本文へジャンプ

成果報告書詳細
管理番号20190000000329
タイトル*平成30年度中間年報 次世代人工知能・ロボット中核技術開発 グローバル研究開発分野 新薬開発を効率化・加速する製剤処方設計AIの開発
公開日2019/6/18
報告書年度2018 - 2018
委託先名国立大学法人京都大学
プロジェクト番号P15009
部署名ロボット・AI部
和文要約
英文要約Title: Next-generation Artificial Intelligence and Robot core technology development / Global research and development / Development of AI for pharmaceutical formulation design to streamline and accelerate drug development (FY2018-FY2019) FY2018 Annual Report

1-1. Survey of pharmaceutical formulation data
We surveyed and reviewed on-site data on drug formulation with pharmaceutical companies that participate in Life Intelligence Consortium (LINC). As a result, we decided to extract pharmaceutical formulation data from documents of the package insert, interview form, and patent. In addition, regarding of the drugs that are supplied from pharmaceutical companies, X-ray diffraction measurement and differential scanning calorimetry (melting point) measurements were performed by an outsourced experiment.

2-1. Construction of specification about algorithm for generating learning data
Data extraction from related literature: In order to extract data related to pharmaceutical formulation from medical package insert and interview form, we developed the following methods: (i) Advanced word segmentation technology, (ii) Traditional rule-base parameter extraction method, and (iii) Deep-learning based parameter extraction method.
A simulation method of a dissolution of a drug tablet: Studying a dissolution profile of a drug tablet in a solid formulation is of critical importance for the formulation of a multiple component drug. We successfully implemented a set of prototypic computational tools for simulating a dissolution process.
Chemical stability prediction of Active Pharmaceutical Ingredient (API): Chemical stability of an API is an important requirement for drug formulation. We developed a systematic and automatic method to predict Bond Dissociation Energy (BDE) values, which is an energy required for a radical formation and is associated with chemical stability, of a compound of interest.
pKa prediction of active pharmaceutical ingredient (API): pKa of an API is a critical factor for its dissolution in water, and hence, an accurate prediction of pKa is needed. We developed a systematic and automatic method to predict pKa values of a compound of interest by using quantum chemical calculations similar to the BDE prediction mentioned above.

3-1. Examination of multimodal-multitask learning model
We developed a prototype system that applied and extended a deep learning based multimodal approach. In this prototype, we implemented deep learning algorithms including Graph Convolutional Network (GCN) and Multilayer Perceptron (MLP), and designed a system that could construct a neural network for each modality. The prototypes were tested using public available data, and usefulness of the multimodal neural networks was confirmed.

4-1. Construction of specification about active learning algorithm
Since experiments and simulations are quite expensive, it is necessary to efficiently obtain distribution data (phase diagram) of physical and chemical properties in the entire parameter space. After examining various algorithms to reduce the number of experiments, we actually developed a new method using active learning and published it in the journal of Physical Review Materials.
ダウンロード成果報告書データベース(ユーザ登録必須)から、ダウンロードしてください。

▲トップに戻る