RIKEN Program for Drug Discovery and Medical Technology Platforms

Drug Discovery Computational Chemistry Platform Unit
High precision in silico screening technology

 Structure Based Drug Design (SBDD) such as in silico screening is an effective approach on exploring the new outstanding seeds of drug discovery. In the case of difficult targets, the silico screening sometimes ended in failure because in the conventional docking methods, flexibility of protein were not taken into consideration. This Unit has been developing high precision in silico screening systems making full advantage of both protein structure and known ligand information(Knowledge-Based Docking Screening, KBDS, Fig.1).


Fig.1 Workflow of Knowledge-Based Docking Screening

 KBDS consists of two systems k-PALLAS and k-MUSES. The k-PALLAS is semiautomated system for docking condition optimization, can select optimal docking conditions suitable for docking of various small lgands by inputing various conditions such as different variations of protein structure and the presence or absence of solvent molecule and so on. Docking using the selected appropriate condition enables us to improve the docking efficiency and avoid an expected failure. The Fig.2 shows the results of applications of k-PALLAS to four targets and indicates that the selected appropriate condition (Best Condition) improve 1% Enrichment Factor (ratio of the number of active compounds recovered in top 1% score compared to that of 1% random selection) compared to average conditions.


Fig.2 Validation results of high precision in silico screening by k-PALLAS

 We developed prediction system by machine learning of interaction patterns (k-MUSES). The k-MUSES can discriminate active compounds from inactive compounds based on new parameters describing the protein-ligand interactions (aPLIED) after reasonable docking using k-PALLAS conditions, to narrow down compounds for a biologocal assay. This k-MUSES system can detect slight difference of interaction patterns between actives and inactives (dummy compounds were used in the case of validations) using cutting edge statistical methods.


Fig.3 Schematic figure of k-MUSES system

 The Fig.4 shows prediction performance of five targets by k-MUSES compared to those of commercial docking score. The figure is called ROC plots that is a widely used for assessing prediction performance. The ROC plots indicate relationship of the number of true positives (vertical axis) and the number of false positives (horizontal axis) in all the ranking by score. The nearer the upper-left of line graph shows higher accuracy. The machine leaning prediction models using aPLIED and Support Vector Machine provided improved precision further compared to commercial docking score (GLIDE score). k-PALLAS and k-MUSES are useful techniques on the front line of real drug discovery projects that can fully utilize both protein and ligand information.


Fig4. Prediction performance by k-MUSES

ADMET prediction technology

 In drug discovery process, it is necessary to improve not only potency for the target but pharmacokinetics (ADME) and toxicity (Tox). Because improvement of ADMET profile is much difficult tasks than enhancement of potency, success of drug discovery largely depends on the ADMET issues. Our unit constructed machine leaning models for ADMET endpoints by Support Vector Machine, Random Forest etc.. We design new drug candidates on hit to lead or lead optimization stages considering the ADMET prediction and/or docking results.


Fig 5. ADMET prediction cycle