以柏雷多效率為核心概念的分子最佳化演算法

Abstract

電腦輔助的藥物設計已經被廣泛的使用在藥物開發。在預測藥物性質這個領域也發展得很成熟了，但電腦虛擬篩選依舊受限於資料庫的大小和多樣性。因此，建立一個能產生全新藥物的生成器勢在必行。在過去幾年，許多基於機器學習 (ML) 的生成器被提出，它們可以生成具有一個特定性質的藥物，像是具有特定生物活性的藥物。然而在現實應用場景，一個具有價值的藥物必須同時滿足多個特性。在這篇研究中，我們延伸前人的方法來同時優化多個不同的藥物特性。我們引入了經濟學中的柏雷多效率來為循環神經網路 (RNN) 生成器設計新的回饋函數。也設計了用來滿足不同優化設定的轉換函數，包含了最大化、最小化、特定數值範圍、特定閥值、藥物結構相似度等等。在驗證實驗中我們選了 QED、ALOGP、MW 三個常見的化學特性作為優化目標。實驗結果表明以柏雷多為基礎的回饋函數能同時最佳化多個目標，生成的分子相較於其他方法更好且更具有多樣性。總體來說，這個方法是一個高層次的抽象概念，可以在不更改架構的前提下套用到其他模型。我們相信這個演算法未來會是藥物開發中重要的一環。

Computer-aided drug design has been widely used for drug discovery. In silico predictions of drug properties are currently well developed, but using entire drug databases to perform virtual screening is limited by the capacity and diversity of the database. Thus, building a generator that can develop entirely new drugs is necessary. Over the past few years, many generators based on machine learning (ML) have been proposed to produce novel drug molecules that fit a particular property, such as a specific biological activity. However, a drug must satisfy multiple properties to be valuable in real-world applications. This study extends the previous method to optimize multiple drug properties simultaneously. We introduce Pareto optimality, commonly utilized in economics, to design a new reward function for recurrent neural network (RNN) generators. Transformation functions are also designed to meet property optimization requirements, such as maximization, minimization, specific value ranges, thresholds, and structure similarity. We selected three common chemical properties as targets for validation experiments: QED, ALOGP, and MW. The experimental results show that the Pareto-based reward function can optimize multiple targets simultaneously, and the generated molecules are better for optimization and diversity than other methods. In conclusion, this method is a high-level concept that can be easily applied to any model architecture without redesign. It is believed that this algorithm will be an essential part of drug discovery.