Research map / Label-efficient and robust learning / Test-time adaptation

Drug-TTA: Test-Time Adaptation for Drug Virtual Screening via Multi-task Meta-Auxiliary Learning

ICML 2025 · Label-efficient and robust learning › Test-time adaptation · approach: Pseudo-label test-time adaptation

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.

Drug-TTA applies per-instance test-time adaptation with self-supervised auxiliary tasks, meta-learning, and an adaptive loss-balance module on top of DrugCLIP for zero-shot drug virtual screening.

Problem

Contrastive feature-alignment virtual screening models trained with overly simple in-batch negatives generalize poorly to hard negatives at test time on unseen pockets and molecules.

Approach

Adapt the pocket and molecule encoders to each test instance by updating normalization layers with five self-supervised auxiliary losses during both training and inference via meta-learning, with a multi-scale feature-based module (ALBM) that predicts per-instance softmax weights for the auxiliary losses.

Key results

• On DUD-E, Drug-TTA reaches 93.16% AUROC and 82.82% BEDROC versus 80.93% and 50.52% for DrugCLIP (zero-shot).
• On LIT-PCBA, Drug-TTA reaches 71.24% AUROC and 45.08% BEDROC versus DrugCLIP's 57.17% and 6.23%.
• Ablation shows removing meta-learning drops DUD-E AUROC to 80.19%, below DrugCLIP, and using only coarse-grained or only fine-grained tasks gives smaller gains.

Measured effects reported in the paper

Surprising findings

• Applying TTA only to the pocket encoder yields worse performance than the DrugCLIP baseline, so both encoders must be adapted together.
• Without meta-learning, the TTA variant performs below DrugCLIP, indicating the auxiliary tasks can harm the primary task.

Limitations

• Per-instance adaptation requires updating encoders at inference for every test pocket and candidate set, adding computational cost.
• Ablations and gains are reported on specific benchmarks; generalization beyond the five benchmarks is not shown.

Experimental setup

Datasets: DrugCLIP training dataset DUD-E LIT-PCBA AD DEKOIS 2.0 CASF-2016

Benchmarks: DUD-E LIT-PCBA AD DEKOIS 2.0 CASF-2016

Metrics: AUROC BEDROC EF0.5% EF1% EF5%

Baselines: Glide-SP Vina NN-score RFscore Pafnucy OnionNet Planet DrugCLIP Surflex Gnina DeepDTA BigBind

Builds on: DrugCLIP Uni-Mol Test-time training (TT++) SimCLR meta-auxiliary learning

Closest papers by meaning

• AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation (NeurIPS 2025 · similarity 0.86)
• Rethinking the generalization of drug target affinity prediction algorithms via similarity aware evaluation (ICLR 2025 · similarity 0.86)
• PharmacoMatch: Efficient 3D Pharmacophore Screening via Neural Subgraph Matching (ICLR 2025 · similarity 0.85)
• DrugHash: Hashing Based Contrastive Learning for Virtual Screening (AAAI 2025 · similarity 0.85)
• UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery (ICLR 2025 · similarity 0.84)
• CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening (NeurIPS 2026 · similarity 0.84)