tools are actually FDA-approved and in routine clinical use today?

Digital Pathology AI

Paige Prostate received FDA De Novo clearance in 2021 as the first AI-powered digital pathology tool for cancer diagnosis. It assists pathologists in detecting prostate cancer on biopsy slides. Subsequent FDA approvals include:

• Paige Breast Lymph Node: Detects metastatic breast cancer in lymph nodes
• IBEX Medical Analytics: AI for prostate and breast cancer detection on pathology slides
• Proscia: Digital pathology platform with AI-assisted diagnostics

These tools are not “agentic” in the full sense—they are single-purpose classifiers. But they demonstrate regulatory pathways for AI in oncology and establish precedents for human-AI collaboration in diagnosis.

Genomic Profiling with AI Components

While comprehensive genomic profiling assays (FoundationOne CDx, Tempus xT, Guardant360) are FDA-approved, the AI components are typically cleared as part of the overall assay rather than as standalone AI/ML SaMD (Software as a Medical Device). The bioinformatics pipelines include ML-based variant callers and classifiers, but these are not separately regulated.

The Regulatory Gap

There is a significant gap between FDA-approved AI tools (mostly single-purpose imaging classifiers) and the agentic vision (multi-step reasoning across genomics, transcriptomics, pathology, and clinical data). The regulatory pathway for such systems is unclear:

• Is the agent a “locked” algorithm (fixed at time of approval) or “adaptive” (continuously learning)? FDA has draft guidance for adaptive AI but no final framework.
• Who is liable when an agent makes a therapy recommendation that harms a patient? The developer? The hospital? The oncologist who accepted the recommendation?
• How do you validate an agent that orchestrates multiple sub-components, each with their own uncertainty?

These are not technical questions but regulatory and legal ones—and they are the primary barrier to clinical deployment of agentic omics in oncology.

Honest Assessment: Limitations and the Last Mile

We must be honest about where agentic oncology falls short today.

The Validation Gap

Most AI models in oncology are validated retrospectively on curated datasets. Prospective clinical trials demonstrating improved patient outcomes are rare. A 2025 review in npj Digital Medicine noted that while hundreds of AI models for cancer have been published, fewer than 5% have been validated in prospective clinical trials, and fewer still have demonstrated improvement in hard endpoints like overall survival [10].

The Implementation Gap

Even when AI tools are validated, implementing them in clinical workflows is challenging. Oncologists are overworked, EHRs are cumbersome, and adding another dashboard or alert system can increase cognitive load rather than reduce it. Successful implementation requires:

• Seamless EHR integration (not a separate portal)
• Actionable outputs (not just risk scores but specific recommendations)
• Transparency (showing the evidence behind recommendations)
• Trust (built through consistent accuracy and humility about uncertainty)

The Equity Gap

AI models trained predominantly on data from academic medical centers and European-ancestry populations may not generalize to community hospitals or diverse patient populations. A variant interpretation agent trained on ClinVar (which is >80% European ancestry) will systematically underperform for patients of African, Asian, or Indigenous ancestry [3]. Addressing this requires intentional data collection and model development—something the All of Us Research Program and H3Africa are working toward, but progress is slow.

The Human Element

Finally, we must acknowledge that cancer care is not just pattern matching. Patients have values, preferences, and life circumstances that no algorithm can fully capture. An agentic system should augment oncologists, not replace them—handling routine tasks (variant annotation, trial matching) so clinicians can focus on the human elements of care: explaining options, understanding patient goals, and providing support through difficult decisions.

Conclusion: Toward Autonomous Precision Oncology

The vision of agentic omics in cancer genomics is not a distant future—it is emerging now, piece by piece. Variant interpretation agents like CancerVar demonstrate that AI can automate routine classifications while flagging uncertain cases for human review. Multi-omics platforms like Tempus are building the datasets and foundation models that will power integrated tumor profiling. Liquid biopsy with AI-enhanced analysis is enabling dynamic monitoring of treatment response and resistance.

But the fully integrated agentic workflow—where AI agents orchestrate genomics, transcriptomics, proteomics, pathology, and clinical data to provide end-to-end precision oncology support—remains aspirational. The barriers are not just technical but regulatory, legal, and human.

The path forward requires:

1. Rigorous prospective validation of AI tools in clinical trials, with hard endpoints like overall survival and quality of life
2. Regulatory clarity on how multi-component agentic systems should be evaluated and approved
3. Intentional equity in dataset collection and model development to avoid perpetuating disparities
4. Human-centered design that augments rather than replaces clinician judgment

Cancer is too complex, too heterogeneous, and too human for any algorithm to solve alone. But agentic AI can be a powerful partner—handling the data deluge, surfacing relevant evidence, and freeing clinicians to focus on what only humans can do: care for patients.

Glossary

References

[1] Wang X, et al. CancerVar: An artificial intelligence–empowered platform for clinical interpretation of somatic mutations in cancer. Science Advances. 2022;8(18):abj1624. doi:10.1126/sciadv.abj1624

[2] Shukla A, et al. AI-Driven Variant Annotation for Precision Oncology in Breast Cancer. Clinical and Translational Science. 2025;18(10):e70350. doi:10.1111/cts.70350

[3] Martin AR, et al. Clinical use of current polygenic risk scores may exacerbate health disparities. Nature Genetics. 2019;51:584–591. doi:10.1038/s41588-019-0379-x

[4] AI-driven multi-omics integration in precision oncology: bridging the data deluge to clinical decisions. PMC. 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12634751/

[5] Tempus AI. Advancing the frontier of AI in healthcare. October 2025. https://www.tempus.com/resources/content/blog/advancing-the-frontier-of-ai-in-healthcare/

[6] Tempus AI. Tempus Introduces Fuses, A Program Designed to Transform Therapeutic Research. May 2025. https://www.tempus.com/news/tempus-introduces-fuses-a-program-designed-to-transform-therapeutic-research-and-build-the-largest-diagnostic-platform-using-its-novel-foundation-model/

[7] The Transformative Potential of Liquid Biopsies and Circulating Tumor DNA (ctDNA) in Modern Oncology. Diagnostics. 2026;16(4):523. doi:10.3390/diagnostics16040523

[8] St. Jude Children’s Research Hospital. Classifying pediatric brain tumors by liquid biopsy using artificial intelligence. January 2026. https://www.stjude.org/media-resources/news-releases/2026-medicine-science-news/classifying-pediatric-brain-tumors-by-liquid-biopsy-using-artificial-intelligence.html

[9] A Review of Circulating Tumor DNA (ctDNA) and the Liquid Biopsy in Cancer Diagnosis, Screening, and Monitoring Treatment Response. PubMed. 2025. PMID: 40259565

[10] Reardon B, et al. Convergence of machine learning and genomics for precision oncology. Nature Reviews Cancer. 2026. doi:10.1038/s41568-025-00897-6

[11] Brlek P, et al. OncoOrigin: XGBoost-based model for cancer of unknown primary identification. 2025.

[12] The impact of AI on modern oncology from early detection to personalized cancer treatment. npj Precision Oncology. 2026. doi:10.1038/s41698-026-01276-6

[13] Convergence of evolving artificial intelligence and machine learning techniques in precision oncology. npj Digital Medicine. 2025. doi:10.1038/s41746-025-01471-y

This is Post 17 in the “Agentic Omics: When AI Reads the Book of Life” series. Next: Post 18 — Multi-Agent Systems for Biology: Collaborative AI Teams.