ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening
ScreenShot employs a hierarchical transformer pretrained on 40 drug screening datasets, enabling few-shot prediction and active learning to significantly improve combination drug screening efficiency.
Key Findings
Methodology
The proposed ScreenShot model utilizes a hierarchical transformer architecture pretrained on a large collection of 40 drug screening datasets, encompassing approximately 3700 drugs and 6000 biological samples. The architecture reflects the natural nested structure of screening data, comprising three levels: drug-dose combination encoding (Level 1), response encoding (Level 2), and sample-level aggregation (Level 3). During pretraining, the model employs self-supervised learning on drug-dose-viability triplets, leveraging cross-attention mechanisms to fuse information across levels. At inference, the model performs in-context learning by conditioning on a small set of observed responses from a new sample, predicting responses for untested drug-dose combinations without fine-tuning or molecular features. Additionally, the model integrates an active learning strategy based on embedding space clustering (k-means++) to select experiments efficiently, reducing the experimental budget while maintaining high hit detection performance.
Key Results
- On four held-out datasets, including monotherapy and combination screens, ScreenShot outperforms baseline models such as XGBoost, TabPFN, and fine-tuned MLP in both correlation metrics (Pearson r > 0.8 with as few as 50 observations) and top-hit recall (exceeding 90% in top 10% predictions). The model demonstrates robust few-shot generalization, especially in low-data regimes, with significant improvements over existing methods.
- Active learning experiments show that using ScreenShot’s embedding-driven sampling strategy achieves comparable hit detection as uniform screening but with only one-third of the experimental budget, validating the efficiency of the approach.
- The internal representations learned by ScreenShot are not only predictive but also interpretable enough to guide experimental design, enabling targeted exploration of promising drug combinations and doses, thus accelerating the drug discovery pipeline.
Significance
This work advances the field of drug response prediction by removing the dependency on molecular profiling, which is often unavailable or inconsistent in clinical settings. The ability to perform accurate few-shot predictions directly from functional measurements addresses a critical bottleneck in personalized medicine, especially for rare or limited samples. The integration of a pretrained foundation model with active learning strategies offers a scalable, cost-effective solution for high-throughput drug screening, with profound implications for accelerating drug development and optimizing personalized treatment regimens. By bridging the gap between large-scale pretraining and practical experimental constraints, this approach paves the way for more adaptive and intelligent screening platforms.
Technical Contribution
The core technical innovation lies in designing a hierarchical transformer architecture that mirrors the nested structure of drug screening data, enabling efficient multi-level attention. The model is pretrained using an in-context learning paradigm, akin to large language models, allowing it to adapt to new samples without fine-tuning. The embedding of drugs, doses, and responses facilitates flexible, modality-agnostic representations. The active learning component employs a weighted k-means++ seeding strategy in the embedding space, explicitly balancing exploration and exploitation. This combination of hierarchical modeling, in-context inference, and embedding-driven experiment selection constitutes a significant advancement over existing drug response models, which often rely on molecular features or require retraining for each new sample.
Novelty
This study is the first to apply a hierarchical transformer framework for few-shot drug response prediction directly from functional measurements, bypassing the need for molecular features. Its in-context learning approach, combined with a pretrained foundation model trained on a large, diverse dataset, distinguishes it from prior transfer learning or fine-tuning-based methods. The embedding-based active learning strategy further enhances experimental efficiency, setting a new standard for cost-effective drug screening. Unlike previous models limited to monotherapy or requiring extensive molecular data, ScreenShot achieves broad applicability across drug combinations and biological contexts, representing a major step forward in AI-driven pharmacology.
Limitations
- The model’s performance diminishes when the number of observations drops below 50, indicating challenges in extremely sparse data scenarios. Its generalization to non-cancerous or non-tumor tissues remains unverified, as the pretraining data predominantly covers cancer cell lines.
- While the model effectively predicts responses based on functional measurements, it does not explicitly incorporate mechanistic or molecular information, which could limit interpretability and biological insight.
- Computational costs associated with pretraining and inference, especially for very large datasets or real-time applications, may pose practical constraints. Further optimization and model compression are needed for deployment in resource-limited settings.
Future Work
Future directions include integrating multi-omics data such as gene expression and mutation profiles to enhance interpretability and predictive accuracy. Developing explainability methods for the internal representations could facilitate biological insights and clinical trust. Extending the framework to incorporate causal inference and reinforcement learning could further optimize experimental strategies. Additionally, applying the model to clinical trial design and real-world patient data will be crucial for translating this technology into clinical practice. Finally, efforts to reduce computational overhead and improve robustness across diverse biological contexts will be essential for broader adoption.
AI Executive Summary
The quest for effective combination therapies in cancer treatment is fraught with challenges. Traditional drug screening methods, while invaluable, are hampered by high costs, lengthy timelines, and technical limitations, especially when dealing with limited patient samples. As the complexity of tumor biology and the diversity of potential drug combinations grow, the need for smarter, more efficient screening strategies becomes urgent.
Enter ScreenShot, a groundbreaking AI framework that leverages the power of hierarchical transformers pretrained on a vast collection of drug screening data. Unlike conventional models that depend heavily on molecular features or require retraining for each new sample, ScreenShot operates directly on functional measurements—drug-dose-viability data—making it highly adaptable and practical for clinical settings with scarce tissue samples.
