PerturbRx: Learning Treatment-Conditioned Latent Transitions for Patient Drug Response Prediction
PerturbRx learns drug-induced latent transitions, boosting patient drug response prediction accuracy.
Key Findings
Methodology
PerturbRx employs a two-stage framework: source domain training on large-scale single-cell perturbation data to learn drug- and dose-conditioned latent transition models, using multi-objective loss with MMD and cosine constraints. The trained model is then transferred to patient data, where it generates drug-induced transition features for response prediction. Key components include scFoundation-based cell embedding, ChemBERTa drug representations, and a multi-target loss to ensure domain generalization.
Key Results
- On TCGA and PDX datasets, PerturbRx achieved AUROCs of 0.626 and 0.692, respectively, outperforming baselines like CODE-AE and DeepSADR, demonstrating strong transferability and predictive power.
- Incorporating pretrained transition features improved AUROC by approximately 0.04 over static features, confirming the value of dynamic transition modeling.
- Ablation studies showed that pretrained transition representations outperform random or shuffled controls, emphasizing their importance for accurate response prediction.
Significance
This work advances personalized cancer therapy by modeling drug effects as dynamic latent state transitions, addressing the limitations of static molecular features. Its cross-domain transfer capability bridges preclinical and clinical data, enabling more accurate predictions even with limited patient data, thus pushing the frontier of precision oncology.
Technical Contribution
PerturbRx introduces a novel drug- and dose-conditioned latent transition predictor trained on unpaired single-cell data, leveraging multi-objective optimization for distributional alignment and transferability. Its core innovation lies in learning dynamic molecular changes rather than static features, providing a more informative representation for drug response prediction. This approach opens new avenues for dynamic modeling in biomedical applications.
Novelty
This is the first framework to explicitly learn and transfer drug-induced latent state transitions from large-scale single-cell perturbation atlases to patient data, moving beyond static feature-based models. It uniquely combines perturbation atlases with domain transfer techniques, enabling dynamic response modeling in clinical settings.
Limitations
- Relies heavily on large-scale single-cell perturbation datasets, which are costly and may not fully capture patient heterogeneity.
- Currently limited to single-drug, single-dose scenarios; multi-drug interactions and temporal dynamics are not yet modeled.
- Performance in rare or extreme cases remains untested; clinical validation is needed for real-world deployment.
Future Work
Future efforts include integrating multi-drug and temporal data, enhancing model interpretability, and validating predictions with prospective clinical trials. Combining multi-modal data such as imaging and clinical metrics could further improve accuracy and robustness.
AI Executive Summary
PerturbRx represents a significant step forward in personalized cancer treatment prediction. Traditional models primarily rely on static molecular features, which often fail to capture the complex biological responses to therapy. Recognizing this gap, the authors developed a novel framework that explicitly models drug-induced molecular changes as latent state transitions. Using large-scale single-cell perturbation data, PerturbRx trains a drug- and dose-conditioned transition predictor, employing multi-objective optimization to ensure robust transferability across domains.
The core innovation lies in learning dynamic, intervention-specific latent transitions that reflect how cells respond to drugs. Once trained, the model is transferred to patient data, where it predicts how a patient's molecular profile would change under a given treatment. These predicted transitions, combined with static patient and drug features, significantly improve the accuracy of drug response predictions. Extensive experiments on TCGA and PDX datasets demonstrate that PerturbRx outperforms existing methods, achieving AUROCs up to 0.692 and 0.787 respectively.
This approach opens new avenues for precision oncology by enabling models to understand and predict the molecular effects of drugs dynamically. It addresses key limitations of static feature-based models, providing a more mechanistic understanding of treatment response. While promising, the method's reliance on large perturbation datasets and its current focus on single-drug scenarios highlight areas for future research. Overall, PerturbRx paves the way for more accurate, generalizable, and mechanistic models in cancer therapy prediction, with potential applications extending to drug development and personalized treatment planning.
