Elix, CBI 2021, 招待講演, ElixにおけるAI創薬と最新動向

Slide 1

Slide 1 text

$PQZSJHIU&MJY *OD"MMSJHIUTSFTFSWFE $#*ֶձট଴ߨԋ &MJYʹ͓͚Δ"*૑ༀͱ࠷৽ಈ޲ גࣜձࣾ&MJY $&0݁৓৳࠸ 2021/10/26 1

Slide 2

Slide 2 text

$PQZSJHIU&MJY *OD"MMSJHIUTSFTFSWFE ૑ༀʹ͓͚Δ՝୊ 2 Scannell et al. (2012) • 10೥Ҏ্ͷظؒɺ1000ԯԁҎ্ ͷίετɺ੒ޭ཰ͷ௿͞   • ૑ༀίετ͸ࢦ਺ؔ਺తʹ૿େ   • 2010೥Ҏ߱͸͜ͷ૿Ճ͕ࢭ·ͬͨ ͱ͢Δݚڀ΋͋Δ΋ͷͷґવͱ͠ ͯߴ͍ Eroom's Law

Slide 3

Slide 3 text

$PQZSJHIU&MJY *OD"MMSJHIUTSFTFSWFE 3 3FUIJOLJOH%SVH%JTDPWFSZ ૑ༀΛ࠶ߟ͢Δ

Slide 4

Slide 4 text

"*ؔ࿈ຽؒ౤ࢿֹ೥WT೥ 4 https://aiindex.stanford.edu/report/ 2020೥Ͱ࠷΋େ͖͔ͬͨͷ͸૑ༀɻ͍ͭʹࣗಈӡసΑΓ΋େ͖ͳ౤ࢿֹʹɻ

Slide 5

Slide 5 text

$PQZSJHIU&MJY *OD"MMSJHIUTSFTFSWFE "*ελʔτΞοϓͱ੡ༀձࣾͷఏܞ਺ 5 https://www.biopharmatrend.com/m/free-reports/ai/ ߹ܭఏܞ਺͸ॱௐʹ૿Ճ

Slide 6

Slide 6 text

"*ར༻༻్ͷ಺༁ 6 Schuhmancher et al. (2020) ௿෼ࢠ͕Ұ൪ଟ͍ ※ ੈքͷ੡ༀձࣾτοϓ21͕ࣾ2014ʙ2019 ʹ͔͚ͯऔΓ૊ΜͰ͍ͨAIϓϩδΣΫτ

Slide 7

Slide 7 text

3FTUSJDUFE&MJY *OD ෼ࢠઃܭ 7 Image Source: Sanchez-Lengeling et al. (2018) Drug-likeͳ෼ࢠ͸ʙ1060ݸ ࣮ݧ/γϛϡϨʔγϣϯ ༧ଌϞσϧ ੜ੒Ϟσϧ

Slide 8

Slide 8 text

$PQZSJHIU&MJY *OD"MMSJHIUTSFTFSWFE &MJY%JTDPWFSZ5. "*%SVH%JTDPWFSZ1MBUGPSN 8 Elix͕ഓ͖ٕͬͯͨज़Λ݁ूͨ͠AI૑ༀϓϥοτϑΥʔϜ ᶃڞಉݚڀɹᶄϥΠηϯεఏڙ w ׆ੑ͋Γ w ׆ੑͳ͠ w O. w ʜ

Slide 9

Slide 9 text

$PQZSJHIU&MJY *OD"MMSJHIUTSFTFSWFE &MJY%JTDPWFSZ5. "*%SVH%JTDPWFSZ1MBUGPSN 9 ࣗࣾݚڀΛؚΉ࠷৽ͷݚڀ੒ՌΛϓϥοτϑΥʔϜʹ࣮૷ɻࣗࣾݚڀͷྫΛަ͑ͭͭ঺հɻ ͜ͷޙͷεϙϯαʔηογϣϯͰ΋ϓϥοτϑΥʔϜΛ঺հ → ← CBIͰ΋5ͭޱ಄ൃද w ׆ੑ͋Γ w ׆ੑͳ͠ w O. w ʜ

