Sign In

Multi-level Self-supervised Pretraining on Compositional Hierarchical Graph for Molecular Property Prediction

์ž‘์„ฑ์ž
  • Haebom
์นดํ…Œ๊ณ ๋ฆฌ
Empty

์ €์ž

Xiayu Liu, Zhengyi Lu, Hou-biao Li

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ๋…ผ๋ฌธ์€ ๊ธฐ์กด ๋ถ„์ž ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜ ์ž๊ธฐ ์ง€๋„ ์‚ฌ์ „ ํ•™์Šต ๋ฐฉ๋ฒ•์ด ๋‹จ์ผ ๊ตฌ์กฐ์  ์„ธ๋ถ„์„ฑ๋งŒ์„ ๋‹ค๋ฃจ๊ณ  ๊ฒฐํ•ฉ ์ •๋ณด๋ฅผ ๋ถ€๊ฐ€์ ์ธ ์—ฃ์ง€ ์†์„ฑ์œผ๋กœ ์ทจ๊ธ‰ํ•˜๋Š” ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด, ๋ถ„์ž ๊ตฌ์กฐ๋ฅผ ๋„ค ๊ฐ€์ง€ ๋…ธ๋“œ ์œ ํ˜•๊ณผ ์„ธ ๊ฐ€์ง€ ์˜๋ฏธ๋ก ์  ์ˆ˜์ค€์œผ๋กœ ๊ตฌ์„ฑํ•˜๋Š” ์ƒˆ๋กœ์šด ํ•ฉ์„ฑ ๊ณ„์ธต ๊ทธ๋ž˜ํ”„๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๋‹ค์ค‘ ๋ ˆ๋ฒจ ์ž๊ธฐ ์ง€๋„ ์‚ฌ์ „ ํ•™์Šต ํ”„๋ ˆ์ž„์›Œํฌ์ธ MolCHG๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์›์ž ๊ทธ๋ž˜ํ”„์™€ ๋ณ‘๋ ฌ์ ์œผ๋กœ ์ž‘๋™ํ•˜๋Š” ๊ฒฐํ•ฉ ๊ทธ๋ž˜ํ”„๋ฅผ ๋„์ž…ํ•˜์—ฌ ๊ฒฐํ•ฉ ์ˆ˜์ค€ ์ •๋ณด๋ฅผ ๋…๋ฆฝ์ ์ธ ๋…ธ๋“œ ํ‘œํ˜„์œผ๋กœ ์Šน๊ฒฉ์‹œํ‚ค๊ณ , ์›์ž ๋ฐ ๊ฒฐํ•ฉ ์˜๋ฏธ๋ฅผ ๋™๋“ฑํ•˜๊ฒŒ ํ†ตํ•ฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
์ƒˆ๋กœ์šด ๋ถ„์ž ํ‘œํ˜„ ํ•™์Šต: ๊ฒฐํ•ฉ ์ •๋ณด๋ฅผ ๋ถ€์ฐจ์ ์ธ ์†์„ฑ์ด ์•„๋‹Œ ๋…๋ฆฝ์ ์ธ ๋…ธ๋“œ ํ‘œํ˜„์œผ๋กœ ๋‹ค๋ฃจ๋Š” ๋‹ค์ค‘ ๋ ˆ๋ฒจ ๊ณ„์ธต์  ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ๋ฅผ ํ†ตํ•ด ๋ถ„์ž์˜ ๊ตฌ์กฐ์ , ํ™”ํ•™์  ์˜๋ฏธ๋ฅผ ๋” ํ’๋ถ€ํ•˜๊ฒŒ ํ•™์Šตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
โ€ข
ํ–ฅ์ƒ๋œ ๋ถ„์ž ํŠน์„ฑ ์˜ˆ์ธก ์„ฑ๋Šฅ: ์ œ์•ˆ๋œ MolCHG๋Š” ๋‹ค์–‘ํ•œ ๋ถ„์ž ํŠน์„ฑ ์˜ˆ์ธก ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ธฐ์กด ์ตœ์‹  ๋ฐฉ๋ฒ• ๋Œ€๋น„ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ, ๋‹ค์ค‘ ๋ ˆ๋ฒจ ์ž๊ธฐ ์ง€๋„ ํ•™์Šต์˜ ํšจ๊ณผ์„ฑ์„ ์ž…์ฆํ•ฉ๋‹ˆ๋‹ค.
โ€ข
ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ: ๋‹ค์–‘ํ•œ ๋ถ„์ž ํ‘œํ˜„ ๋ฐ ์‚ฌ์ „ ํ•™์Šต ๋ชฉํ‘œ๋ฅผ ํƒ์ƒ‰ํ•˜๊ฑฐ๋‚˜, ๋” ๋ณต์žกํ•œ ํ™”ํ•™์  ์ƒํ˜ธ์ž‘์šฉ์„ ๋ชจ๋ธ๋งํ•˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ์˜ ํ™•์žฅ ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค.
๐Ÿ‘