Iyini indima ye-Generated AI ekutholakaleni kwezidakamizwa? [Ividiyo kanye nombuzo]

Impendulo emfushane: I-AI ekhiqizayo isheshisa kakhulu ukutholakala kwezidakamizwa kusenesikhathi ngokukhiqiza ama-molecule noma ukulandelana kwamaprotheni, iphakamisa imizila yokwenziwa, kanye nokuveza imibono evivinywayo, ukuze amaqembu akwazi ukwenza izivivinyo ezimbalwa "ezingaboni". Isebenza kahle kakhulu uma uphoqelela imikhawulo eqinile futhi uqinisekisa imiphumela; uma iphathwa njenge-oracle, ingadukisa ngokuzethemba.

Izinto ezibalulekile okufanele uzicabangele:

Ukusheshisa: Sebenzisa i-GenAI ukuze wandise ukwenziwa kwemibono, bese unciphisa ngokuhlunga okuqinile.

Izithiyo: Zidinga ububanzi bezindawo, imithetho ye-scaffold, kanye nemikhawulo emisha ngaphambi kwesizukulwane.

Ukuqinisekiswa: Phatha imiphumela njengemibono; qinisekisa ngezivivinyo kanye namamodeli ahambisanayo.

Ukulandelelwa: Imiyalelo yelogi, imiphumela, kanye nezizathu ukuze izinqumo zihlale zihlolwa futhi zibuyekezwa.

Ukumelana nokusetshenziswa kabi: Vimbela ukuvuza kanye nokuzethemba ngokweqile ngokuphatha, izilawuli zokufinyelela, kanye nokubuyekezwa kwabantu.

Izihloko ongase uthande ukuzifunda ngemva kwalesi:

🔗 Indima ye-AI ekunakekelweni kwezempilo
Indlela i-AI ethuthukisa ngayo ukuxilongwa, imisebenzi, ukunakekelwa kweziguli, kanye nemiphumela.

🔗 Ingabe i-AI izothatha indawo yodokotela bemisebe?
Ihlola ukuthi ukuzenzakalela kuthuthukisa kanjani i-radiology nokuthi yini ehlala ingumuntu.

🔗 Ingabe i-AI izothatha indawo yodokotela?
Ukubheka ngobuqotho umthelela we-AI emisebenzini nasemsebenzini wodokotela.

🔗 Amathuluzi elabhorethri e-AI amahle kakhulu okuthola isayensi
Amathuluzi elabhorethri e-AI aphezulu okusheshisa izivivinyo, ukuhlaziya, kanye nokuthola.

Indima ye-AI yokukhiqiza ekutholakaleni kwezidakamizwa, ngokuphefumula okukodwa 😮💨

I-AI ekhiqizayo isiza amaqembu emithi ukudala ama-molecule afanelekayo, ukubikezela izakhiwo, ukuphakamisa izinguquko, ukuphakamisa izindlela zokwenziwa, ukuhlola imibono yezinto eziphilayo, nokucindezela imijikelezo yokuphindaphinda - ikakhulukazi ekutholakaleni kwasekuqaleni kanye nokwenza ngcono i-lead. I-Nature 2023 (ukubuyekezwa kokutholwa kwe-ligand) Ukubuyekezwa kwe-Elsevier 2024 (amamodeli okukhiqiza ekwakhiweni kwezidakamizwa okusha)

Futhi yebo, kungadala futhi ngokuzethemba okungenangqondo. Lokho kuyingxenye yesivumelwano. Njengomfundi okhuthele kakhulu onenjini yerokhethi. Umhlahlandlela wodokotela (ingozi yokubona izinto ezingekho) i-npj Digital Medicine 2025 (ukubona izinto ezingekho + uhlaka lokuphepha)

Kungani lokhu kubaluleke ngaphezu kwalokho abantu abakuqaphelayo 💥

Umsebenzi omningi wokuthola ulwazi “ukusesha.” Isikhala samakhemikhali sokusesha, i-biology yokusesha, izincwadi zokusesha, ubudlelwano besakhiwo nomsebenzi wokusesha. Inkinga ukuthi isikhala samakhemikhali… empeleni asinamkhawulo. Ama-Akhawunti Ocwaningo Lwamakhemikhali 2015 (isikhala samakhemikhali) U-Irwin noShoichet 2009 (isikali sesikhala samakhemikhali)

Ungachitha impilo yakho yonke isikhathi eside uzama nje ukushintshashintsha "okunengqondo".