Deep Analysis
Background
Cancer治疗中的药物反应预测一直是个挑战。早期方法主要依赖静态的分子特征，难以反映药物作用引起的动态分子变化。随着单细胞测序技术的发展，研究者开始利用单细胞扰动数据（如scGen、CPA）模拟药物作用的细胞状态变化，试图捕获药物的动态影响。近年来，深度学习模型如CODE-AE、DeepSADR在跨域迁移方面取得一定进展，但仍受限于静态特征和缺乏对药物诱导状态变化的显式建模。现有方法多关注特征提取和域适应，未能充分利用单细胞扰动数据中的动态信息，限制了其临床应用潜力。
Core Problem
现有模型多基于静态分子特征，难以反映药物引起的分子状态变化，导致预测准确性和泛化能力不足。缺乏有效利用单细胞扰动数据中动态信息的机制，限制了模型在实际临床中的应用。如何从无配对的单细胞扰动数据中学习药物诱导的潜在状态转移，并迁移到患者数据中，成为亟待解决的核心难题。
Innovation
PerturbRx的创新在于引入药物-剂量条件下的潜在状态转移学习框架，利用大规模单细胞扰动数据训练转移模型，结合多目标损失（MMD、余弦相似性）确保模型在不同域的泛化能力。其核心创新包括：
- �� 设计药物-剂量条件的潜在转移预测机制，捕获药物作用引起的细胞状态变化；
- �� 利用无配对的细胞群体数据，通过分布匹配实现模型训练；
- �� 将预训练模型迁移到患者层面，利用患者预处理表达谱预测药物诱导的潜在转移。这些创新突破了静态特征的局限，为个性化治疗提供了新思路。
Methodology
- �� 采集Tahoe-100M单细胞扰动数据，利用scFoundation编码器将细胞表达谱映射到高维潜在空间。
- �� 设计药物-剂量条件编码器g，将SMILES字符串和剂量信息转化为潜在条件向量。
- �� 构建药物-剂量条件下的潜在转移预测模型Pθ，输入控制状态zX和药物条件αu，输出细胞状态变化向量。
- �� 使用多目标损失（MMD、余弦相似性、均方误差）训练模型，确保模型在不同数据域的迁移能力。
- �� 在源域训练后，将模型迁移到患者数据，利用患者表达谱潜在嵌入预测药物诱导的转移。
- �� 最后，将转移特征与患者和药物表示结合，用于二分类药物反应预测，训练响应头hψ。
Experiments
- �� 采用TCGA-186、TCGA-508和独立的PDX数据集，验证模型迁移和泛化能力。
- �� 比较PerturbRx与多种基线模型（CODE-AE、DeepSADR、TransDRP、WISER），指标包括AUROC和AUPRC。
- �� 进行消融实验，验证预训练转移特征的贡献，控制随机和扰动转移的影响。
- �� 采用不同划分策略（患者组别、药物留出）确保模型稳健性。
Results
- �� 在TCGA-186中，PerturbRx的平均AUROC达0.626，AUPRC为0.600，优于对比模型如WISER（AUROC 0.580）和CODE-AE（AUROC 0.534）。
- �� 在TCGA-508中，模型表现更佳，AUROC达0.692，AUPRC达0.787，显示出强大的跨患者迁移能力。
- �� 消融分析显示，预训练的转移特征优于随机或拼接控制，验证其在预测中的关键作用。
- �� 预测的潜在转移在响应组和非响应组中表现出不同的几何特征，支持其临床潜在应用。
Applications
- �� 该模型可用于临床药物反应预测，帮助制定个性化治疗方案，特别在缺乏配对样本时。
- �� 未来可结合多模态数据（影像、临床指标）提升预测性能，推动精准医疗。
Limitations & Outlook
- �� 依赖大量单细胞扰动数据，成本高，实验条件差异可能影响迁移效果。
- �� 目前仅考虑单药单剂量，未充分建模多药联合和时间动态变化。
- �� 在极端或罕见病例中的表现尚未验证，需结合临床多模态信息优化。
Plain Language Accessible to non-experts
想象你在厨房做菜，每次用不同的调料（药物）会让菜的味道发生变化。以前的方法就像只看菜的原始样子（预处理数据），难以预测加调料后会变成什么味道。而PerturbRx就像学会了调料如何影响菜的味道（潜在状态转移），通过分析很多调料对菜的影响，建立了一个“调料-菜变味”的模型。这样，即使没有实际试验（没有配对样本），也能预测菜变味的趋势，帮厨师（医生）选择最合适的调料组合，做出美味佳肴（个性化治疗）。它就像掌握了调料的秘密武器，让厨房变得更智能、更高效。
ELI14 Explained like you're 14
