Commentary|Articles|October 9, 2026

AI in Precision Oncology: A Paradigm Shift

Author(s)Shivansh Sahni, Casey Reulbach, DO, Madison Brown, DO

"AI will not replace oncologists; it will help interpret complex data and design more precise, adaptive, personalized cancer treatment," the authors write.

Artificial intelligence (AI) has become increasingly visible in clinical practice, including in evidence retrieval and clinical documentation. AI-based tools are being evaluated for their ability to answer clinical questions and assist with digital triage workflows.^1-3As these technologies become more common across medicine, oncology will begin to see similar applications in drug development, therapy selection, and planning for personalized treatments.

Traditional cancer treatment techniques are mainly based on tumor site and stage, pathology results, prior lines of therapy, and the physician’s clinical judgment. However, the shift to precision oncology involves therapy based more on molecular profiles and the context of an individual patient. The rapid growth in detectable actionable targets, the complex interplay between different markers, and the multitude of treatment options for each target have created a need for a more algorithmic approach to treatment selection.

Between 2017 and 2022, standard-care predictive biomarkers grew from 8.9% to 31.6% in MSK-IMPACT/OncoKB analysis.⁴ The main issue is that precision oncology is producing substantially too much data for complete manual interpretation. The emerging role of AI in this field includes organizing genomic, transcriptomic, single-cell, drug-response, resistance, and clinical trial data into more usable, simplified information for treatment. AI is moving oncology to a stage where treatment design is based on response probability and resistance risk, among other patient-specific factors.

Mutation-to-Drug Matching

Next-generation sequencing can identify mutations, amplifications, fusions, copy-number alterations, and tumor-agnostic biomarkers. Variant databases are used to connect alterations with therapies, evidence levels, and trial options. An example of a cancer variant interpretation data source is CIViCdb, an open-access database to help standardize variant-to-treatment evidence.⁵ Mutation matching is not automatic because the same alteration can act differently depending on various factors, such as the tumor origin or histology, comutations, and prior therapy approaches. Analysis of the NCI-MATCH clinical trials showed tumor-specific sensitivity in 6 of 10 groups, demonstrating that histology remains an important part of treatment efficacy.⁶

In the KOSMOS pilot, the feasibility of a centralized molecular tumor board was established, with actionable genomic alterations in 75.1% and a molecular profiling-guided therapy rate of 51.3%.⁷ The phase 2 ROME trial showed that genomically matched and tailored therapy significantly improves the rate of response and progression-free survival, when compared to standard care for advanced solid tumors.⁸ Together, genomic matching works best when interpreted in combination with tumor context and patient history.

AI for Molecular Tumor Boards and Trial Matching

Molecular tumor boards—panels of medical experts who meet to analyze a patient’s tumor and biomarker data—require review of sequencing reports, clinical notes, PubMed evidence, variant databases, prior therapy, and available trials. Manual processes are often slow, require substantial resources, and take effort to scale, so new technologies, most notably retrieval-augmented generation (RAG) systems, quickly summarize literature and suggest treatment options based on evidence. In one study, outputs produced by RAG-supported large language models (LLMs) for molecular tumor boards included clinician-equal recommendations and some plausible alternatives, but hallucinations, a phenomenon where AI models generate highly confident and plausible-sounding information that is actually fabricated and incorrect, remained a limitation.⁹

Therefore, AI is extremely useful as an assistant if supervised, but as of now, it cannot be used to make decisions independently. Clinical trial matching is another challenge in medicine due to the complex, unstructured eligibility criteria across trials. LLMs can extract logic and details from oncology trial descriptions, thereby improving the quality and efficiency of matching.¹⁰ Although these techniques can lead to faster trial search and fewer missed options, eligibility also depends on more complex criteria such as prior therapies and performance status, which are more difficult for LLMs to extract, leading to higher chances of errors and hallucinations.

Drug Response Prediction

The presence of mutations does not guarantee a response to treatment, because the state of the tumor is often more important than a single associated biomarker in determining outcomes. Response can be affected by gene expression, copy number, methylation, pathway activity, immune environment, and drug structure, among other factors. Models based on the transcriptome can predict targeted therapy and immunotherapy responses from tumor expression profiles, adding functional context beyond DNA mutations alone.¹¹ The field of deep learning for drug response prediction is still growing rapidly, but benchmarking remains highly inconsistent across different data sets and evaluation metrics.¹²

Transfer learning approaches from large RNA-sequence drug-response data to single-cell data, as seen in scDEAL, help address the issue of limited patient-level training data.¹³ In addition, NDSP is a computational framework that can predict drug sensitivity by combining different data types using similarity network fusion and deep learning.¹⁴ MMCL-CDR is another multimodal model that uses copy number, gene expression, cell morphology, and chemical structure for drug response prediction.¹⁵ However, many of these models face the samechallenge in their evaluation: Retrospective accuracy measurements do not automatically prove clinical benefit without further validation. Once we have prospective validation, these tools can enhance patient outcomes.

