Yang Xia

Background: Pathological response following neoadjuvant chemoimmunotherapy is a robust prognostic predictor for non-small cell lung cancer (NSCLC). It remains impossible for preoperative non-invasive assessment. Tumor treatment response and host immune status are the major contributors to chemoimmunotherapy. Thus, we develop a multimodal deep learning framework that integrates longitudinal CT imaging with host body composition dynamics to guide optional neoadjuvant plan. Methods: 901 patients with stage II-IIIB NSCLC undergoing surgical resection following neoadjuvant immunochemotherapy were collected from five institutions, consisting of a training cohort (n=686), validation cohort (n=120), and two test cohorts (n=70 and 25). The Deep Host-Tumor Dynamic(DeepHTD) platform, integrating immuno-metabolic and tumor evolutionary modules, which was design to automatically extract and analyze body composition dynamics and longitudinal tumor imaging features using multi-head attention neural networks and learnable principal component analysis (L-PCA). The model provides real-time prediction of major pathological response (MPR). Results: The DeepHTD platform demonstrated robust predictive performance with AUCs of 0.847. The Deep-Tumor module along (excluding immuno-metabolic module) resulted in reduced AUCs (declined from 0.847 to 0.809). the Deep-Host module along (excluding tumor evolutionary module) coursed a performance drop (AUC:0.708). Moreover, there were significant difference in EFS between MPR and non-MPR. The HRs were 0.349 in validation set and 0.332 in test cohorts. Conclusion: The DeepHTD platform, combines host-tumor dynamics during neoadjuvant chemoimmunotherapy,offers an accurate and robust tool for the preoperative prediction of MPR. It guided the selection of surgical timing.

Intracranial efficacy of crizotinib and optimal postprogression therapies in c‐ros oncogene 1 (ROS1)‐positive non–small cell lung cancer(NSCLC) remain under‐explored. We performed a multicenter real‐world study across 17 Chinese hospitals (2015–2024) in 325 ROS1‐positive NSCLC patients to evaluate the intracranial efficacy of crizotinib and postprogression treatment strategies. Crizotinib achieved an overall response rate(ORR) of 78.5% and a median progression‐free survival (PFS) of 22.4 months. Patients with baseline brain metastases had a 66.7% intracranial response rate and a median central nervous system (CNS)‐PFS of 22.2 months, demonstrating intracranial benefit. After progression, next‐generation ROS1 inhibitors (active against the G2032R resistance mutation, for example, taletrectinib and repotrectinib) significantly prolonged PFS compared to older‐generation targeted agents (median 13.9 vs. 4.4 months), especially among patients with secondary ROS1 resistance mutations (ORR 66.7% vs. 14.3%). However, due to greater availability in China, earlier agents like lorlatinib and entrectinib were commonly used after crizotinib failure. In our cohort, lorlatinib demonstrated superior efficacy to entrectinib, with a trend toward longer PFS and higher overall (34% vs. 7%) and intracranial (50% vs. 0%) response rates. This large real‐world study confirms crizotinib's intracranial benefit and underscores the importance of next‐generation ROS1 inhibitors and older‐generation agents like lorlatinib—for improved patient outcomes after crizotinib resistance.

Background: Transbronchial lung cryobiopsy (TBLC) is increasingly used to diagnose interstitial lung disease (ILD). However, the learning curve associated with this technique remains unclear. Methods: This retrospective analysis included the first 100 patients suspected of having ILD for whom TBLC was performed by six physicians (three chief physicians and three attending physicians) across three centers in China. The diagnostic yield, procedural information, and complications were compared between consecutive groups of patients. Logistic regression was employed to analyze the median diagnostic yields across ten subgroups of operative experience, categorized in intervals of ten patients, to delineate the learning curve. Results: The overall multidisciplinary discussion (MDD)-based diagnostic rate for ILD after TBLC was 75% (450 of 600 procedures). The diagnostic yield increased with time. Learning curves revealed that chief physicians achieved proficiency after approximately 60 procedures and attending physicians after approximately 90. The sample biopsy was larger for chief physicians than for attending physicians (median, 19 mm2 vs. 15.25 mm2, P<0.001). The complication rates were acceptable, with no significant difference between the chief and attending physicians over their first 100 procedures. Conclusions: TBLC is a safe and effective diagnostic approach for ILD, but operators require extensive training to achieve proficiency. Senior physicians reached competency earlier than junior colleagues (after several dozen procedures), underscoring the importance of structured training programs and standardized protocols for facilitating safe implementation.

