Learning to Describe for Predicting Zero-shot Drug-Drug Interactions

TL;DR

Leveraging textual information from online databases like DrugBank and PubChem, an innovative approach TextDDI with a language model-based DDI predictor and a reinforcement learning~(RL)-based information selector is proposed, enabling the selection of concise and pertinent text for accurate DDI prediction on new drugs.

Abstract

Adverse drug-drug interactions~(DDIs) can compromise the effectiveness of concurrent drug administration, posing a significant challenge in healthcare. As the development of new drugs continues, the potential for unknown adverse effects resulting from DDIs becomes a growing concern. Traditional computational methods for DDI prediction may fail to capture interactions for new drugs due to the lack of knowledge. In this paper, we introduce a new problem setup as zero-shot DDI prediction that deals with the case of new drugs. Leveraging textual information from online databases like DrugBank and PubChem, we propose an innovative approach TextDDI with a language model-based DDI predictor and a reinforcement learning~(RL)-based information selector, enabling the selection of concise and pertinent text for accurate DDI prediction on new drugs. Empirical results show the benefits of the proposed approach on several settings including zero-shot and few-shot DDI prediction, and the selected texts are semantically relevant. Our code and data are available at \url{https://github.com/zhufq00/DDIs-Prediction}.

Figures (5)

• Figure 1: A graphical illustration of zero-shot DDI prediction based on the textual descriptions. The green box in the upper left contains known interactions between a set of known drugs. The blue box in the upper right includes several new drugs that are not seen during training. The task is to use drug descriptions from websites to predict interactions for the new drugs.
• Figure 2: The inference procedure of the proposed TextDDI. (a). Given a drug pair, we obtain the drug descriptions from websites. (b). The previous prompt $\bm p_{t-1}[u,v]$ (where $\bm p_0[u,v]$ is "[CLS] $u^{name}:$ {} $v^{name}:$ {} [Task Instruction] [SEP]") is fed into the selector. (c). The information selector samples one sentence from descriptions based on their priority score. (d). The prompt $\bm p_t[u,v]$ is constructed with $\bm p_{t-1}[u,v]$ and the new sentence $u_j$. (e). When the prompt $\bm p_t[u,v]$ exceeds the length limit $L$, we stop the loop and use $\bm p_{t-1}[u,v]$ as the final prompt $\bm p[u,v]$ to predict the interaction type of drug pair $(u,v)$.
• Figure 3: Comparison with baselines under the different number of shots of DDIs seen during training.
• Figure 4: The impact of maximum prompt length $L$ on model performance, and computational costs.
• Figure 5: Distribution of prompt length for complete drug-pair prompts constructed with raw drug descriptions.

Theorems & Definitions (1)

• Definition 1: Zero-shot DDI prediction