Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DTI-GP: Bayesian operations for drug-target interactions using deep kernel Gaussian processes

Bence Bolgár, András Millinghoffer, Péter Antal

TL;DR

The paper tackles uncertainty-aware DTI prediction by introducing DTI-GP, a deep kernel Gaussian process that fuses chemical fingerprints and protein embeddings to deliver probabilistic predictions. By sampling from the predictive distribution, it constructs a Bayesian precedence matrix to enable top-$K$ selection, ranking, and rejection-based enrichment, with two practical heuristics (Score and Eigen). Compared to DeepDTA and SparseChem on the KIBA benchmark with scaffold-based splits, DTI-GP achieves superior performance in early-discovery regimes and provides calibrated posterior distributions. The work demonstrates the practical value of Bayesian operations for decision-making in drug discovery and outlines avenues for active learning and ranking-focused extensions.

Abstract

Precise probabilistic information about drug-target interaction (DTI) predictions is vital for understanding limitations and boosting predictive performance. Gaussian processes (GP) offer a scalable framework to integrate state-of-the-art DTI representations and Bayesian inference, enabling novel operations, such as Bayesian classification with rejection, top-$K$ selection, and ranking. We propose a deep kernel learning-based GP architecture (DTI-GP), which incorporates a combined neural embedding module for chemical compounds and protein targets, and a GP module. The workflow continues with sampling from the predictive distribution to estimate a Bayesian precedence matrix, which is used in fast and accurate selection and ranking operations. DTI-GP outperforms state-of-the-art solutions, and it allows (1) the construction of a Bayesian accuracy-confidence enrichment score, (2) rejection schemes for improved enrichment, and (3) estimation and search for top-$K$ selections and ranking with high expected utility.

DTI-GP: Bayesian operations for drug-target interactions using deep kernel Gaussian processes

TL;DR

The paper tackles uncertainty-aware DTI prediction by introducing DTI-GP, a deep kernel Gaussian process that fuses chemical fingerprints and protein embeddings to deliver probabilistic predictions. By sampling from the predictive distribution, it constructs a Bayesian precedence matrix to enable top- selection, ranking, and rejection-based enrichment, with two practical heuristics (Score and Eigen). Compared to DeepDTA and SparseChem on the KIBA benchmark with scaffold-based splits, DTI-GP achieves superior performance in early-discovery regimes and provides calibrated posterior distributions. The work demonstrates the practical value of Bayesian operations for decision-making in drug discovery and outlines avenues for active learning and ranking-focused extensions.

Abstract

Precise probabilistic information about drug-target interaction (DTI) predictions is vital for understanding limitations and boosting predictive performance. Gaussian processes (GP) offer a scalable framework to integrate state-of-the-art DTI representations and Bayesian inference, enabling novel operations, such as Bayesian classification with rejection, top- selection, and ranking. We propose a deep kernel learning-based GP architecture (DTI-GP), which incorporates a combined neural embedding module for chemical compounds and protein targets, and a GP module. The workflow continues with sampling from the predictive distribution to estimate a Bayesian precedence matrix, which is used in fast and accurate selection and ranking operations. DTI-GP outperforms state-of-the-art solutions, and it allows (1) the construction of a Bayesian accuracy-confidence enrichment score, (2) rejection schemes for improved enrichment, and (3) estimation and search for top- selections and ranking with high expected utility.
Paper Structure (13 sections, 8 equations, 8 figures, 2 tables)

This paper contains 13 sections, 8 equations, 8 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Materials and methods
  3. Method overview
  4. Data set and preparation
  5. Proposed model
  6. Input representations
  7. Learning the covariance function
  8. Bayesian prediction through Gaussian processes
  9. Bayesian top-$K$ selection with sampling
  10. Baseline methods
  11. Experiments and results
  12. Discussion
  13. Conclusion

Figures (8)

  • Figure 1: DTI-GP model with two neural networks learning molecule and sequence representations and a Gaussian process classification model computing probabilistic outputs.
  • Figure 2: ROC and PR curves in the global evaluation setting. The GP-based model outperforms DeepDTA, especially in the region relevant for early discovery and top-$K$ selection.
  • Figure 3: Reliability plot of the proposed model. Note that both the Bayesian and MAP estimations suffer from overconfidence in the relevant high-probability region.
  • Figure 4: Difference between the score and eigenvector-based heuristics. The eigenvector-based method assigns values close to $0$ to most interactions; the score-based method distributes values more evenly.
  • Figure 5: Relationship between the set size $K$ and false discovery rates. The score and eigenvector-based heuristics outperform DeepDTA, SparseChem and DTI-GP-MAP in the small set size region. The right side of the figure is somewhat noisy, as expected with such small set sizes.
  • ...and 3 more figures