$30 off During Our Annual Pro Sale. View Details »

Elix, 第50回構造活性相関シンポジウムランチョンセミナー, SMILES言語モデルSmilesFormerの開発とその応用例

Elix
November 10, 2022

Elix, 第50回構造活性相関シンポジウムランチョンセミナー, SMILES言語モデルSmilesFormerの開発とその応用例

Elix

November 10, 2022
Tweet

More Decks by Elix

Other Decks in Research

Transcript

  1. Copyright © Elix, Inc. All rights reserved.
    4.*-&4ݴޠϞσϧ4NJMFT'PSNFSͷ
    ։ൃͱͦͷԠ༻ྫ

    גࣜձࣾ&MJYϦαʔνΤϯδχΞ
    Ҫ্ وԝ

    View Slide

  2. ෼ࢠߏ଄ͷઃܭͱਂ૚ߏ଄ੜ੒ث
    2

    View Slide

  3. Copyright © Elix, Inc. All rights reserved. 3
    ҩༀ඼Λ͸͡Ίͱ͢ΔػೳੑԽ߹෺ͷߏ଄ઃܭʹ͸ࢼߦࡨޡ͕ඞཁͰɺ࣌ؒͱίετ͕͔͔Δ
    ➔ ߏ଄ઃܭͷखॱΛޮ཰ԽͰ͖Δͱศར
    ػೳੑԽ߹෺ͷ։ൃϓϩηε
    ੒ޭ
    ߏ଄ઃܭ
    ߹੒
    ࢼݧ
    ࣦഊʜʜ

    View Slide

  4. Copyright © Elix, Inc. All rights reserved. 4
    ߏ଄ੜ੒ث͋Δछͷߏ଄ੜ੒نଇΛ༻͍ͯίϯϐϡʔλ্Ͱ෼ࢠߏ଄Λൃੜͤ͞ΔϓϩάϥϜ
    ߏ଄ੜ੒ثΛ༻͍Δ͜ͱͰଟ਺ͷ෼ࢠߏ଄Λੜ੒Ͱ͖Δ
    ଟ਺ͷ෼ࢠߏ଄ΛؚΉੜ੒ߏ଄܈͔Βɺॴ๬ͷੑ࣭Λ࣋ͭͱظ଴͞ΕΔ༗๬ͳԽ߹෺܈Λநग़͢Δ
    ➔ ༗๬Խ߹෺܈ʹରͯ͠ࢼݧ͢Δ͜ͱͰɺߏ଄ઃܭͷखॱΛޮ཰ԽͰ͖Δͱظ଴͞ΕΔ
    ߏ଄ੜ੒ث
    ߏ଄ੜ੒ث ੜ੒ߏ଄܈
    ༗๬ߏ଄܈
    நग़
    ੜ੒

    View Slide

  5. Copyright © Elix, Inc. All rights reserved. 5
    ਂ૚ߏ଄ੜ੒ث
    ༩͑ΒΕͨσʔληοτ͔Βࣗಈநग़͞Εͨ ֬཰Ͱදݱ͞ΕΔ

    ߏ଄ੜ੒نଇΛ༻͍ͨਂ૚ੜ੒Ϟσϧ
    ෼ࢠߏ଄ͷݱΕ΍͢͞ σʔλੜ੒෼෍
    Λදݱ͢Δ

    һ؀ͱһ؀Ͱ͸ɺһ؀ͷํ͕ݱΕ΍͍͢
    $ݪࢠͱ4ݪࢠͰ͸ɺ$ݪࢠͷํ͕෼ࢠ಺ʹଟؚ͘·Ε͍ͯΔ
    ෼ࢠߏ଄ͷಛ௃ʹ͸Ұఆͷ܏޲͕͋Δ
    ➔ ݱΕ΍͍͢෼ࢠߏ଄ʹରͯ͠ɺߴ͍֬཰஋ׂ͕Γ౰ͯΒΕ͍ͯΔ
    ➔ ܇࿅ࡁΈϞσϧ͔ΒͷαϯϓϦϯάͰɺ܇࿅αϯϓϧΒ͍͠
    ෼ࢠߏ଄Λੜ੒Ͱ͖Δ
    ߏ଄ੜ੒نଇ
    ੜ੒
    ੜ੒ߏ଄܈
    ਂ૚ߏ଄ੜ੒ث
    ෼ࢠߏ଄
    σʔληοτ
    نଇͷநग़