The core innovation of ScreenShot lies in its architecture, which mimics the nested structure of screening data. It encodes drug-剂量组合、反应和样本信息在三个层级，通过交叉注意力机制实现信息融合。这种设计不仅符合数据的自然生成过程，还大大降低了计算复杂度。预训练过程中，模型通过自监督学习在大规模药物筛选数据上学习到丰富的药物和样本表征，支持在新样本上进行即插即用的少样本预测。
在实际应用中，ScreenShot展现出优异的性能。在四个未见过的数据集上，它在少样本条件下的预测相关性（Pearson r）超过0.8，命中率达到90%以上，显著优于XGBoost、TabPFN和微调的MLP。更重要的是，结合其嵌入空间的主动学习策略，模型能在只用三分之一预算的情况下，达到与全量筛查相当的筛查效果。这一创新极大地提高了药物筛查的效率，降低了成本。
此外，模型的内部表示不仅支持预测，还能指导实验设计。通过加权k-means++策略，模型在探索潜在有效药物组合时，兼顾了多样性和预测效果，极大地提升了筛查的智能化水平。未来，结合多模态数据和因果推断，ScreenShot有望在临床药物筛查、个性化治疗方案制定中发挥更大作用。
尽管如此，模型在极端少样本和跨疾病应用方面仍存在挑战。未来的研究将致力于增强模型的泛化能力、解释性和临床适用性，推动AI在精准医疗中的深度融合。
Deep Dive
Abstract
Treating patients with combinations of drugs reduces the risk of resistance to any individual drug. Finding effective combinations is difficult because the large search space makes combinatorial screens prohibitively expensive, time consuming, and often technically infeasible. Predictive models can fill this gap, yet existing methods typically require molecular profiling of each sample and per-cohort training, limiting their applicability when time and tissue are scarce. To address this challenge, we introduce ScreenShot, a hierarchical transformer pretrained on 40 drug screening datasets covering 3,700 drugs and 6,000 biological samples, whose architecture mirrors the nested structure of screening data. Given a few-shot context of observations from a new patient, ScreenShot predicts the response of the sample to combination therapies through in-context learning, operating directly on functional measurements with no fine-tuning and no molecular profiling. On four held-out datasets, ScreenShot outperforms all baselines in both prediction accuracy and identification of selectively effective treatments. ScreenShot's internal representations are directly useful for experimental design: we use them to drive a weighted k-means++ active learning strategy that selects which experiments to run, achieving the same hit detection as uniform screening with a third of the budget. Source code and interactive dashboard: https://github.com/tansey-lab/screenshot.
References (20)
Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning
Y. Gal, Zoubin Ghahramani
Accurate predictions on small data with a tabular foundation model
N. Hollmann, Samuel G. Müller, Lennart Purucker et al.
A Pan-Cancer Ex Vivo Drug Screen Atlas for Functional Precision Oncology
K. Pichotta, Jessica White, Jeffrey Quinn et al.
Clonal somatic copy number altered driver events inform drug sensitivity in high-grade serous ovarian cancer
F. C. Martins, Dominique-Laurent Couturier, Ines de Santiago et al.
Drug responses are conserved across patient-derived xenograft models of melanoma leading to identification of novel drug combination therapies
R. Ice, Michelle Chen, M. Sidorov et al.
An Unbiased Oncology Compound Screen to Identify Novel Combination Strategies
J. O'neil, Yair Benita, I. Feldman et al.
Precision medicine treatment in acute myeloid leukemia using prospective genomic profiling: feasibility and preliminary efficacy of the Beat AML Master Trial
A. Burd, R. Levine, A. Ruppert et al.
Organoid screening reveals epigenetic vulnerabilities in human colorectal cancer
Kohta Toshimitsu, Ai Takano, Masayuki Fujii et al.
Predicting Drug Response and Synergy Using a Deep Learning Model of Human Cancer Cells
B. Kuenzi, Jisoo Park, Samson H. Fong et al.
Tumor evolution and drug response in patient-derived organoid models of bladder cancer
Suk Hyung Lee, Wenhuo Hu, J. Matulay et al.
Language Models are Few-Shot Learners
Tom B. Brown, Benjamin Mann, Nick Ryder et al.
The National Cancer Institute ALMANAC: A Comprehensive Screening Resource for the Detection of Anticancer Drug Pairs with Enhanced Therapeutic Activity
S. Holbeck, R. Camalier, J. Crowell et al.
Identification of Novel Therapeutic Targets for Fibrolamellar Carcinoma Using Patient Derived Xenografts and Direct from Patient Screening
G. Lalazar, David Requena, Lavoisier S. Ramos-Espiritu et al.
High-Throughput Drug Screening of Primary Tumor Cells Identifies Therapeutic Strategies for Treating Children with High-Risk Cancer
C. Mayoh, Jie Mao, Jinhan Xie et al.
The landscape of drug sensitivity and resistance in sarcoma
Ahmad Al Shihabi, Peyton J. Tebon, H. Nguyen et al.
k-means++: the advantages of careful seeding
David Arthur, Sergei Vassilvitskii
Integrative Analysis of Drug Response and Clinical Outcome in Acute Myeloid Leukemia
D. Bottomly, Nicola Long, Anna Reister Schultz et al.
Scikit-learn: Machine Learning in Python
Fabian Pedregosa, G. Varoquaux, Alexandre Gramfort et al.
Effective drug combinations in breast, colon and pancreatic cancer cells
P. Jaaks, E. Coker, D. Vis et al.
Small Cell Lung Cancer Screen of Oncology Drugs, Investigational Agents, and Gene and microRNA Expression.
E. Polley, M. Kunkel, David Evans et al.