I-AI ekhiqizayo ishintsha umsebenzi kusuka ku:

• “Ake sihlole ukuthi yini esingayicabanga”

ku:

• “Masidale isethi enkulu nehlakaniphile yezinketho, bese sihlola ezingcono kakhulu”

Akukhona mayelana nokuqeda izivivinyo. Kumayelana nokukhetha izivivinyo ezingcono. 🧠 Imvelo 2023 (isibuyekezo sokutholakala kwe-ligand)

Futhi, futhi lokhu akuxoxwa ngakho kakhulu, kusiza amaqembu ukuthi axoxe ngemikhakha eyahlukene. Osokhemisi, izazi zezinto eziphilayo, abantu be-DMPK, ososayensi bekhompyutha… wonke umuntu unezinhlobo ezahlukene zengqondo. Uhlelo oluhle lokukhiqiza lungasebenza njengephedi yokudweba eyabiwe. Isibuyekezo se-Frontiers in Drug Discovery 2024

Yini eyenza inguqulo enhle ye-AI yokukhiqiza ekutholakaleni kwezidakamizwa? ✅

Akuzona zonke i-AI ezikhiqizayo ezidalwe zilingana. Inguqulo "enhle" yalesi sikhala ayimayelana nama-demo akhangayo kodwa imayelana nokuthembeka okungafanele (ukungafaneleki kuyimfanelo lapha). I-Nature 2023 (isibuyekezo sokutholakala kwe-ligand)

Ukusethwa okuhle kwe-AI okukhiqizayo kuvame ukuba nalokhu:

• Ukusekelwa kwesizinda: kuqeqeshwe noma kulungiswe ngedatha yamakhemikhali, yezinto eziphilayo, kanye neyemithi (hhayi umbhalo ojwayelekile nje kuphela) 🧬 Isibuyekezo sika-Elsevier 2024 (amamodeli akhiqizayo)

• Izithiyo-isizukulwane sokuqala: ingalalela imithetho efana nobubanzi be-lipophilicity, imikhawulo ye-scaffold, izici zesayithi ezibophayo, imigomo yokukhetha i- JCIM 2024 (amamodeli okusabalalisa ekwakhiweni kwezidakamizwa ezintsha) REINVENT 4 (uhlaka oluvulekile)

• Ukuqwashisa ngempahla: kukhiqiza ama-molecule angagcini nje ngokuba amasha kodwa futhi “angahlekisayo” ngokwe-ADMET i-ADMEtlab 2.0 (kungani i-ADMET yokuqala ibalulekile)

• Ukubika ukungaqiniseki: kubonisa ukuthi iqagela nini uma iqhathaniswa nokuthi iqinile nini (ngisho nebhendi yokuqiniseka engacwengekile iyasiza) Izimiso zokuqinisekisa ze-OECD QSAR (isizinda sokusebenza)

• Ukulawula okungaphakathi komuntu: osokhemisi bangaqondisa, benqabe, futhi baqondise imiphumela ngokushesha Imvelo 2023 (ukuhamba komsebenzi + umongo wobuchwepheshe bokuthola)

• Ukulandeleka: ungabona ukuthi kungani kwenzeke isiphakamiso (okungenani ingxenye), noma uphaphazela nge -OECD QSAR isiqondiso (ukubonakala kwemodeli + ukuqinisekiswa)

• I-harness yokuhlola: i-docking, i-QSAR, izihlungi, ukuhlolwa kwe-retrosynthesis - konke kuxhunywe ku-🔧 I-Nature 2023 (isibuyekezo sokutholwa kwe-ligand) Ukufunda Komshini ku-CASP (Coley 2018)

• Ukukhetha kanye nokulawula ukuvuza: ukugwema ukubamba ngekhanda idatha yokuqeqesha ngokunyenya (yebo, kuyenzeka) USENIX 2021 (ukuqeqeshwa kokukhipha idatha) Vogt 2023 (izinkinga ezintsha/ezihlukile)

Uma i-AI yakho yokukhiqiza ingakwazi ukubhekana nemingcele, empeleni iyisikhiqizi sezinto ezintsha. Kumnandi emaphathini. Akumnandi kakhulu ohlelweni lwezidakamizwa.

Lapho i-AI ekhiqizayo ifanelana khona nomzila wokuthola izidakamizwa 🧭

Nansi imephu yengqondo elula. I-AI ekhiqizayo ingaba negalelo cishe kuzo zonke izigaba, kodwa isebenza kahle kakhulu lapho ukuphindaphinda kubiza khona futhi isikhala se-hypothesis sikhulu. I-Nature 2023 (isibuyekezo sokutholwa kwe-ligand)

Izindawo zokuthinta ezivamile:

• Ukutholwa kanye nokuqinisekiswa kwenhloso (izimvo, imephu yendlela, iziphakamiso ze-biomarker) Isibuyekezo se-Frontiers in Drug Discovery 2024

• Ukuhlonza i-hit (ukwengezwa kokuhlolwa okubonakalayo, ukukhiqizwa okusha kwe-hit) I-Nature Biotechnology 2019 (GENTRL)