Slide 10

Slide 10 text

$PQZSJHIU&MJY *OD"MMSJHIUTSFTFSWFE &MJY1SFEJDUʢ༧ଌʣ 10 Properties Model • ׆ੑɺ෺ੑɺADMETͳͲΛ༧ଌ͢ΔϞδϡʔϧ • ैདྷϞσϧ͔Β࠷৽ͷάϥϑܥͷϞσϧ·Ͱ • ࣗಈͰϋΠύʔύϥϝʔλௐ੔Λͯ͠ϕετͳϞσϧΛબ୒ Case Studyɿ   Ξϯυϩήϯड༰ମʹର͢ΔόʔνϟϧεΫϦʔχϯά • ༧ଌ஋͚ͩͰͳ͘ɺCon fi denceείΞΛߟྀ͢Δ͜ͱͰ ΑΓ·ͱ΋ͳԽ߹෺Λબ୒ • ߜΓࠐΜͩ53ݸͷ͏ͪ34ݸ͕ಛڐऔಘࡁΈͷ΋ͷͩͬͨ   ʢ༧ଌϞσϧͷੑೳΛࣔ͢͜ͱ͕໨తͷ࣮ݧʣ Con fi denceείΞ͋Γ Con fi denceείΞͳ͠ 10/27 ޱ಄ൃද O4-3 Romeo Cozac   “Graph Convolutional Networks for Ligand-based Virtual Screening against the Androgen Receptor”

Slide 11

Slide 11 text

$PQZSJHIU&MJY *OD"MMSJHIUTSFTFSWFE &MJY$SFBUFʢ෼ࢠੜ੒ʣ 11 Model • ੜ੒ϞσϧʹΑΓߏ଄ൃੜΛߦ͏Ϟδϡʔϧ • Elix Predictͱ૊Έ߹Θͤͯॴ๬ͷ׆ੑɾ෺ੑΛ࣋ͭ෼ࢠΛੜ੒ • άϥϑϕʔεɺSMILESϕʔεɺϑϥάϝϯτϕʔεͳͲಠࣗϞσϧΛ ؚΉ༷ʑͳϞσϧΛαϙʔτ • ࢦఆͨ͠෦෼ߏ଄Λར༻͢Δ͜ͱ΋Մೳ eGEGLɿElixͷಠࣗϞσϧʢWüthrich et al. 2021ʣ • χϡʔϥϧωοτϫʔΫͱҨ఻తΞϧΰϦζϜΛ ߹ΘͤͨϞσϧ͕SOTAͱ͍͏എܠ • Ҩ఻తΞϧΰϦζϜ෦෼ʹυϝΠϯ஌ࣝΛೖΕΔ ͜ͱΛۃྗഉআͨ͠ΑΓҰൠԽͨ͠Ϟσϧ • ϕϯνϚʔΫͰಉ౳Ҏ্ Wüthrich et al. (2021) 10/27 ޱ಄ൃද O4-3 Pierre Wüthrich   “Using Attribution-based Explainability to Guide Deep Molecular Optimization”

Slide 12

Slide 12 text

$PQZSJHIU&MJY *OD"MMSJHIUTSFTFSWFE 3FUSPTZOUIFTJTʢٯ߹੒ղੳʣ 12 Model • ٯ߹੒ղੳΛߦ͏Ϟδϡʔϧ • େྔͷԽ߹෺Λ·ͱΊͯॲཧ΋Մೳ • Elix CreateͰੜ੒͞ΕͨԽ߹෺ͷ߹੒༰қੑͷݕ౼ʹ΋ • ڞ௨͢ΔதؒମΛܦ༝͢Δϧʔτͷఏࣔ • Ձ֨ɺར༻Մೳͳࢼༀɺऩ཰౳Λߟྀͨ͠ϧʔτͷఏࣔ • ౦ژ޻ۀେֶͱͷڞಉݚڀ • ٯ߹੒ղੳπʔϧ܈ • Φʔϓϯιʔεͱͯ͠ެ։༧ఆ Elix Synthesize 10/27 ޱ಄ൃද O4-2 Haris Hasic     “RetroSynthWAVE: An Open- Source Software Platform for E ffi cient Chemical Synthesis Research”

