PO.TB03.01 · 肿瘤生物学
拦截转移：通过多模态基础模型进行8步CRISPR设计
Intercepting metastasis: 8-step CRISPR design via multi-modal foundation models
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：转移占癌症死亡的>90%，然而CRISPR设计工具仅优先考虑靶向效率（CRISPOR、Benchling），忽略了阶段特异性生物学（EMT、外渗、定植）。现有工具均未将必需性（DepMap）、染色质可及性（ENCODE）和转移级联相关性整合到向导选择中。我们提出Interception，一个针对转移级联各阶段脆弱性的阶段感知CRISPR框架。
方法：我们计算三个正交信号：(i) 功能性（Evo2）：预测变异对蛋白质功能的影响；(ii) 必需性（DepMap）：优先考虑对转移存活至关重要的基因；(iii) 调控性（GTEx eQTLs）：预测表达变化。Target-Lock评分使用文献衍生的权重量化基因与特定转移阶段（如EMT与外渗）的相关性。在PAM约束下生成向导候选，对其疗效（通过sigmoid变换的Evo2预测表达变化）和安全性（采用指数错配衰减的minimap2全基因组比对）评分，然后按Assassin评分（0.40×疗效 + 0.30×安全性 + 0.30×任务契合度）排序。权重通过对304个基因-阶段对的消融研究进行优化。所有输出均包含完整来源信息，并可通过冻结的脚本/环境重现。
结果：Target-Lock显著优于单指标基线：AUROC 0.976（本研究）对0.61（单独DepMap）、0.58（单独Evo2）、0.52（随机）；AUPRC 0.948对0.42（DepMap）、0.39（Evo2）；Precision@3 = 1.000（完美的前3富集）对0.33（随机）。所有8个转移步骤均显示显著富集（8步中6步的Fisher精确检验p < 0.001），效应量大（Cohen's d > 2.0）。对20个设计的向导验证显示平均疗效0.548 ± 0.119，安全性0.771 ± 0.210。通过AlphaFold 3 Server对15个向导:DNA复合物进行的结构验证实现了100%的通过率（pLDDT 65.6 ± 1.8，iPTM 0.36 ± 0.01）——这是首个具有结构验证的转移靶向CRISPR框架。
结论：Interception为转移提供了可重现、任务感知的CRISPR设计，整合了多模态信号、全基因组安全性和结构验证。未来工作将 (i) 整合Enformer进行染色质预测，(ii) 将结构验证扩展至40个向导（完成8步全覆盖），以及 (iii) 在转移性小鼠模型（PDX/CDX）中进行验证以评估体内疗效。
查看英文原文 English abstract
Background: Metastasis accounts for >90% of cancer deaths, yet CRISPR design tools prioritize on-target efficiency alone (CRISPOR, Benchling), ignoring stage-specific biology (EMT, extravasation, colonization). No existing tool integrates essentiality (DepMap), chromatin accessibility (ENCODE), and metastatic cascade relevance into guide selection. We present Interception, a stage-aware CRISPR framework targeting vulnerabilities across the metastatic cascade.
Methods: We compute three orthogonal signals: (i) Functionality (Evo2): predict variant impact on protein function; (ii) Essentiality (DepMap): prioritize genes critical for metastatic survival; (iii) Regulatory (GTEx eQTLs): predict expression changes. Target-Lock score quantifies gene relevance to specific metastatic stages (e.g., EMT vs. extravasation) using literature-derived weights. Guide candidates are generated with PAM constraints, scored for efficacy (Evo2-predicted expression change via sigmoid transformation) and safety (minimap2 genome-wide alignment with exponential mismatch decay), then ranked by Assassin score (0.40×efficacy + 0.30×safety + 0.30×mission fit). Weights were optimized via ablation study on 304 gene-stage pairs. All outputs include full provenance and are reproducible via frozen scripts/environment.
Results: Target-Lock significantly outperformed single-metric baselines: AUROC 0.976 (ours) vs. 0.61 (DepMap alone), 0.58 (Evo2 alone), 0.52 (random); AUPRC 0.948 vs. 0.42 (DepMap), 0.39 (Evo2); Precision@3 = 1.000 (perfect top-3 enrichment) vs. 0.33 (random). All 8 metastatic steps showed significant enrichment (Fisher's exact p < 0.001 for 6/8 steps) with large effect sizes (Cohen's d > 2.0). Guide validation on 20 designs showed mean efficacy 0.548 ± 0.119, safety 0.771 ± 0.210. Structural validation of 15 guide:DNA complexes via AlphaFold 3 Server achieved 100% pass rate (pLDDT 65.6 ± 1.8, iPTM 0.36 ± 0.01)-the first metastasis-targeted CRISPR framework with structural validation.
Conclusions: Interception delivers reproducible, mission-aware CRISPR design for metastasis, integrating multi-modal signals, genome-wide safety, and structural validation. Future work will (i) integrate Enformer for chromatin predictions, (ii) expand structural validation to 40 guides (complete 8-step coverage), and (iii) validate in metastatic mouse models (PDX/CDX) to assess in vivo efficacy.
利益披露 Disclosure
F. Kiani, None.