• Ukulungiswa kwe-lead (okuphakamisa ama-analog, ukulungiswa kwamapharamitha amaningi) REINVENT 4

• Usekelo lwangaphambi kokwelashwa (ukubikezela impahla ye-ADMET, amacebiso okwakha ngezinye izikhathi) I-ADMETlab 2.0

• Ukuhlela i-CMC kanye nokuhlanganiswa (iziphakamiso zokwenziwa kabusha, ukuhlelwa kabusha komzila) I-AiZynthFinder 2020 Coley 2017 (ukwenziwa kabusha okusiza ngekhompyutha)

• Umsebenzi wolwazi (ukuhlanganiswa kwezincwadi, izifinyezo zezindawo ezincintisanayo) 📚 Amaphethini ka-2025 (ama-LLM ekutholakaleni kwezidakamizwa)

Ezinhlelweni eziningi, impumelelo enkulu ivela ekuhlanganisweni komsebenzi, hhayi kumodeli eyodwa "eyingcweti." Imodeli iyinjini - umzila wepayipi yimoto. I-Nature 2023 (isibuyekezo sokutholwa kwe-ligand)

Ithebula Lokuqhathanisa: izindlela ezidumile zokukhiqiza i-AI ezisetshenziswa ekutholakaleni kwezidakamizwa 📊

Itafula elingaphelele kancane, ngoba impilo yangempela ayiphelele kancane.

Amanothi: amanani ayahlukahluka kakhulu kuye ngokuthi ingakanani, ukubala, ilayisensi, nokuthi ithimba lakho lifuna "ukuxhuma nokudlala" noma "asakhe umkhumbi-mkhathi."

Ukubheka eduze: I-AI Ekhiqizayo yokuthola izinto ezintsha kanye nomklamo omusha 🧩

Nasi isibonelo esiyinhloko sokusetshenziswa: khiqiza ama-molecule afanelekayo kusukela ekuqaleni (noma esikhaleni) afana nephrofayili eqondiwe. I-Nature Biotechnology 2019 (GENTRL) IYAVULA 4

Indlela esebenza ngayo ngokuvamile:

1. Chaza imikhawulo

   • ikilasi eliqondiwe, ukuma kwephakethe elibophayo, ama-ligand aziwayo

   • ububanzi bezakhiwo (ukuncibilika, i-logP, i-PSA, njll.) I-Lipinski (Umthetho wesi-5 womongo)

   • imikhawulo yobusha (gwema izindawo ze-IP ezaziwayo) 🧠 I-Vogt 2023 (ukuhlolwa kwentsha)

2. Khiqiza abantu abazongenela ukhetho

   • ukugxuma kwe-scaffold

   • ukukhula kwengxenye

   • iziphakamiso "zokuhlobisa lesi sihluthulelo"

   • ukukhiqiza okunezinhloso eziningi (ukubopha + okuvulekayo + okungenabo ubuthi) UKUVULA KABUSHA 4 Isibuyekezo sika-Elsevier 2024 (amamodeli akhiqizayo)

3. Hlunga ngokucindezela

   • imithetho yekhemistri yezokwelapha

   • Izihlungi ze-PAINS kanye neqembu elisabelayo i-Baell & Holloway 2010 (PAINS)

   • ukuhlolwa kokwenziwa kwezinto i- AiZynthFinder 2020

   • ukudokha / ukufaka amaphuzu (akuphelele kodwa kuyasiza) Imvelo 2023 (isibuyekezo sokutholakala kwe-ligand)

4. Khetha isethi encane yokwenziwa

   • abantu basakhetha, ngoba abantu banganuka iphunga elingenangqondo ngezinye izikhathi

Iqiniso elixakile: inani alikho nje “ama-molecule amasha.” Ama -molecule amasha anengqondo ngemikhawulo yohlelo lwakho. Leyo ngxenye yokugcina iyikho konke. I-Nature 2023 (isibuyekezo sokutholakala kwe-ligand)

Futhi, kukhona ukweqisa okuncane: uma kwenziwe kahle, kungazwakala sengathi uqashe ithimba losokhemisi abancane abangakhathali abangalali futhi abangalokothi bakhononde. Futhi, abaqondi ukuthi kungani isu elithile lokuvikela liyiphupho elibi, ngakho… bhalansi 😅.

Ukubheka eduze: Ukulungiswa kwe-lead nge-AI ekhiqizayo (ukulungiswa kwamapharamitha amaningi) 🎛️

Ukuthuthukisa i-lead yilapho amaphupho eba yinkimbinkimbi khona.

