AUTHORS
Yun Kang, Shuai Ni2, Zhihan Chen3, Fengjiao Zhu4, kejia Kan5, Jia Xu6, Qiaolin Liu7, Lin Liu8, Wei Xie9, Ning Wu10, Heng Mei, MD, PhD11
AFFILIATIONS
1 Hubei Provincial Clinical Medical Research Center of Cell Therapy for Neoplastic Disease, Wuhan, China, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China
2 Waker Bioscience Co. Ltd., Shanghai, China
3 Hubei Provincial Clinical Medical Research Center of Cell Therapy for Neoplastic Disease, Wuhan, China, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China
4 Waker Bioscience Co. Ltd., Shanghai, China
5 Waker Bioscience Co. Ltd., Shanghai, China
6 Hubei Provincial Clinical Medical Research Center of Cell Therapy for Neoplastic Disease, Wuhan, China, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China
7 Hubei Provincial Clinical Medical Research Center of Cell Therapy for Neoplastic Disease, Wuhan, China, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China
8 Hubei Provincial Clinical Medical Research Center of Cell Therapy for Neoplastic Disease, Wuhan, China, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China
9 Hubei Provincial Clinical Medical Research Center of Cell Therapy for Neoplastic Disease, Wuhan, China, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China
10 Waker Bioscience Co. Ltd., Shanghai, China
11 Hubei Provincial Clinical Medical Research Center of Cell Therapy for Neoplastic Disease, Wuhan, China, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China
ABSTRACT
Background
The depth and durability of response to chimeric antigen receptor (CAR) T-cell therapy depend on the dynamics of transduced clones. Disproportionate or sustained outgrowth of vector integration-site-defined clones may reflect altered clonal fitness and raise concerns about insertional mutagenesis and secondary malignancies. Integration-site analysis enables clonal tracking and vector safety evaluation; conventional methods seldom combine genomic and functional context to distinguish expansion from persistence. We developed a seven-domain approach to characterize both behaviors across hematologic malignancies, assign predefined clonal-risk categories, and test associations between integration-site features and relapse in longitudinal relapsed/refractory multiple myeloma (R/R MM) samples.
Methods
Clonal characterization included 671 samples comprising 2,164,635 integration sites (ISs). IS_Ratio was defined for each IS within a sample as Unique_IS_count/Total_Seq_Count. Dominant clones (DCs) had IS_Ratio ≥0.20; lower values defined nondominant comparators. Maximum and Top1 IS ratios represented the most abundant clone; gene evenness summarized integration burden across genes.
Long-term persistent clones (LTPCLs) were ISs detected at ≥12 months, and short-term persistent clones (STPCLs) were ISs detected earlier; both were treated as time-point-based descriptive groups. DC-specific exon bias was evaluated using unadjusted exon-to-intron frequencies and gene-level odds ratios (ORs).
Seven domains covered clonal structure; genomic, functional, and epigenetic context; tissue expression; oncogenic potential; and internal DC/comparator genes.
Pre-existing labels combined dominance and persistence and were binarized as High_Risk versus Low/Medium_Risk. Feature filtering and eight-model comparison were confined to the internal training cohort; the selected model was then locked. Youden’s J from internal-training leave-one-out cross-validation (LOOCV) predictions set the 0.67 threshold used in the 89- and 800-sample test cohorts. L1-regularized logistic regression predicted relapse in 54 outcome-annotated samples from 24 patients with R/R MM using sample-level LOOCV.
Results
The approach generated 105 IS features; 25 remained after variance filtering, correlation pruning, and Wilcoxon rank-sum testing. High-risk profiles showed a greater difference in integration burden between predefined high-risk candidate genes and the low-risk reference gene (P=2.3×10?5), higher maximum and Top1 IS ratios (P=1.8×10?4 and P=3.1×10?4), and lower gene evenness (P=5.7×10?³), indicating concentrated clonal expansion. Unadjusted gene-level exon-to-intron analysis identified 20 genes with DC-specific exon enrichment, including CD8A (OR, 6.49; P=.0022) and IFIH1 (OR, 43.13; P=.0019); these genes mapped to immune-signaling and chromatin-regulatory pathways. Among 18,858 RefSeq genes, 77.0% had higher raw occurrence counts in the ≥12-month LTPCL group than in the STPCL group. LTPCL analyses were descriptive and excluded from classifier development.
Against predefined clonal-risk labels, the classifier achieved an area under the receiver operating characteristic curve (AUC-ROC) of 0.954 (95% CI, 0.925–0.983), 95.5% accuracy, 95.5% sensitivity, and 95.6% specificity in the 89-sample internal test set. In the independent multicenter cohort of 800 samples (117 high-risk, 683 low-risk; 2,869,464 ISs), it achieved an AUC-ROC of 0.977 (95% CI, 0.968–0.986), with 78.6% sensitivity and 98.3% specificity at the fixed threshold of 0.67.
The R/R MM relapse model assigned the largest-magnitude coefficients to maximum and Top1 IS ratios, cancer-gene integration burden, lower clonal evenness, and DC-derived candidate-gene features involving NOTCH1, SMARCA4, and RB1. Under sample-level LOOCV, it achieved an AUC-ROC of 0.8788, an AUPRC of 0.9363, 74.07% accuracy, 81.82% sensitivity, 61.90% specificity, and an F1 score of 0.7941.
Conclusions
Across diseases, multidimensional IS features identified concentrated expansion and more frequent gene occurrence among clones detected at ≥12 months. The locked classifier retained discrimination in the held-out internal and multicenter external cohorts, and the R/R MM model discriminated relapse status at the sample level. These results provide a quantitative basis for vector-safety monitoring and post–CAR-T relapse risk assessment.
DISCLOSURES
Yun Kang: No relevant conflicts of interest to declare.