Conformal Prediction for Uncertainty Estimation in Drug-Target Interaction Prediction

Abstract: Accurate drug-target interaction (DTI) prediction with machine learning models is essential for drug discovery. Such models should also provide a credible representation of their uncertainty, but applying classical marginal conformal prediction (CP) in DTI prediction often overlooks variability across drug and protein subgroups. In this work, we analyze three cluster-conditioned CP methods for DTI prediction, and compare them with marginal and group-conditioned CP. Clusterings are obtained via nonconformity scores, feature similarity, and nearest neighbors, respectively. Experiments on the KIBA dataset using four data-splitting strategies show that nonconformity-based clustering yields the tightest intervals and most reliable subgroup coverage, especially in random and fully unseen drug-protein splits. Group-conditioned CP works well when one entity is familiar, but residual-driven clustering provides robust uncertainty estimates even in sparse or novel scenarios. These results highlight the potential of cluster-based CP for improving DTI prediction under uncertainty.

Conformal Prediction for Uncertainty Estimation in Drug-Target Interaction Prediction

Accurate predictions of drug-target interactions (DTI) are of significant importance in the field of drug discovery. Traditional approaches often focus on point estimates of interaction strengths, neglecting the uncertainty which can impede experimental validations. This paper introduces advanced methodologies for incorporating uncertainty estimates into DTI predictions using Conformal Prediction (CP). The authors analyze three cluster-conditioned CP methods, contrasting these with marginal and group-conditioned CP approaches.

In their experiments using the KIBA dataset, the authors demonstrated that cluster-conditioned CP methods, particularly those based on nonconformity scores (CCP-NC), are highly effective in narrowing prediction intervals and providing reliable subgroup coverage. This is particularly evident in scenarios where drug-protein splits are random and incorporate entirely new data pairs. Notably, nonconformity-based clustering outperformed others by accurately reflecting uncertainty and producing the tightest prediction intervals under conditions of random and previously unseen splits.

Group-conditioned CP methods, which work well when data includes familiar entities, were proficient at generating promising uncertainty estimates but struggled in novel data scenarios. Here, residual-driven clustering strategies proved robust, ensuring valid coverage despite sparse samples. This highlights the efficacy of cluster-based approaches in adapting to various contexts of heterogeneity, especially their strength in improving DTI prediction by refining uncertainty estimates at the subgroup level.

From a broader perspective, this research provides detailed annotations on the conditions under which cluster-based CP methods excel. Additionally, it bridges theoretical constructs with practical implementations, illustrating that cluster-conditioned prediction frameworks can significantly enhance machine learning models in drug discovery. The implications extend to computational biology and other domains that hinge on interaction predictions, aligning predictive confidence with real-world complexities such as unseen or sparsely represented entities.

Future research in artificial intelligence may continue to explore the integration of structured clustering approaches within uncertainty quantification frameworks, further optimizing interaction predictions under the increasing complexity of biological data. Enhancements in model scalability and efficiency are anticipated as new clustering algorithms and feature representations emerge, offering refined uncertainty estimates tailored to specific domains.