Technical information-Mutual Interaction Machine Learning (CGBDD: Chemical-Genomics-based Drug Design)

TOP ＞ Technical information

Our novel method: Mutual Interaction Machine Learning (CGBDD: Chemical-Genomics-based Drug Design)

An unique method produced by Kyoto Constella Technologies predict activated compounds based on bonding patterns extracted from protein-compound interaction data (chemical genomics data) using cutting-edge pattern recognition technology.
This approach can predict without three-dimensional target protein structure data unlike conventional methods like SBDD (Docking Simulation) and LBDD (QSAR-Similarity Search). Besides that, it can produce results at extremely low calculation cost in a short period like a couple of weeks. It also has advantage of discovering new compounds with novel scaffolds, so it is useful for application like increasing variation of hit compounds at an early stage.

Features of our CGBDD and Benefits of technological inventions

Features	Benefits
1. No need of three-dimensional protein structure	It can expand an applicable range of computational technology. It is useful for R & D especially at an early stage.
2. High rate of prediction: more than 10% (Conventional methods: 1%)	High reliability for prediction results. High-activated compounds of μM-nM order can be searched for.
3. Ability to discover novel scaffolds	New applications like variation of candidate compounds or discovery of new usages can be created.
4. Low calculation cost: 2 or 3 weeks (Conventional methods: 3 months)	Even phases with a limited amount of time and money like small and medium-sized companies, ventures or researchers can really order. So, they can promote R & D and expand service coverage.

Mutual Interaction Learning Machine with Support Vector Machine (SVM)

	STEP1: Describe interaction as vector and calculate various attributes (descriptors) regarding chemical structure of each compound and amino-acid sequence of each protein. STEP2: Comprise feature vectors combining each descriptor corresponding to positive pair (bind compounds-protein pair) or negative pair (non-bind compounds-protein pair), and construct a learning model with Support Vector Machine. STEP3: Predict which classes, bind or non-bind, new vectors corresponding to unknown compound-protein pairs belong in.

STEP1: Describe interaction as vector and calculate various attributes (descriptors) regarding chemical structure of each compound and amino-acid sequence of each protein.

STEP2: Comprise feature vectors combining each descriptor corresponding to positive pair (bind compounds-protein pair) or negative pair (non-bind compounds-protein pair), and construct a learning model with Support Vector Machine.

STEP3: Predict which classes, bind or non-bind, new vectors corresponding to unknown compound-protein pairs belong in.

Searching ability for novel scaffolds (Comparison between CGBDD and a conventional method regarding prediction results of β2-AR ligand)

Each point depicts compound. Interaction prediction score by CGBDD (a nearest neighbor algorithm in a principal component space) is on a vertical axis and that by a conventional method is on a horizontal axis. A dotted line depicts the top 50 scores of each method andof upper compounds means novel scaffolds discovered by CGBDD which is experimentally-confirmed of binding.