    View Slide

  6. Copyright © Elix, Inc. All rights reserved.
    ఆྔతʹ͸ɺ࣍ͷΑ͏ͳࢦඪ͕Α͘ར༻͞Ε͍ͯΔ
    • ෼ࢠߏ଄ͷ༗ޮੑ 7BMJEJUZ

    ੜ੒͞Εͨαϯϓϧͷ͏ͪɺਖ਼͘͠෼ࢠߏ଄ΛදݱͰ͖͍ͯΔ΋ͷͷׂ߹
    • ෼ࢠߏ଄ͷ৽نੑ /PWFMUZ

    ੜ੒ߏ଄܈ͷ͏ͪɺ܇࿅σʔληοτʹؚ·Ε͍ͯͳ͍΋ͷͷׂ߹
    • ෼ࢠߏ଄ͷଟ༷ੑ %JWFSTJUZ

    ೚ҙͷੜ੒ߏ଄ϖΞʹର͢Δ5BOJNPUPڑ཭ ෼ࢠߏ଄ͷඇྨࣅ౓
    ͷฏۉ
    ͨͩ͠ɺ͜͏ͨ͠ࢦඪ͕ྑ͔ͬͨͱͯ͠΋ɺ࣮ࡍʹྑ͍෼ࢠߏ଄͕ಘΒΕΔ͔Ͳ͏͔͸ෆ໌
    ࣮ࡍͷ૑ༀͰͷ࢖༻Λ૝ఆͨ͠໰୊ઃఆͰɺਂ૚ߏ଄ੜ੒ثΛఆੑతʹධՁ͍ͨ͠
    ➔ 4NJMFT'PSNFSͱ͍͏ਂ૚ߏ଄ੜ੒ثΛɺ૑ༀͷϦʔυδΣωϨʔγϣϯͷஈ֊ʹద༻ͯ͠ධՁ
    ਂ૚ߏ଄ੜ੒ثͷੑೳධՁͱͦͷ՝୊
    6

    View Slide

  7. 4NJMFT'PSNFS
    7

    View Slide

  8. Copyright © Elix, Inc. All rights reserved. 8
    จࣈྻʹΑΔ෼ࢠߏ଄ͷදݱ
    4.*-&4จࣈྻ෼ࢠߏ଄Λදݱ͢ΔͨΊͷจࣈྻ
    • Ұఆͷจ๏ʹैͬͯදه͞ΕΔ ҎԼ͸จ๏ͷྫ

    ◦ ೋॏ݁߹͸ l=zͰɺࡾॏ݁߹͸ l#zͰදه ಛʹॻ͔ͳ͍৔߹͸୯݁߹ͱղऍ

    ◦ ओ࠯͔Βͷࢬ෼͔Ε͸ؙΧοί l()zͰදه
    ◦ ؀͸ద౰ͳลΛ੾அͯ͠ߟ͑ɺ੾அՕॴΛରԠ͢Δ਺ࣈͰදه
    ◦ ๕߳؀Λߏ੒͢Δݪࢠ͸খจࣈͰදه
    • ϞσϧͰੜ੒͠΍͍͢෼ࢠߏ଄දݱ
    ◦ 4.*-&4จࣈྻΛઌ಄͔ΒτʔΫϯ จࣈྻͷ࠷খߏ੒୯Ґ
    ͣͭੜ੒
    ▪ τʔΫϯ͸ࣗવݴޠʹ͓͚Δ୯ޠʹ૬౰
    ▪ ݩૉ͝ͱʹҰͭͷτʔΫϯΛͳ͢ͱߟ͑Δ͜ͱ͕ଟ͍
    ◦ ੜ੒͞ΕΔจࣈྻ͕4.*-&4จ๏ʹԊΘͳ͍ ແޮͰ͋Δ
    ͜ͱ΋͋ΓಘΔ
    CC(=O)Nc1ccccc1
    4.*-&4จࣈྻ