想象你在学校的实验室里做科学实验。每次你用不同的化学品（药物）处理一堆细胞（像小实验品），会让它们变得不一样。以前的科学家只看细胞原本的样子（预处理数据），想猜猜加了化学品后会发生什么，但这个猜测很难，因为每个细胞都不一样。而PerturbRx就像学会了化学品怎么影响细胞的“内部状态”，它通过分析很多实验数据，学习到药物引起的细胞变化的“秘密”。然后，把这个秘密带到实际的病人身上，帮医生预测药物对病人的效果。就像你学会了化学品的魔法，可以提前知道药物会让细胞变成什么样，从而帮病人找到最合适的药。这个方法让药物治疗变得更聪明、更精准，也让科学家更了解药物的作用机制。
Glossary
潜在转移 (Latent Transition)
指药物作用引起的细胞状态变化的隐含表示，反映药物影响下的分子动态。
用于描述PerturbRx中药物诱导的细胞状态变化模型。
scFoundation (单细胞基础模型)
一种用于将单细胞表达谱映射到高维潜在空间的深度编码器。
作为PerturbRx的细胞表达编码器基础。
ChemBERTa (药物表示模型)
基于Transformer的药物分子表示模型，利用SMILES字符串编码药物结构。
用于生成药物的条件向量。
最大均值差异 (MMD)
一种衡量两个分布差异的统计指标，用于模型分布匹配。
在源域训练中确保预测分布与真实分布一致。
响应预测 (Response Prediction)
根据患者和药物特征预测治疗效果的任务。
PerturbRx的最终目标。
Open Questions Unanswered questions from this research
- 1 如何结合临床多模态数据提升模型泛化？
- 2 多药联合扰动模型的研究尚不充分。
- 3 模型在极端病例中的表现和解释性待加强。
Applications
Immediate Applications
个性化药物反应预测
帮助医生根据患者预处理数据和药物信息，预测治疗效果，优化用药方案。
药物筛选
利用模型快速评估新药在不同细胞类型中的潜在效果，加快药物研发。
Long-term Vision
精准医疗普及
结合多模态信息，建立全方位的个性化治疗模型，实现癌症等疾病的精准治疗。
Abstract
Scarce data and tumor heterogeneity limit patient-level cancer treatment-response prediction. Existing approaches predict response from pretreatment molecular profiles and drug representations, without explicitly modeling the molecular changes expected under treatment. We propose PerturbRx, a treatment-conditioned representation learning framework that learns intervention-induced latent transitions and uses them as patient-drug response features. PerturbRx trains a drug- and dose-conditioned transition predictor from context-matched but cell-unpaired control and treated single-cell populations, then freezes and transfers the predictor to pretreatment patient profiles without requiring post-treatment measurements. The transition is combined with patient and drug representations to predict response. Across TCGA and patient-derived xenograft benchmarks, PerturbRx achieves the strongest aggregate predictive performance among the evaluated methods. These results support perturbation-pretrained latent transitions as useful representations for patient-level drug-response prediction.