Single-Cell and Organoid-Based Personalization

Bulk tumor sequencing averages molecular signals across tumor cell populations, which can hide resistant or sensitive subclones. Single-cell RNA sequencing can identify specific tumor subpopulations and vulnerabilities in individual cells. Tools such as scDrug, scDrug+, and scDR use single-cell transcriptomic data to connect tumor-cell clustering with drug-response prediction, including the ability to estimate response differences across tumor subgroups, and in some cases, incorporate molecular structures of drugs for newer agents.^16-18 More patient-specific approaches are emerging: scTherapy uses single-cell transcriptomes to identify coinhibition strategies tailored to a patient, and PharmaFormer uses transfer learning with patient-derived organoid data to approximate drug response better than traditional cell-line-only models.^19,20

Current barriers to progress include tissue requirements, cost, turnaround time, and standardization. In the future, models should combine sequencing, single-cell profiling, and organoids for more patient-specific treatment, while new methods are in development to improve the efficiency and price of sequencing.

Combination Therapy Optimization

Although single-agent therapy still may select resistant clones, combinations of therapies can affect multiple pathways simultaneously and delay resistance in the body. Testing all drug pairs experimentally is infeasible due to the rapidly expanding number of possibilities when new treatments are approved for a given condition. PDSP is an approach that personalizes drug synergy prediction by incorporating cell-line synergy information alongside data from patient-specific single-drug responses.²¹ Another key tool in the field is SynerGNet, which employs graph neural networks to accurately identify synergistic or antagonistic relationships for anticancer drug combinations.²² MFSynDCP uses graph attention and multisource feature learning to predict synergistic combinations and identify important drug substructures.²³ scTherapy, which was previously mentioned for its use of single-cell transcriptomes, also supports patient-specific coinhibition of malignant clones while considering toxicity.²⁰

These tools can minimize the need for empirical testing of drug combinations and make the evaluation process far more efficient, creating an opportunity to find more effective treatment options by combining existing drugs. The main clinical challenge before more widespread implementation is that the predicted synergy has to be balanced by calculations for toxicity, dosing, sequencing, and patient comorbidities.

Challenges, Ethics, and Implementation

Many AI models are trained on cell lines, public data sets, and retrospective cohorts, which can yield impressive performance metrics, but these results are not representative of real-world prospective patient care. Common problems faced include small cohorts, missing data, batch effects, inconsistent response labels, and sequencing-platform variation, which leads to the use of methods such as data imputation and synthetic data creation, reducing the reliability of models trained on these data sets. Standardized benchmarking remains challenging, making it extremely difficult to accurately compare models.¹² Explainability in AI recommendations is also essential, since clinicians cannot feel confident in using information without valid reasoning.

However, even when justifications are provided, LLMs can create hallucinated citations, which are identified only with clinician oversight.⁹ Ethical concerns regarding AI-augmented treatment include privacy, autonomy, data protection, data set biases, and equitable access.²⁴ Biased training data can worsen existing disparities in health care if models perform better for already represented populations. Lastly, the practical barriers that remain for the use of AI involve reimbursements, infrastructure, sequencing access, and adoption in more community oncology settings.

Future Directions and Conclusion

Future precision oncology will likely have AI models training on more multimodal data from genomics, transcriptomics, organoids, pathology, imaging, and electronic health records. AI-assisted tumor boards will be able to review evidence and prioritize treatment options more efficiently. Additionally, LLMs may mature to a level where they can be used for biomarker extraction, accurate literature review, and trial matching without a high risk of hallucinations. Response-aware treatment selection can advance to a stage where therapy will be chosen based on current sensitivity and a patient’s specific information. One major end point will be the optimization of combination therapies, in which multidrug treatment decisions will be substantially automated and trial-and-error approaches will be largely obsolete. All in all, AI will not be used to replace oncologists, but will be used to help interpret complex data and design more precise, adaptive, personalized cancer treatment.

The authors collaborated through their affiliation with The Cancer & Hematology Centers in Flint, MI.