NEW YORK — A multimodal approach encompassing standard clinical markers, imaging, and genomic features can better predict immunotherapy response among non-small cell lung cancer (NSCLC) patients than any one biomarker alone, a new study has found.
While PD-1/PD-L1 immunotherapies have become a common treatment approach for advanced NSCLC patients, only a portion of patients have a durable response, suggesting to a Memorial Sloan Kettering Cancer Center-led team that better biomarkers to predict response are needed. Using a cohort of nearly 250 patients for whom they had a range of data types, the team generated a multimodal risk prediction model.
"Our study represents a proof of principle that information content present in routine diagnostic data, including baseline CT scans, histopathology slides, and clinical next-generation sequencing can be combined to improve prognostication for response to PD-(L)1 blockade over any one modality alone and over standard clinical approaches," researchers led by MSKCC's Sohrab Shah wrote in their paper, which appeared in Nature Cancer on Monday.
For their analysis, the researchers turned to a cohort of 247 patients with NSCLC who underwent PD-(L)1-blockade-based therapy and for whom they had baseline data on standard clinical biomarkers like PD-L1 tumor proportion score and tumor mutational burden, CT scan data, and sequencing data as well as known outcomes — a quarter of patients in this cohort responded to the therapy.
The researchers first assessed the ability of standard biomarkers to predict treatment response on their own. Most, they noted, had modest ability to identify treatment responders. For instance, CT-based predictions could distinguish responders and non-responders with an area under the curve of 0.64. Genomic predictors alone had similar abilities, with an AUC of 0.65
By combining these factors using a dynamic deep attention-based multiple-instance learning model with masking (DyAM) approach, the researchers could better distinguish responders and non-responders. The DyAM model enabled them to combine multimodal data with different attention weights and could mask factors that are not present in a patient, such when a tumor has no PD-L1 expression.
In all, the researchers generated a DyAM model with an AUC of 0.80. By contrast, if they averaged the different factors, that approach resulted in an AUC of 0.72.
Further, in comparison to established biomarkers of immunotherapy response, the DyAM model performed better, the researchers found. It in particular outperformed measures of tumor mutational burden and PD-L1 immunohistochemistry scores. The DyAM model could further better predict short-term response. At four months, the highest-risk quartile identified by the DyAM model had 79 percent progression, while the lowest-risk quartile had 21 percent progression.
The researchers noted, though, that their analysis has limitations, particularly as it relied on data from one institution to ensure consistent training data quality. The inclusion of external data would have required considerable work to incorporate, they added, noting that multi-institutional training datasets could be incorporated in the future.
"Our DyAM model is a promising approach to integrate multimodal data and future models using larger datasets will make it possible to augment current precision oncology practices in treatment decision-making," they wrote.