    View Slide

  9. Copyright © Elix, Inc. All rights reserved. 9
    4NJMFT'PSNFS
    ೖྗ
    ग़ྗ
    જࡏม਺
    Τϯίʔμ
    σίʔμ
    5SBOTGPSNFS<>ͱݺ͹ΕΔχϡʔϥϧωοτϫʔΫΛ༻͍ͨਂ૚ߏ଄ੜ੒ث
    • छྨͷχϡʔϥϧωοτϫʔΫΛ༻͍ͯೖྗΛʮ຋༁ʯ͢Δ͜ͱͰ
    ෼ࢠߏ଄Λੜ੒͢Δ
    ◦ Τϯίʔμ4.*-&4จࣈྻΛ਺஋ϕΫτϧ જࡏม਺
    ʹม׵
    ◦ σίʔμજࡏม਺Λ΋ͱʹͯ͠4.*-&4จࣈྻΛੜ੒
    • ೖྗ͞Εͨγʔυߏ଄ ෼ࢠߏ଄ɾϑϥάϝϯτߏ଄
    ͔Βજࡏม਺Λܭࢉ
    ➔ ܭࢉ͞Εͨજࡏม਺Λى఺ͱͯ͠ɺજࡏม਺Λগͣͭͣ͠Β͠ͳ͕Β
    ॴ๬ͷ෺ੑɾ׆ੑ஋Λຬͨ͢Α͏ͳ෼ࢠߏ଄Λग़ྗͰ͖Δม਺Λ୳ࡧ
    *
    <>"7BTXBOJ FUBM/FVS*14

    View Slide

  10. Copyright © Elix, Inc. All rights reserved.
    • ܇࿅αϯϓϧͷ4.*-&4จࣈྻΛม׵ͨ͠จࣈྻ͔Βɺ
    ΋ͱͷ4.*-&4จࣈྻΛ෮ݩͰ͖ΔΑ͏ʹ܇࿅
    • 4.*-&4จࣈྻͷม׵ํ๏ ϥϯμϜʹબ୒͞ΕΔ

    ◦ ෳ਺ϑϥάϝϯτ΁ͷ෼ׂ
    3&$"1ΞϧΰϦζϜͳͲΛར༻

    ◦ 4.*-&4จࣈྻͷผදه΁ͷม׵
    ◦ μϛʔจࣈྻ l*z ΁ͷஔ׵
    • े෼ͳ਺ͷ܇࿅αϯϓϧΛ༻͍Δͱɺೖྗ͞Εͨ
    γʔυߏ଄Λద੾ʹิ׬Ͱ͖ΔΑ͏ʹͳΔ
    ◦ ܇࿅ʹ͸෺ੑɾ׆ੑ஋͕ඞཁͳ͍
    ➔ େྔʹ͋Δ
    ϥϕϧͳ͠αϯϓϧΛར༻Մೳ
    4NJMFT'PSNFSͷ܇࿅
    10
    CC1=CC=C(C=C1)C1=CC
    (=NN1C1=CC=C(C=C1)
    S(N)(=O)=O)C(F)(F)F
    ܇࿅αϯϓϧͷ4.*-&4จࣈྻ

    c1(C)ccc(-c2cc(C(F)
    (F)F)nn2-c2ccc(S(N)
    (=O)=O)cc2)cc1
    ผͷ4.*-&4จࣈྻදه

    [*]c1ccc(S(N)(=O)=O)cc1
    [*]c1cc(C(F)(F)F)nn1[*]
    [*]c1ccc(C)cc1
    ϑϥάϝϯτ෼ׂ

    *
    μϛʔจࣈྻ

    4NJMFT'PSNFSͰ
    ෮ݩ
    ϥϯμϜʹ
    ม׵

    View Slide

  11. έʔεελσΟ
    11

    View Slide

  12. Copyright © Elix, Inc. All rights reserved.
    ໨త c࣮ࡍͷ૑ༀͰͷ࢖༻Λ૝ఆͨ͠໰୊ઃఆͰɺਂ૚ߏ଄ੜ੒ثΛఆੑతʹධՁ
    ໰୊ઃఆ cώοτԽ߹෺͔Βطଘༀͷυωϖδϧ΍ɺͦͷ෦෼ߏ଄ΛؚΉΞηνϧίϦϯΤεςϥʔθ
    "$I&
    ્֐ࡎͷߏ଄Λੜ੒
    υωϖδϧ͕ൃݟ͞Εͨͷ͕೥ $I&.#-ʹه࿥͞Ε͍ͯΔ೥

