MuCoS: Efficient Drug-Target Prediction through Multi-Context-Aware Sampling

TL;DR

This work tackles the challenge of predicting drug-target interactions in biomedical knowledge graphs where unseen relations and the overhead of negative sampling hinder traditional KG embedding methods. It introduces MuCoS, a Multi-Context-Aware Sampling approach that uses density-based neighbor selection to form context-rich inputs, which are then processed by BERT to predict relations and tails with a negative-sample-free training regime. Compared to CAB-KGC and other baselines on the KEGG50k dataset, MuCoS achieves higher MRR and Hits@k for drug-target prediction while offering roughly a 10x reduction in computational cost. The approach enables efficient, scalable DTI prediction by leveraging contextualized, structurally informed representations, with potential gains in speeding up drug discovery workflows.

Abstract

Drug-target interactions are critical for understanding biological processes and advancing drug discovery. However, traditional methods such as ComplEx-SE, TransE, and DistMult struggle with unseen relationships and negative triplets, which limits their effectiveness in drug-target prediction. To address these challenges, we propose Multi-Context-Aware Sampling (MuCoS), an efficient and positively accurate method for drug-target prediction. MuCoS reduces computational complexity by prioritizing neighbors of higher density to capture informative structural patterns. These optimized neighborhood representations are integrated with BERT, enabling contextualized embeddings for accurate prediction of missing relationships or tail entities. MuCoS avoids the need for negative triplet sampling, reducing computation while improving performance over unseen entities and relations. Experiments on the KEGG50k biomedical dataset show that MuCoS improved over existing models by 13\% on MRR, 7\% on Hits@1, 4\% on Hits@3, and 18\% on Hits@10 for the general relationship, and by 6\% on MRR, 1\% on Hits@1, 3\% on Hits@3, and 12\% on Hits@10 for prediction of drug-target relationship.

Figures (4)

• Figure 1: A concise view of the MuCoS model pipeline for predicting the relation (given a head $h$ and a tail $t$ entity) and tail entity (given a head entity $h$ and a relationship $r$). The box on the left illustrates the input sequence to the BERT model, containing a union of $\mathcal{R}(h)$ (set of all relations) and $\mathcal{E}(h)$ (set of all neighboring entities). The BERT model, combined with a linear classifier and softmax, generates probabilities for tail entities.
• Figure 2: MuCoS $\mathcal{H}_c$ construction. The left box illustrates the extraction of $\mathcal{H}_c$, which consists of the set of relationships $\mathcal{R}(h)$ ($r_1, r_2, r_3, r_4, r_5, r_6$) and the set of neighboring entities $\mathcal{E}(h)$ ($e_1, e_2, e_3, e_4, e_5, e_6$) associated with the head entity $h$. The middle box shows optimization, where each entity $e$ density $d(e)$ is computed based on its frequency in triples. The optimized head context $H_c*$ includes the top-$k$ most influential entities and relationships. Lastly, the optimized $\mathcal{H}_c^*$ is then aggregated into a unified representation $\mathcal{H}_{\text{agg}}$ using concatenation, as shown in the right box.
• Figure 3: $\mathcal{R}_c$ construction, the left section shows the appearance of $r_1$ and lists the entities $\mathcal{E}(r)$ that are connected to the relationship $r_1$. The middle section is used for optimization; we suppose that the current top $k$ for $r$ selection is two, so retaining both relevant entity pairs $(e_2, e_3)$ and $(e_6, e_7)$. The optimized relationship context $\mathcal{R}_c^*$ is aggregated $\mathcal{R}_{\text{agg}}$ using concatenation, as depicted in the right box.
• Figure 4: Similarly, the left box represents the extraction of $\mathcal{T}_c$ the set of relationships $\mathcal{R}(t)$ ($r_1, r_2, r_3, r_4, r_5, r_6$) and neighboring entities $\mathcal{E}(t)$ ($e_1, e_2, e_3, e_4, e_5, e_6$) associated with the tail entity $t$. The middle box illustrates the optimization step, where the density $d(e)$ of each entity $e$ is calculated based on its frequency in triples. The top-$k$ most influential entities and relationships are selected to form the optimized tail context $T_c^*$. The concatenation results in a unified representation $T_{\text{agg}}$.