Efficiently and accurately predict DDI, the explanatory drug AI model of Fuzhou University and Yuanxing Intelligent Drug Team was published in Nature sub-journal

Editor | Radish Peel

Unexpected drug interactions (DDIs) are an important issue in drug research and clinical application because they are highly likely to cause serious adverse drug effects. reaction or drug withdrawal.

While many deep learning models have achieved good results in DDI prediction, model interpretability to reveal the underlying causes of DDI has not been widely explored.

Researchers from Fuzhou University, the First Affiliated Hospital of Fujian Medical University and Yuanxing Intelligent Medicine proposed MeTDDI - a deep learning framework with local-global self-attention and joint attention for learning based on DDI prediction plot of subject.

Regarding interpretability, researchers conducted an extensive evaluation on 73 drugs (13,786 DDIs), and MeTDDI can accurately explain the structural mechanisms of 5,602 DDIs involving 58 drugs. Furthermore, MeTDDI shows potential to explain complex DDI mechanisms and reduce DDI risk.

MeTDDI provides a new perspective for exploring DDI mechanisms, which will facilitate drug discovery and polypharmacy, thereby providing safer treatments for patients.

The study was titled "Learning motif-based graphs for drug–drug interaction prediction via local–global self-attention" and was published in "Nature Machine Intelligence" on August 27, 2024.

Due to aging and multimorbidity, drug combinations or polypharmacy are widely used and may have consequences for public health and the economy. Despite the therapeutic benefits of polypharmacy, there is a risk of unintended drug-drug interactions (DDIs), which may lead to serious adverse drug reactions (ADRs) or even discontinuation.
Thus, predicting DDIs in advance will bring huge benefits to drug research and clinical settings, thereby improving drug safety and protecting patient health. DDI assessment through in vitro and in vivo experiments is useful but is costly, time-consuming, and laborious, hindering the practicality of large-scale DDI screening.
Today, deep learning models have emerged as a promising alternative for high-throughput accurate DDI prediction as well as root cause explanation.
In the latest study, the research team of Fuzhou University, the First Affiliated Hospital of Fujian Medical University, and Yuanxing Intelligent Medicine focused on the prediction of metabolism-mediated drug interactions (MMDDI) and proposed a deep molecular structure-based Learning framework MeTDDI for predicting MMDDI.
This method is mainly used to solve three challenges in DDI prediction: (1) learning intra- and intermolecular substructural interactions, (2) predicting DDI-related drug metabolism, (3) widely providing and evaluating the model Interpretability.

Illustration: MeTDDI architecture overview. (Source: paper)

Benefiting from local-global self-attention and joint attention structures, MeTDDI can effectively learn intra- and intermolecular substructure interactions within/between graphs based on motifs , thereby performing DDI reasoning.

Evaluation results show that it achieves competitive performance compared to baselines in both classification and regression tasks. MeTDDI can also accurately identify the mechanistic role of a drug (perpetrator or victim) in DDI and quantify the impact of the perpetrator on the victim PK, which is very beneficial for both drug research and clinical applications.

Illustration: Performance comparison of models in predicting AUC FC values. (Source: paper)

Regarding model interpretability, MeTDDI demonstrates the ability to identify key mechanistic substructures relevant to DDI.

First, the key substructures visualized by MeTDDI roughly match those reported in the literature from the analysis of 73 representative compounds (with 13,786 DDI pairs).

Illustration: Interpretability analysis of MeTDDI is used to explain the DDI mechanism. (Source: paper)

Second, the researchers evaluated the model interpretability of MeTDDI and two state-of-the-art models, namely CIGIN and CGIB. The results show that MeTDDI also exhibits excellent performance in terms of model interpretability.

Additionally, MeTDDI can highlight metabolic sites of chemicals associated with enzyme inhibition.

Advantages of MeTDDI

Traditional methods only explain the mechanism of DDI by testing the metabolic enzyme inhibition of the perpetrator in vitro, without fully considering the victim. This is problematic because the potency of enzyme inhibition by the perpetrator can vary depending on the chemical identity of the victim.

被害者は、代謝酵素 (特に CYP) と加害者の結合または相互作用パターンを変更し、その結果、さまざまな酵素阻害メカニズムが生じる可能性があります。これは、インビトロで単独で使用すると代謝酵素の強力な阻害剤であるエチニルエストラジオールやゲストデンなどの一部の化学物質が、それらの犠牲者と組み合わせると効果が低くなる理由を説明する可能性があります。これは、エチニルエストラジオールを用いた研究でなぜ 2 つの反応しか観察されなかったのかを説明する可能性があり、これが in vitro で CYP3A4 を不活化するメカニズムであると考えられています。

さらに、パロキセチンとイトラコナゾールのケーススタディでは、MeTDDI が化学物質のモチーフの変化を正確に予測し、生物学的実験の結果と一致していることが示されており、研究者が薬物の構造を変更して MMDDI のリスクを軽減するのに役立つ可能性が実証されています。

要約すると、MeTDDI は DDI 予測機能を強化し、DDI メカニズムを理解して探索するための新しい視点を提供します。これにより、医薬品開発とポリファーマシーが促進され、それによって患者により安全な治療が提供されます。

図: MeTDDI を使用した DDI 軽減の 2 つのケーススタディ。 （出典：論文） MeTDDI の改善方向
MeTDDI には多くの利点がありますが、同時にいくつかの制限もあります。
まず、困難なシナリオでは正確な予測が困難です。これは、DDI メカニズムの多様性と複雑さ、および薬物構造のみに依存することの限界に起因している可能性があります。
MMDDI では両方の薬物が同じ代謝酵素上で相互作用する必要があるため、酵素の特徴をモデルに組み込んで学習を改善できます。ただし、一部の代謝酵素 (CYP など) は薬物と酵素の相互作用部位に非常に高い柔軟性を示すため、酵素特性のモデル化は依然として課題です。
第二に、MeTDDI でトレーニングされたデータセットは FDA の医薬品ラベルに基づいています。これは集団の統計的観察であり、個々の患者の特徴を反映していない可能性があります。したがって、モデルを開発し、将来的により正確な予測を行うために、利用可能な場合は個々の患者データを考慮する必要があります。第三に、MeTDDI は 3 つ以上の薬物の相互作用を同時に予測することが難しい可能性があります。
ただし、ポリファーマシーを確保するための一般的な方法は、考えられるすべての薬物ペア間のペアごとの DDI を検索することです。MeTDDI を直接展開して、すべての薬物ペアを列挙することで複数の薬物間の DDI を予測できます。
最後に、DDI の基礎となる新たに発見された部分構造については、分子ドッキングなどの代替技術を補完的なアプローチとして採用して、MeTDDI 可視化機能の信頼性を高めることができます。そして研究者らは、分子ドッキングはMeTDDIを補完する貴重なツールであると述べている。
論文リンク: https://www.nature.com/articles/s42256-024-00888-6

The above is the detailed content of Efficiently and accurately predict DDI, the explanatory drug AI model of Fuzhou University and Yuanxing Intelligent Drug Team was published in Nature sub-journal. For more information, please follow other related articles on the PHP Chinese website!