    ➔ ೥Ҏલʹ஌ΒΕ͍ͯͨ"$I&્֐ࡎΛத৺ʹͯ͠ϞσϧΛ܇࿅
    ໰୊ઃఆ
    12
    υωϖδϧͷൃݟ

    ࣌ؒ࣠

    View Slide

  13. Copyright © Elix, Inc. All rights reserved.
    ࣄલֶश༻σʔληοτ dສԽ߹෺

    • $I&.#-͔Βऔಘ ׆ੑσʔλ͸ར༻͠ͳ͍

    • 4NJMFT'PSNFSͷࣄલֶशͰ൚༻తͳϞσϧΛ࡞Δ
    • "$I&્֐ࡎ ສԽ߹෺
    ͱυωϖδϧࠎ֨Λ
    ΋ͭߏ଄ Խ߹෺
    ͸আڈ
    "$I&્֐ࡎσʔληοτ
    • $I&.#-͔Β*$
    ஋ͷ͍ͭͨ Խ߹෺Λऔಘ
    ◦ ͏ͪɺ೥Ҏલʹط஌ͷԽ߹෺Λத৺ʹ
    ܇࿅σʔληοτ "#$ Խ߹෺

    Λ࡞੒
    ➔ 4NJMFT'PSNFSͷϑΝΠϯνϡʔχϯά΍
    ׆ੑ༧ଌϞσϧͷߏஙʹར༻
    ◦ υωϖδϧʹؔ࿈͢Δߏ଄͔ΒͳΔ
    σʔληοτ%͸ݕূσʔληοτͱ͢Δ
    σʔληοτͷ४උ
    13
    " #
    $
    FUD
    FUD FUD
    %
    FUD
    ೥Ҏલʹൃݟ͞Εͨ
    "$I&્֐ࡎ

    "ͷࠎ֨Λ࣋ͭߏ଄

    ೥Ҏ߱ʹൃݟ͞Εͨ΋ͷ΋ؚΉ

    ώοτԽ߹෺ͱ
    ͦͷ༠ಋମ

    ϐϖϥδϯ ϐϖϦδϯ
    Πϯμϯࠎ֨Λ࣋ͭ΋ͷ

    View Slide

  14. Copyright © Elix, Inc. All rights reserved.
    ਂ૚ߏ଄ੜ੒ث c&MJY %JTDPWFSZäʹ࣮૷͞Ε͍ͯΔ4NJMFT'PSNFSΛར༻
    • $I&.#-ͷԽ߹෺ dສԽ߹෺
    Ͱࣄલ܇࿅ͨ͋͠ͱɺ"$I&σʔληοτ "#$ Խ߹෺

    ͰϑΝΠϯνϡʔχϯά
    • ੜ੒࣌͸ෳ਺ͷධՁ஋Λ༻͍ͨଟ໨త࠷దԽΛ࣮ࢪ
    ◦ ߹੒ՄೳੑείΞ 4"TDPSF

    ◦ υϥοάϥΠΫωεࢦඪ 2&%

    ◦ ๬·͍͠෺ཧԽֶతੑ࣭
    ◦ ߏ଄ͷ৽نੑ ܇࿅σʔληοτ͔Βͷڑ཭

    ◦ "$I&્֐׆ੑ༧ଌ஋
    ༧ଌϞσϧ c&MJY %JTDPWFSZäʹ࣮૷͞Ε͍ͯΔάϥϑ৞ΈࠐΈωοτϫʔΫ ($/
    Λར༻
    • "$I&્֐׆ੑͷճؼϞσϧ"$I&σʔληοτ "#$ Խ߹෺
    Ͱ܇࿅
    ➔ 4NJMFT'PSNFSͷ࠷దԽࢦඪɾੜ੒ߏ଄ͷϑΟϧλϦϯάʹར༻
    • ݂ӷ೴ؔ໳ ###
    ಁաੑͷ෼ྨϞσϧผ్༻ҙͨ͠σʔληοτ Խ߹෺
    Ͱ܇࿅
    ➔ ੜ੒ߏ଄ͷϑΟϧλϦϯάʹར༻
    Ϟσϧͷ४උ
    14