Background: Neoadjuvant immunochemotherapy achieves higher pathological complete response (pCR) rates than chemotherapy alone in limited-stage Small Cell Lung Cancer (LS-SCLC). However, whether pCR translates into a survival benefit remains unclear, and the underlying biological mechanisms have yet to be investigated.Patients and Methods: This is a multicenter retrospective cohort study. 85 patients with stage I-III LS-SCLC who underwent radical resection following neoadjuvant immunochemotherapy (NIC, n=28) or neoadjuvant chemotherapy (NC, n=57) were included. The primary objectives were to compare pCR rates between treatment groups and to assess the prognostic impact of pCR stratified by treatment group, using event-free survival (EFS) and overall survival (OS) as clinical endpoint. Seven-plex multiplex immunofluorescence (mIF) was performed on nine surgical specimens selected via purposive sampling. Public transcriptomic datasets (GSE281525, GSE60052) served as independent validation.Results: The pCR rate in the NIC group was significantly higher than the NC group (35.7% vs 14.0%, P=0.027). With the median follow-up of 20.8 months in the NIC group, pCR patients had longer EFS than non-pCR patients (not reached vs 20.3 months, P=0.007). No such difference was seen in the NC group (19.8 vs 18.0 months, P=0.945). Spatial immune profiling revealed that NIC pCR tumors were uniformly immune-hot (2/2) with high CD8+ T cell infiltration and low PD-1 exhaustion. Nearest-neighbor distance analysis revealed closer macrophage-CD8 T cell co-localization in NIC pCR compared with NC pCR specimens. Transcriptomic analysis of independent cohorts confirmed that CD8 activation and composite antitumor immune scores were significantly associated with favorable survival.Conclusions: pCR after neoadjuvant immunochemotherapy, but not chemotherapy, translates to survival benefit in LS-SCLC. Spatial immune profiling reveals distinct immune-mediated versus drug-mediated mechanisms underlying SCLC tumor clearance. These findings highlight the need to consider treatment type when interpreting pCR in SCLC and provide rationale for integrating spatial immune biomarkers in prospective neoadjuvant trials.

Background: Pathological response following neoadjuvant chemoimmunotherapy is a robust prognostic predictor for non-small cell lung cancer (NSCLC), however, which remains impossible for preoperative non-invasive assessment. Since radiographic response often diverges from pathological regression, we aimed to develop an interpretable dynamic multimodal deep-learning framework that integrates longitudinal tumor imaging with host immuno-metabolic (IM) surrogates to predict pathological response after neoadjuvant chemoimmunotherapy. Methods: 901 patients with stage II-IIIB NSCLC undergoing surgical resection following neoadjuvant immunochemotherapy were collected from five institutions, consisting of a training (n=686), a validation (n=120), and two test cohorts (n=70 and 25). Deep HostTumor Dynamic (DeepHTD) platform combines a tumor-evolutionary (TE) module derived from serial CT imaging with an IM module incorporating fully automated body-composition features, laboratory variables, and clinical characteristics. The model provides real-time prediction of major pathological response. Findings: Thirteen immuno-metabolic patterns and 128-dimensional image features were eventually employed to establish DeepHTD. The platform demonstrated robust predictive performance across all cohorts, with AUCs of 0·851 in the training, 0·847 in validation, 0·876 and 0·849 in two test cohorts. Removing either IM or TE module consistently impaired predictive performance, supporting the complementary value of host-side and tumor-side dynamics. DeepHTD also outperformed static baseline or preoperative tumor imaging models. Model predicted response groups were significantly associated with EFS in the training, validation, and test cohorts. Decision curve analysis further suggested potential clinical utility for guiding surgery versus continued-therapy decisions after neoadjuvant treatment. Interpretation: We developed an interpretable platform integrating longitudinal tumor imaging and host immuno-metabolic surrogates, enabling dynamic prediction of MPR after before surgery. This multimodal deep-learning approach may inform personalized neoadjuvant strategies and surgical timing, pending prospective validation.

Although immunochemotherapy sheds light on neoadjuvant strategies, about two‐thirds of patients still respond poorly to perioperative chemoimmunotherapy. Hence, it is crucial to investigate the underlying response mechanism to improve the prognosis of these patients. In this study, we utilized paired pre‐ and post‐neoadjuvant immunochemotherapy samples from non‐small‐cell lung cancer (NSCLC) patients with single‐cell RNA and T‐cell receptor (TCR) sequencing to characterize the dynamic changes of T cells in tumor microenvironment. Within nine enrolled patients with distinct pathological assessments, we identified bi‐directional mechanisms associated with their pathological responsiveness. One is mediated by a batch of CD8 ⁺ T‐cell subsets such as effector memory T cells (Tem), effector T cells (Teff), tissue‐resident memory T cells (Trm), and exhausted T cells (Tex), exhibiting higher TCR clonality and diversity in responders. CD8 ⁺ Tem cells with both novel and pre‐existing TCR clonal expansion patterns particularly contributed to improved pathological responses. The other mechanism is through inhibitory Tregs, which showed more novel clonal expansion and enhanced functional profiles in nonresponsive tumors. In conclusion, our findings proposed the bidirectional characteristics of T‐cell dynamics for in‐depth interpretation of responding mechanisms to neoadjuvant immunochemotherapy of NSCLC.