TL;DR
This paper introduces TAPB, a novel framework that tackles "target prior bias" in drug-target interaction (DTI) prediction. By employing causal inference and interventional training, TAPB significantly improves model generalization and provides more reliable predictions, moving beyond simply memorizing data patterns.
What Problem Does This Paper Address?
Predicting how drugs interact with biological targets (DTI prediction) is crucial for drug discovery and repurposing. Existing DTI models, particularly those using sequence data, often suffer from biases that hinder their ability to generalize to new drugs or targets. While previous work focused on "drug bias," this paper argues that "target prior bias" is the more significant issue. This bias arises from "prior tendency"—an imbalanced distribution of positive and negative interaction labels for certain targets in the training data. This imbalance causes models to learn spurious correlations rather than true biological mechanisms, leading to poor performance on unseen data.
Methods
TAPB introduces a novel interventional debiasing framework that moves beyond traditional DTI prediction. Key components include:
- Amino Acid Randomization: This technique involves randomly deleting (70%) and mutating (20%) residue features in target sequences. This disrupts spurious correlations and prevents models from simply memorizing target patterns.
- Confounder Alignment Module (CAM): TAPB identifies "target confounders" by clustering target features. The CAM then aligns these confounder representations with target features, establishing a causal path from targets to confounders.
- Interventional Training with Backdoor Adjustment: Instead of predicting P(Y|D, T) (the probability of interaction given drug D and target T), TAPB aims to compute P(Y|D, do(T)) (the probability of interaction given drug D and an intervention on target T). This is achieved through a theoretically exact backdoor adjustment, effectively isolating the causal effect of the drug on the target interaction.
Datasets
The study utilizes several publicly available datasets for training and evaluation:
- BioSNAP: A dataset for DTI prediction.
- BindingDB: Another widely used dataset for DTI prediction.
- Davis: Used for in-domain evaluation.
- Human: Used for cold-split evaluation (testing generalization to entirely new drugs and targets).
The authors also re-evaluate existing datasets and create counter-prior datasets to demonstrate the impact of "prior tendency."
Key Insights and Ideas
- "Target Prior Bias" is the Dominant Issue: The paper convincingly demonstrates that imbalanced target label distributions ("prior tendency") are a more significant source of bias in DTI prediction than previously emphasized "drug bias."
- Causal Inference for Debiasing: Applying causal inference principles to DTI prediction offers a powerful new lens for understanding and mitigating biases, moving beyond correlational approaches.
- Interventional Training for Robustness: By training models to predict the effect of an intervention on a target (do(T)), TAPB forces models to learn genuine interaction mechanisms rather than relying on spurious correlations.
- Amino Acid Randomization as a General Technique: The randomization approach is shown to be effective not only within TAPB but also when applied to existing DTI models, suggesting its broader applicability.
Key Results
- Improved Generalization: TAPB consistently outperforms existing state-of-the-art models across multiple datasets and evaluation settings, particularly in cross-domain and cold-split scenarios. For instance, on the BioSNAP in-domain split, TAPB showed significant improvements over the next best baseline (PSICHIC) in AUROC (2.3%), AUPRC (2.2%), and other metrics.
- Competitive Performance: Even on datasets where "drug bias" is more prevalent, TAPB maintains strong competitiveness, sometimes surpassing established models.
- Interpretability: TAPB provides interpretable insights at both molecular and amino acid levels, visualizing attention maps that highlight potential binding sites and interactions, which are then validated against experimental data.
Limitations and Future Work
- Unobservable Confounders: The paper acknowledges that the "target prior bias" stems from unobservable confounders. While TAPB uses a proxy (target confounder dictionary), perfectly capturing all confounding biological mechanisms remains a challenge.
- Computational Challenges: Integrating causal inference methods, especially with unobserved confounders, presents significant computational hurdles, including numerical instability.
- Generalizability of Causal Models: While TAPB's causal framework is theoretically sound, adapting it to different DTI prediction tasks and exploring optimal adjustment sets for causal inference are areas for future work.
- Multimodal Data Integration: The authors suggest that incorporating multimodal data (e.g., structural or functional annotations) could enhance the quality of proxy variables for confounder adjustment.
- Broader Applications: The paper posits that the "prior tendency" phenomenon might be prevalent in other domains, suggesting future work could explore TAPB's application in areas like drug-target affinity prediction or visual question answering.
Summary updated