• Àüü
  • ÀüÀÚ/Àü±â
  • Åë½Å
  • ÄÄÇ»ÅÍ
´Ý±â

»çÀÌÆ®¸Ê

Loading..

Please wait....

±¹³» ÇÐȸÁö

Ȩ Ȩ > ¿¬±¸¹®Çå > ±¹³» ÇÐȸÁö > µ¥ÀÌÅͺ£À̽º ¿¬±¸È¸Áö(SIGDB)

µ¥ÀÌÅͺ£À̽º ¿¬±¸È¸Áö(SIGDB)

Current Result Document :

ÇѱÛÁ¦¸ñ(Korean Title) Åè½¼ »ùÇøµ ±â¹ÝÀÇ ½Å¾à Èĺ¸ ¹°Áú µðÀÚÀÎÀ» À§ÇÑ ½ÉÃþ »ý¼º ¸ðµ¨
¿µ¹®Á¦¸ñ(English Title) A deep generative model for de novo drug design based on Thompson sampling
ÀúÀÚ(Author) ÃÖ°Ç¿ì   ÀåÈ¿¼ø   ¼­»ó¹Î   ÃÖÁ¾È¯   ±èº´ÁÖ   ¹Ú»óÇö   ÃÖ»ó¹Î   ¹ÚÄ¡Çö   Geon-Woo Choi   Hyosoon Jang   Sangmin Seo   Jonghwan Choi   Byung-ju kim   Sanghyun Park   Sang-min Choi   Chihyun Park  
¿ø¹®¼ö·Ïó(Citation) VOL 38 NO. 02 PP. 0047 ~ 0061 (2022. 12)
Çѱ۳»¿ë
(Korean Abstract)
½ÉÃþ »ý¼º ¸ðµ¨, °­È­ÇнÀ, ÃÖÀûÈ­ ¾Ë°í¸®Áò ±â¹ÝÀ¸·Î »õ·Î¿î ½Å¾à Èĺ¸ ¹°ÁúÀ» »ý¼ºÇÏ´Â ¾à¹° µðÀÚÀÎ ¿¬±¸´Â ´ëºÎºÐ ´Ü¹éÁú ÇÑ °³ÀÇ ±â´ÉÀ» Á¶ÀýÇÒ ¼ö ÀÖ´Â ÀúºÐÀÚÈ­ÇÕ¹° µðÀÚÀÎ ¸ðµ¨À» Á¦½ÃÇÏ°í ÀÖÁö¸¸, ¸ñÀûÇÏ´Â ÀûÀÀÁõÀÇ ¸ÞÄ¿´ÏÁò¿¡ µû¶ó ¿©·¯ ´Ü¹éÁúÀÇ ±â´ÉÀ» Á¶ÀýÇÏ°í, ´Ü¹éÁú ¿Ü¿¡ È­ÇÐÀû Ư¼ºÀ» °í·ÁÇÒ ¼ö ÀÖ´Â ´ÙÁß ¸ñÀû ÃÖÀûÈ­ ±â¹ÝÀÇ ¾à¹° µðÀÚÀÎ ¹æ¹ý·ÐÀÌ ¿ä±¸µÇ°í ÀÖ´Ù. ½ÉÃþ »ý¼º ¸ðµ¨·Î À̸¦ ±¸ÇöÇÒ ¼ö ÀÖ´Â ÇÑ °¡Áö ¹æ¹ýÀº »ùÇøµÀ» ÅëÇØ »çÈÄ È®·ü ºÐÆ÷¸¦ ÃßÁ¤ÇÑ ÈÄ ´ÙÁß ¸ñÀûÀ» ¸¸Á·ÇÏ´Â ¹æÇâÀ¸·Î »ý¼º ¸ðµ¨À» ÃÖÀûÈ­ÇÏ´Â °ÍÀÌ´Ù. À̸¦ À§ÇØ º» ¿¬±¸¿¡¼± ½ÉÃþ »ý¼º ¸ðµ¨ÀÎ Variational Auto Encoder¿¡ Åè½¼ »ùÇøµ ±â¹ýÀ» Á¢¸ñÇÏ¿©, ÀÚü µ¥ÀÌÅÍ Áõ°­°ú ÇнÀÀ» ÅëÇØ ´ÙÁß ¸ñÀûÀ» ¸¸Á·ÇÏ´Â µ¥ÀÌÅ͸¦ »ý¼ºÇÏ°í, À̸¦ ÅëÇØ »ý¼º ¸ðµ¨ÀÇ »çÈÄ ºÐÆ÷¸¦ ÃÖÀûÈ­ÇÏ´Â ¹æ¹ýÀ» Á¦¾ÈÇÑ´Ù. º» ¿¬±¸¿¡¼­´Â BCL2 Æйи® ´Ü¹éÁúµéÀ» Ç¥ÀûÀ¸·Î ¾à¹° »ý¼ºÀ» Á¦¾ÈÇÏ´Â ¸ðµ¨°ú ¿©·¯ º£À̽º¶óÀÎ ¸ðµ¨ÀÌ »ý¼ºÇÑ È­ÇÕ¹°µéÀÇ °áÇÕ Ä£È­µµ¸¦ ºñ±³ÇÏ¿´À¸¸ç, »ý¼ºµÈ ¹°Áú¿¡ ´ëÇÑ È­ÇÐÀûÀΠƯ¼º ÁöÇ¥¿Í ¹°ÁúÀÇ ÇÕ¼º °¡´É¼º ½ºÄھ ÅëÇØ Á¦¾ÈÇÏ´Â ¹æ¹ýÀÇ ½ÇÁ¦ È°¿ë °¡´É¼ºÀ» Á¦½ÃÇÏ¿´´Ù.
¿µ¹®³»¿ë
(English Abstract)
De novo drug design approach, which aims to create a completely novel molecules non-existent in the previous compound library, adopts various machine learning methods such as generative model, reinforcement learning and optimization algorithm. Most of the existing studies suggest the generative model which controls one property. Along with creating a generator, one of the most important factors to be considered in designing an algorithm is an optimization method which can simultaneously guarantee various properties that the molecules must have. One way to implement it on a deep generative model is to estimate the posterior distribution through sampling and then optimize the generator in a direction to satisfy multiple purposes. In this manner, we propose a novel method which optimizes the posterior distribution of the generative model with the generated data which satisfy multiple objectives through self-data augmentation and learning by incorporating the Thomson sampling technique into VAE (Variational Auto-Encoder). To demonstrate the feasibility of the proposed method, we aimed at generating molecules for BCL2 family proteins and compared the binding affinity of the proposed method with several baseline models. The practical usability of the proposed method was confirmed by measuring the various chemical properties and synthesizability of the molecules.
Å°¿öµå(Keyword) µå ³ëº¸ ¾à¹° µðÀÚÀΠ  Åè½¼ »ùÇøµ   ½ÉÃþ »ý¼º ¸ðµ¨   ´ÙÁß ¸ñÀû ÃÖÀûÈ­   de novo drug design   Thompson sampling   Deep generative model   Multi-object optimization  
ÆÄÀÏ÷ºÎ PDF ´Ù¿î·Îµå