Research map / Probabilistic and scientific modeling / Molecules and proteins

ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Protein interface and docking prediction · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces ScreenShot, a hierarchical transformer pretrained on 40 drug screening datasets that predicts combination drug responses for new samples via in-context learning from functional measurements only, plus an efficacy-weighted k-means++ active learning policy.

Problem

Predicting responses to drug combinations in new patient samples from few functional measurements, without molecular profiling or per-sample retraining, and choosing which experiments to run under tight sample budgets.

Approach

A three-level hierarchical transformer (drug-dose combination, response, sample) that conditions on a context of (drug, dose, viability) observations through cross-attention in a TabPFN-style in-context paradigm, paired with delta-weighted k-means++ selection in its learned embedding space to find selective hits.

Key results

• ScreenShot achieves the highest Pearson correlation on every dataset and budget and the highest hit recall on 16 of 20 dataset/budget combinations (e.g., 63.9% vs 45.9% for MLP-FT at n=10 on BATCHIE).
• Adaptive delta-weighted k-means++ selection reaches ~85% hit recall on BATCHIE at budget 300 vs ~76% for random, and needs 2-3x less budget than random to match its recall.
• ScreenShot is 5-50x faster than baselines, completing inference in 8-68 seconds per experiment on A100 GPUs.

Surprising findings

• Uncertainty-based (MC dropout) selection gives best MAE and Pearson but 6 points lower hit recall than delta-weighted selection on BATCHIE, indicating global accuracy and hit detection are not directly linked.

Limitations

• Ex vivo pretraining data obtained on request from a third party, and drug coverage for novel compounds relies on UNK tokens (4-11% of drugs uncovered).
• Control samples are not designated in most datasets; a proxy control chosen by highest average viability is used.
• Level 3 self-attention can be dropped with modest hit-recall loss, suggesting its contribution is limited.

Experimental setup

Datasets: Beat AML Nair et al. CCLE CTRP1 CTRP2 GDSC1 GDSC2 Merck Malani et al. Polley et al. NCI-ALMANAC GDSC-SQ

Benchmarks: NCI-ALMANAC GDSC-SQ BATCHIE PDO-Breast

Metrics: Pearson correlation top-hit recall wall-clock time

Baselines: XGBoost TabPFN MLP (target-only) MLP (fine-tuned)

Builds on: TabPFN XGBoost k-medoids MC dropout

Closest papers by meaning

• Enhancing Treatment Effect Estimation via Active Learning: A Counterfactual Covering Perspective (ICML 2025 · similarity 0.84)
• Blend the Separated: Mixture of Synergistic Experts for Data-Scarcity Drug-Target Interaction Prediction (AAAI 2025 · similarity 0.84)
• Drug-TTA: Test-Time Adaptation for Drug Virtual Screening via Multi-task Meta-Auxiliary Learning (ICML 2025 · similarity 0.84)
• Learning Identifiable Factorized Causal Representations of Cellular Responses (NeurIPS 2024 · similarity 0.84)
• WISER: Weak Supervision and Supervised Representation Learning to Improve Drug Response Prediction in Cancer (ICML 2024 · similarity 0.83)
• Efficient Biological Data Acquisition through Inference Set Design (ICLR 2025 · similarity 0.83)