Fig. 5

Multimodal model enables fully noninvasive outcome classification. a Two parameters (normalized bTMB and ΔctDNA) were used to build a model for predicting DCB in our discovery cohort, with a sensitivity of 75.0% and a specificity of 88.9%. Stacked column chart showed the proportion of patients predicted to achieve DCB (Pred-DCB) or NDB (Pred-NDB) by the model. Patients with higher predict scores had significantly longer PFS than those with lower scores (median 13.6 months versus 4.3 months). b Performance of the two parameters model in DIREct-On validation cohort, achieving a sensitivity of 80.5% and a specificity of 67.6%. Stacked column chart showing the proportion of patients with Pred-DCB or Pred-NDB. Patients with higher predict scores had longer PFS than those with lower scores (median 8.5 months versus 2.5 months). c Three parameters (normalized bTMB, ΔctDNA and the first RECIST response) were used to build a model for predicting DCB in our discovery cohort, with a sensitivity of 79.2% and a specificity of 86.4%. Stacked column chart showing the proportion of patients with Pred-DCB or Pred-NDB. Patients with higher predict scores had significantly longer PFS than those with lower scores (median 13.5 months versus 4.2 months). d Performance of the three parameters model in DIREct-On validation cohort, achieving a sensitivity of 94.7% and a specificity of 85.3%. Stacked column chart showing the proportion of patients with Pred-DCB or Pred-NDB. Patients with higher predict scores had longer PFS than those with lower scores (median undefined versus 2.2 months). bTMB blood-based TMB, DCB durable clinical benefit, NDB no durable benefit, RECIST The response evaluation criteria in solid tumors, PFS progression-free survival; P < 0.05 represents statistical significance