Virtual tumor model predicts combination therapy candidates and radiation response in NSCLC
Background: Exploring treatment combinations that reflect genetic differences
A “virtual tumor” model that recreates key signaling pathways in lung adenocarcinoma, a subtype of non-small cell lung cancer (NSCLC), predicted combination therapy candidates and responses to radiation. The researchers explored about 66,000 treatment combinations and compared selected predictions with experiments in cancer cell lines and normal lung organoids, as well as with patient data. This was not a clinical trial in which patients were assigned treatments to assess their effects.
Even within the same type of lung cancer, mutations that drive a tumor can change how it responds to drugs and radiation. Testing every possible combination would be costly and time-consuming. This study incorporated biological relationships linking mutations to treatment responses into a model, aiming to narrow down which combinations to test first.
Key findings: Predicting radiosensitization and combination effects with a discrete network
Using information from 171 publications, the researchers connected 160 nodes representing genes, proteins, and biological processes through 360 regulatory relationships. Each node’s activity was represented on a four-level scale from 0 to 3. This qualitative network examines how signaling relationships affect cell proliferation and death; it does not calculate molecular binding energy or drug concentrations.
The model was calibrated using 69 experimental conditions gathered from 26 separate publications. It matched the direction of response in 91% of those conditions. This describes its performance in reproducing experiments used for calibration; it does not represent accuracy in predicting patient treatment responses. Combination effects were assessed as additive effects, defined here as how much better a combination performed than the most effective single treatment. Because continuous dose information was unavailable, pharmacological synergy was not estimated.
Among radiosensitization candidates, loss of 53BP1 increased cancer-cell sensitivity to radiation in a focused CRISPR gene screen, but the same sensitization was not observed in the normal lung organoids used for comparison. However, the predictions did not fully match the experimental results regarding which cancer cell lines showed an effect. Loss of ATM induced sensitization in the tested A549 and SW1573 cells regardless of p53 status. ATM loss also sensitized normal lung organoids, making it difficult to interpret this finding as evidence of cancer-cell selectivity.
In the search for drug-target combinations, simultaneous inhibition of MYC and ATR emerged as a promising candidate. The model predicted that this combination would restrict cancer-cell proliferation and increase DNA damage-related cell death, while having a relatively smaller effect on normal cells. The proposed explanation is that MYC inhibition weakens DNA repair, while ATR inhibition disrupts the response to stress during DNA replication. This mechanism and selectivity are computational predictions; they have not been confirmed as clinical efficacy of an actual drug combination.
Meaning and outlook: A 19-gene signature remains a candidate for further validation
A signature comprising 19 genes derived from the model was also evaluated using lung adenocarcinoma and lung squamous cell carcinoma data from The Cancer Genome Atlas (TCGA). Classification as likely to be more radiation-sensitive based on combined gene-expression and copy-number data was associated with longer survival after radiation therapy. The hazard ratio in the survival analysis was 0.5842, with a P value of 0.03. By contrast, separate analyses of expression and copy number yielded P=0.06 and P=0.08, respectively, which did not meet the usual threshold for statistical significance. Among patients who did not receive radiation, there was no significant survival difference between the classified groups.
This retrospective analysis of patient data points to the potential of a biomarker for radiation response. Some subgroups were small, limiting statistical power, and the study did not prospectively establish its usefulness for treatment selection. The model simplified normal cells and did not include the tumor microenvironment or metabolic and epigenetic mechanisms. Its current steady-state approach also cannot model fractionated radiation. In addition, the inhibitory effects of actual drugs may be weaker than the complete loss represented in the model. Follow-up studies need to assess both the strength of candidate drugs’ inhibitory effects and their toxicity to normal tissues.
The methods and results are available in the full Cancer Research paper.
The disease burden from non-small cell lung cancer (NSCLC) adenocarcinoma is substantial, with a million new cases diagnosed globally each year and a 5-year survival rate of less than 20%. The lack of therapeutic options personalized to individual patients leads to high variation in survival. The combination of patient stratification with personalized treatment has the potential to improve outcomes; however, the variation in mutations found in NSCLC adenocarcinoma patients makes experimentally determining treatment combinations time-consuming and expensive. Here, we developed an interpretable mechanistic model to decipher complex signaling interplay and guide personalized therapy in NSCLC adenocarcinoma. This 'virtual tumor' model encompassed key tumor intrinsic oncogenic signaling pathways, for efficiently predicting rational drug-drug and drug-radiotherapy combination therapies in NSCLC. Diverse genetic profiles were simulated for testing over 10,000 therapeutic strategies to identify optimal approaches to overcome resistance mechanisms specific to genetic profiles and p53 status. The virtual tumor model reproduced drug additivity screens, predicted radio-sensitizing genes validated in a CRISPR screen, and identified 53BP1 as a potential drug target that improved the therapeutic window during radiotherapy. A 19-gene signature derived from the virtual tumor framework stratified patients most likely to benefit from radiotherapy, which was validated using TCGA data. These results demonstrate the utility of virtual tumors to predict effective therapeutic combinations and present a computational resource for large-scale screening of personalized therapies to guide clinical decision-making in NSCLC patients.
The model’s practical value lies in prioritizing treatment combinations for experiments. For example, researchers could compare combined MYC and ATR inhibition or radiosensitization candidates in virtual tumors reflecting lung-cancer mutation profiles, then design studies that administer actual drugs to cancer cells and normal lung organoids to assess effects and toxicity. This is a possible follow-up application proposed on the basis of the study’s findings.
The 19-gene signature is also a candidate biomarker for distinguishing radiation responses, but it has not been established as a test for deciding whether patients should receive treatment. Its predictive performance must be assessed in independent patient cohorts and prospective studies. Researchers also need to evaluate whether toxicities overlap when radiation, sensitizing agents, and existing anticancer drugs are used together before its clinical potential can be judged.