Scholarlink

Purpose <p>Accurate quantitative survival prediction in advanced non-small cell lung cancer (NSCLC) remains an unmet clinical need. While liquid biopsy is widely used, single circulating tumor DNA (ctDNA) shows limited predictive power. We developed an interpretable deep-learning model to quantitatively predict outcomes.</p> Methods/patients <p>We integrated data from 1373 advanced NSCLC patients profiled by two ultra-deep ctDNA sequencing assays (MSK-ACCESS and ctDx Lung). Features associated with overall survival (OS) were incorporated into a deep-learning network (DeepSurv), which estimates time-to-event survival probabilities. Model performance was evaluated by time-dependent area under the curve (AUC). SHapley Additive exPlanations (SHAP) were employed to interpret model output.</p> Results <p>A total of 1373 patients were analyzed, with 1012 using MSK-ACCESS (discovery) and 361 using ctDx Lung (validation). Among over 40 clinicopathological features, ctDNA status, cell-free DNA (cfDNA) concentration, age, blood-based <i>TP53</i>, <i>EGFR</i>, <i>PIK3CA</i>, <i>ARID1A</i>, <i>STK11</i> and <i>MET</i> mutations significantly predicted OS. In ctDNA-positive patients, <i>TP53/PIK3CA/ARID1A/STK11/MET</i>-mutated patients had significantly inferior OS compared with wildtype patients (<i>P</i> &lt; 0.001). Using above variables, DeepSurv was trained and tested in the MSK-ACCESS cohort (12-month AUC = 0.75), outperforming single cfDNA (AUC = 0.66) or ctDNA (AUC = 0.59), and externally validated in the ctDx Lung cohort. Compared with high-risk patients, DeepSurv-identified low-risk patients had significantly longer OS in both discovery (12-month OS 87.8% vs 53.8%, HR 0.32, <i>P</i> &lt; 0.001) and validation cohorts (73.2% vs 48.4%, HR 0.42, <i>P</i> &lt; 0.001). SHAP revealed <i>TP53</i> and cfDNA concentration &gt; 4.8&#xa0;ng/mL had the most important contributions.</p> Conclusions <p>The interpretable DeepSurv model, integrating multimodal features, enables quantitative survival prediction and risk stratification in advanced NSCLC, facilitating personalized decision-making.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable deep learning model of circulating genomics for quantitative survival prediction in advanced non-small cell lung cancer

  • 期刊论文

  • 发表日期：

    2026年2月6日

查看全文
  • Yu Wang,
  • Yi-Tong Li,
  • Ming-Hao Wang,
  • Cheng-Yi Zhang,
  • Ying Jiang,
  • Qi Xu,
  • Ying-Ping Liu,
  • Can-Jun Li,
  • Ye-Xiong Li,
  • Nan Bi

摘要

Purpose

Accurate quantitative survival prediction in advanced non-small cell lung cancer (NSCLC) remains an unmet clinical need. While liquid biopsy is widely used, single circulating tumor DNA (ctDNA) shows limited predictive power. We developed an interpretable deep-learning model to quantitatively predict outcomes.

Methods/patients

We integrated data from 1373 advanced NSCLC patients profiled by two ultra-deep ctDNA sequencing assays (MSK-ACCESS and ctDx Lung). Features associated with overall survival (OS) were incorporated into a deep-learning network (DeepSurv), which estimates time-to-event survival probabilities. Model performance was evaluated by time-dependent area under the curve (AUC). SHapley Additive exPlanations (SHAP) were employed to interpret model output.

Results

A total of 1373 patients were analyzed, with 1012 using MSK-ACCESS (discovery) and 361 using ctDx Lung (validation). Among over 40 clinicopathological features, ctDNA status, cell-free DNA (cfDNA) concentration, age, blood-based TP53, EGFR, PIK3CA, ARID1A, STK11 and MET mutations significantly predicted OS. In ctDNA-positive patients, TP53/PIK3CA/ARID1A/STK11/MET-mutated patients had significantly inferior OS compared with wildtype patients (P < 0.001). Using above variables, DeepSurv was trained and tested in the MSK-ACCESS cohort (12-month AUC = 0.75), outperforming single cfDNA (AUC = 0.66) or ctDNA (AUC = 0.59), and externally validated in the ctDx Lung cohort. Compared with high-risk patients, DeepSurv-identified low-risk patients had significantly longer OS in both discovery (12-month OS 87.8% vs 53.8%, HR 0.32, P < 0.001) and validation cohorts (73.2% vs 48.4%, HR 0.42, P < 0.001). SHAP revealed TP53 and cfDNA concentration > 4.8 ng/mL had the most important contributions.

Conclusions

The interpretable DeepSurv model, integrating multimodal features, enables quantitative survival prediction and risk stratification in advanced NSCLC, facilitating personalized decision-making.