    View Slide

  15. Copyright © Elix, Inc. All rights reserved.
    ੜ੒༻ͷγʔυߏ଄ͷछྨ
    γʔυߏ଄ͳ͠ γʔυߏ଄ʹμϛʔจࣈྻ * Λೖྗ

    ୯Ұͷγʔυߏ଄ υωϖδϧൃݟͷݩʹͳͬͨώοτԽ߹෺<>

    ෳ਺ͷγʔυߏ଄ ͷߏ଄ʹྨࣅͨ͠Խ߹෺

    ߏ଄ੜ੒ͱϑΟϧλϦϯάɾϥϯΩϯά
    γʔυߏ଄Λ΋ͱʹສݸͷߏ଄Λੜ੒ º ճ
    ➔ ֤ηοτʹରͯ͠ɺϑΟϧλϦϯά ߏ଄ͷϥϯΩϯά
    • ϑΟϧλϦϯά
    -JQJOTLJͷ30 ෦෼ߏ଄ʹΑΔϑΟϧλϦϯά ߏ଄ͷ৽نੑ ###ಁա֬཰
    • ϥϯΩϯά
    "$I&׆ੑ༧ଌ஋ͱυϥοάϥΠΫωε 2&%
    Λ΋ͱʹϥϯΫ෇͚
    ߏ଄ੜ੒ํ๏
    15
    <>)4VHJNPUP FUBM+QO+1IBSNBDPM r



    View Slide

  16. Copyright © Elix, Inc. All rights reserved.
    ࣮ݧ݁Ռ
    16
    υωϖδϧࠎ֨ υωϖδϧࠎ֨ͷྨࣅߏ଄
    γʔυߏ଄ͳ͠
    γʔυߏ଄ ݸ
    γʔυߏ଄ ෳ਺ݸ


    Λ࣋ͭ΋ͷ
    *$
    O.
    3BOL
    *$
    O.
    3BOL
    *$
    O.
    3BOL
    *$
    O.
    3BOL
    υωϖδϧ
    ݸͷγʔυߏ଄Λ༻͍ͯੜ੒͞Εͨ"$I&્֐ࡎ ෳ਺ݸͷγʔυߏ଄Λ༻͍ͯੜ੒͞Εͨ"$I&્֐ࡎ
    *$
    O.
    3BOL
    *$
    O.
    3BOL
    *$
    O.
    3BOL
    *$
    O.
    3BOL
    υωϖδϧ
    ➔ ώοτԽ߹෺Λγʔυߏ଄ʹར༻͢Δ͜ͱͰɺυωϖδϧΛؚΉ"$I&્֐ࡎͷ࠶ൃݟʹ੒ޭ

    View Slide

  17. Copyright © Elix, Inc. All rights reserved. 17
    • 4NJMFT'PSNFS5SBOTGPSNFSΛ༻͍ͨਂ૚ߏ଄ੜ੒ث
    ◦ ೖྗ͞Εͨ෼ࢠߏ଄ɾϑϥάϝϯτߏ଄͔Βજࡏม਺Λܭࢉ
    ➔ ܭࢉ͞Εͨજࡏม਺Λى఺ͱͯ͠ɺજࡏม਺Λগͣͭͣ͠Β͠ͳ͕Β
    ॴ๬ͷ෺ੑɾ׆ੑ஋Λຬͨ͢Α͏ͳ෼ࢠߏ଄Λग़ྗͰ͖Δม਺Λ୳ࡧ
    • 4NJMFT'PSNFSΛར༻ͯ͠υωϖδϧͷݩͱͳͬͨώοτԽ߹෺͔Βυωϖδϧؔ࿈Խ߹෺Λੜ੒
    ◦ ࣮ࡍͷ૑ༀͰͷ࢖༻Λ૝ఆͨ͠໰୊ઃఆͰɺ4NJMFT'PSNFSΛఆੑతʹධՁ
    ◦ ώοτԽ߹෺Λγʔυߏ଄ʹར༻͢Δ͜ͱͰɺυωϖδϧΛؚΉ"$I&્֐ࡎͷ࠶ൃݟʹ੒ޭ
    ·ͱΊ

    View Slide

  18. Elix, Inc.
    https://elix-inc.com/
    Copyright © Elix, Inc. All rights reserved.

    View Slide