Slide 13

Slide 13 text

$PQZSJHIU&MJY *OD"MMSJHIUTSFTFSWFE σʔλ͕গͳ͍໰୊΁ͷରԠ 13 ࿈߹ֶशʢkMoL, Elix Milaʣ • ࿈߹ֶशɿσʔλΛ֎෦ʹग़͢͜ͱͳ͘ɺෳ਺اۀͷσʔλΛ׆ ༻ֶͯ͠श͢Δ͜ͱΛՄೳʹ͢Δٕज़ • kMoL: ࿈߹ֶशͱ༧ଌϞσϧͰߏ੒͞ΕΔϥΠϒϥϦ • ژ౎େֶͱڞಉ։ൃ • ࿈߹ֶशϞδϡʔϧElix MilaΛϕʔεʹ։ൃ • Φʔϓϯιʔεͱͯ͠ϦϦʔεࡁΈ (https://github.com/elix-tech/kmol) ࣗݾڭࢣ͋ΓֶशʢSelf-Supervised Learningʣ • ࣮ݧ஋͕ଘࡏ͠ͳ͍Խ߹෺σʔλ΋׆༻ • ༧ଌਫ਼౓Λ޲্ͤ͞ΔΞϓϩʔν • σʔλ͕গͳ͍৔໘Ͱಛʹ༗༻ 10/27 ޱ಄ൃද O4-6 Laurent Dillard   “Improving Molecular Property Prediction using Self-supervised Learning”

Slide 14

Slide 14 text

$PQZSJHIU&MJY *OD"MMSJHIUTSFTFSWFE ௿෼ࢠҎ֎ͷϞμϦςΟ 14 ௿෼ࢠ͕ϝΠϯͰ͋Δ΋ͷͷɺͦΕҎ֎ͷϞμϦςΟ΋ѻ͍ͬͯΔ ߅ମ H3ϧʔϓͷ3࣍ݩߏ଄༧ଌ • λϯύΫ࣭഑ྻσʔλʹΑΔڭࢣͳֶ͠शΛ׆༻ • ͜ͷಛ௃நग़ʹΑΓH3ϧʔϓߏ଄༧ଌͷਫ਼౓Λ޲্ 10/27 ޱ಄ൃද O4-4 David Jimenez   “Leveraging Self-Supervised Contextual Language Models for DNN Antibody CDR-H3 Loop Predictions”

Slide 15

Slide 15 text

$PQZSJHIU&MJY *OD"MMSJHIUTSFTFSWFE ϝϯόʔ 15 🌏 ੈքத͔Β༏लͳϝϯόʔΛ࠾༻ • AIݚڀऀɺAIΤϯδχΞ • έϛετɺόΠΦϩδετ • ത࢜߸औಘऀଟ਺ • ӳޠެ༻ޠ 🇯🇵 🇭🇺 🇫🇷 🇦🇺 🇲🇾 🇰🇬 🇺🇸 🇳🇬 🇨🇭 🇧🇦 🇨🇴 🏢 ૑ۀऀ ݁৓৳࠸ɺPh.D   ڞಉ૑ۀऀɾCEO େٱอୡ໼ ڞಉ૑ۀऀɾCOO ౦ژʹू·ͬͯ࢓ࣄ Λ͍ͯ͠·͢ 📍

Slide 16

Slide 16 text

Thank you! Contact us: [email protected]