| 髙下 大貴 | D, 中間発表 | 計算システムズ生物学 | 金谷 重彦, 宮尾 知幸, 小野 直亮, MD.ALTAF-UL-AMIN, 黄 銘 |
|
title: Design of molecular descriptor by Deep Kernel Gaussian Process and its application to QSPR models
abstract: In Quantitative Structure-Property Relationships model (QSPR model), which predicts the target properties of molecular design from the molecular structure using a machine learning model, it is important to design molecular feature vectors called descriptors, which represent the molecular structure as a numerical vector. In recent years, research has been actively conducted to automate the design of descriptors using deep learning techniques such as Graph Neural Network. However, data of interest for specific drug discovery or material design (e.g., data on biological activity or catalytic reactions) are often very small data (several hundred samples), and the design of appropriate descriptors in deep learning models is itself a difficult task due to the number of samples. Therefore, transfer learning technique is frequently used to improve the generalization performance of predictions on small data by generating in advance a large number of quantum chemical properties for which values close to the true value can be easily obtained by simulation, etc., pre-training a deep learning model using such big data, and then transferring the feature representations obtained by the model to small data. In this study, we make a hypothesis about the possible targets for transfer learning and the structure of their feature representations in situations where such molecular feature representations are to be transferred and introduce a deep learning model for obtaining ideal feature representations for transfer learning when there is a potential correlation between the target properties for transfer learning and the properties used for pre-training. We introduce a deep learning model to obtain the ideal feature representation for transfer learning when there is a potential correlation between the target properties for transfer learning and the properties for pre-training. This model is based on a model called deep kernel gaussian process, which integrates neural network and gaussian processes. We demonstrate that the feature representations extracted by the proposed model are effective in transfer learning through simulations using hypothetical numerical data. Finally, we discuss the prospects for future experiments on molecular properties. language of the presentation: Japanese 発表題目: 深層カーネルガウス過程モデルによる分子特徴表現の設計と構造物性相関モデルへの応用 発表概要: 分子の構造から機械学習モデルによって分子設計のターゲットとなる物性を予測する構造物性相関モデル(QSPRモデル)は、記述子と呼ばれる分子構造を数値ベクトルとして表現した分子特徴ベクトルの設計が重要であり、近年では、この記述子の設計をGraph Neural networkをはじめとした深層学習によって自動化する研究が積極的に行われている。しかし、特定の創薬や材料設計に興味あるデータ（生物活性のデータや触媒反応に関するデータなど）は非常にスモールデータ（数百件ほど）であることが多く、サンプル数の都合から、深層学習モデルで適切な記述子を設計すること自体が難しい課題となっている。よって、あらかじめ真の値に近い値がシミュレーション等によって簡単に得られるような量子化学的な物性を大量に生成しておき、そうして作ったビッグデータを用いて深層学習モデルを事前に学習させ、そのモデルで得られた特徴表現をスモールデータへ転移させることでスモールデータでの予測の汎化性能を向上させる転移学習がよく用いられる。本研究では、こうした分子の特徴表現を転移させるような状況において、転移学習が可能なターゲットとその特徴表現の構造について仮説を立て、転移学習の目的となる物性と事前学習に用いる物性の間に潜在的な相関関係が存在する場合において、転移学習に理想となる特徴表現を得るための深層学習モデルを紹介する。このモデルは、ニューラルネットワークとガウス過程を統合した深層カーネルガウス過程と呼ばれるモデルを応用したものである。本発表では、提案モデルによって抽出される特徴表現が転移学習において有効となることを、仮想的な数値データを用いたシミュレーションによって実証する。最後に、分子の物性を対象とした今後の実験の展望について述べていく。 | |||
| 横山 慎太朗 | D, 中間発表 | 計算システムズ生物学 | 金谷 重彦, 安本 慶一, 小野 直亮, MD.ALTAF-UL-AMIN, 黄 銘 |
|
title: Development of a smartphone application based on Constitution Theory of Traditional Chinese Medicine and Theory of Behavioral Change for the prevention and improvement of Mibyou -Aiming to build a custom-made curing system
abstract: [200~The goal is to prevent health problems associated with the aging of society, and to contribute to the reduction of medical costs, extension of healthy life expectancy, which are social issues.In order to achieve this, it is important to have a system that enables people to continue to practice "Cure Mibyou" on their own. To this end, in this study we aim to develop an application based on constitution theory of traditional Chinese medicine and theory of behavioral change. Currently a prototype of the smartphone application is under development. language of the presentation: Japanese | |||
| AHMAD KAMAL NASUTION | D, 中間発表 | 計算システムズ生物学 | 金谷 重彦, 峠 隆之, MD.ALTAF-UL-AMIN, 小野 直亮, 黄 銘 |
|
title: Natural Antibiotics Prediction based on Traditional Herbal Medicine at Plant and Molecular Level using Machine Learning Approach: Jamu and Unani Medicine
abstract: Jamu is the traditional Indonesian herbal medicine system that is considered to have many benefits, such as serving as a cure for diseases or maintaining sound health. A Jamu medicine is generally made from a mixture of several herbs. Natural antibiotics can provide a way to handle the problem of antibiotic resistance. This research aims to discover the potential of herbal plants as natural antibiotic candidates based on a machine learning approach. Our input data consists of a list of herbal formulas with plants and metabolites as constituents. The target class corresponds to bacterial diseases that herbal formulas can cure. The best model has been observed by implementing the Random Forest (RF) algorithm. For 10-fold cross-validations, the maximum accuracy, recall, and precision are 91.10%, 91.10%, and 90.54%, with standard deviations of 1.05, 1.05, and 1.48, respectively, which imply that the model obtained is excellent and robust. This study has shown that 14 plants can be potentially used as natural antibiotic candidates. Furthermore, according to scientific journals, 10 of the 14 selected plants have direct or indirect antibacterial activity. At the metabolite level, We extracted the potential compounds based on the best model as candidate antibiotics corresponding to five efficacies, e.g., digestive systems, respiratory systems, reproductive systems, skin and soft tissue, and urinary systems. Overall, we mined 111 compounds, many of which could be validated by published literature and considering structural similarities with known antibiotics. language of the presentation: English | |||