Uyafuna:

• amandla aphezulu

• ukukhetha kuphezulu

• ukuzinza kwe-metabolic kuyanda

• ukuncibilika kuphezulu

• izimpawu zokuphepha ziphansi

• ukuvuleka “kulungile”

• FUTHI kusazokwenziwa

Lokhu kuyindlela yokwenza ngcono ye-multi-objective yakudala. I-AI ekhiqizayo inhle kakhulu ekuphakamiseni isethi yezixazululo zokuhwebelana kunokuba yenze sengathi kukhona inhlanganisela eyodwa ephelele. BUYISA Ukubuyekezwa kwe-Elsevier 2024 okungu-4 (amamodeli akhiqizayo)

Izindlela ezisebenzayo amaqembu ayisebenzisa ngazo:

• Isiphakamiso se-analog: "Yenza izinhlobo ezingu-30 ezinciphisa ukukhishwa kodwa zigcine amandla"

• Ukuskena okushintshiwe: ukuhlola okuqondisiwe esikhundleni sokubalwa kwamandla aluhlaza

• Ukugxuma kwe-scaffold: lapho i-core ishaya udonga (ubuthi, i-IP, noma ukuzinza)

• Chaza iziphakamiso: “Leli qembu le-polar lingasiza ekuncibilikeni kodwa lingalimaza ukungena kwamanzi” (akulungile ngaso sonke isikhathi, kodwa kuyasiza)

Isexwayiso esisodwa: izibikezeli zezakhiwo zingaba buthakathaka. Uma idatha yakho yokuqeqesha ingahambisani nochungechunge lwakho lwamakhemikhali, imodeli ingaba nephutha ngokuqiniseka. Njengokuthi, ayilungile kakhulu. Futhi ngeke ihlazeke. Izimiso zokuqinisekisa ze-OECD QSAR (isizinda sokusebenza) Weaver 2008 (isizinda sokusebenza se-QSAR)

Ukubheka eduze: I-ADMET, ubuthi, kanye nokuhlolwa kokuthi “ngicela ungabulali uhlelo” 🧯

I-ADMET yilapho abantu abaningi abafaka izicelo behluleka khona buthule. I-AI ekhiqizayo ayixazululi i-biology, kodwa inganciphisa amaphutha angagwenywa. I-ADMETlab 2.0 Waring 2015 (ukulahlekelwa)

Izindima ezivamile:

• ukubikezela izikweletu ze-metabolic (izindawo ze-metabolism, izitayela zokususa)

• ukumaka izisusa ezinobuthi ezingaba khona (izaziso, ama-proxies asebenzayo)

• ukulinganisa ububanzi bokuncibilika kanye nokuvuleka kwamanzi

• esikisela izinguquko zokunciphisa ingozi ye-hERG noma ukuthuthukisa ukuzinza 🧪 I-FDA (ICH E14/S7B Q&A) EMA (ukubuka konke kwe-ICH E14/S7B)

Iphethini ephumelela kakhulu ivame ukufana nalokhu: sebenzisa i-GenAI ukuphakamisa izinketho, kodwa sebenzisa amamodeli akhethekile kanye nokuhlolwa ukuqinisekisa.

I-AI ekhiqizayo iyinjini yokucabanga. Ukuqinisekisa kusasebenza ekuhlolweni.

Ukubheka eduze: I-AI Ekhiqizayo ye-biologics kanye nobunjiniyela bamaprotheni 🧬✨

Ukutholakala kwezidakamizwa akugcini nje ngokutholakala kwama-molecule amancane. I-AI ekhiqizayo isetshenziselwa futhi:

• ukukhiqizwa kokulandelana kwama-antibody

• iziphakamiso zokuvuthwa kobuhlobo

• ukuthuthukiswa kokuzinza kwamaprotheni

• ubunjiniyela be-enzyme

• ukuhlolwa kwe-peptide therapeutics I -ProteinMPNN (Isayensi 2022) I-Rives 2021 (amamodeli olimi lweprotheni)

Ukukhiqizwa kwamaprotheni nokulandelana kungaba namandla ngoba "ulimi" lwezinhlu luhambisana kahle ngendlela emangalisayo nezindlela ze-ML. Kodwa nansi indlela engavamile yokubuyela emuva: ihambisana kahle… kuze kube yilapho ingasebenzi. Ngoba ukuzivikela komzimba, ukubonakaliswa, amaphethini e-glycosylation, kanye nemikhawulo yokuthuthuka kungaba yingozi. AlphaFold (Nature 2021) ProteinGenerator (Nat Biotech 2024)

Ngakho-ke, izinketho ezinhle kakhulu zifaka:

• izihlungi zokuthuthuka

• ukuhlolwa kwengozi yokuzivikela komzimba

• imikhawulo yokukhiqiza

• ama-loop elebhu amanzi ukuze kuphindwe ngokushesha 🧫

Uma uzigwema lezo, uthola uhlu oluhle kakhulu oluziphatha njenge-diva ekukhiqizweni.

Ukubheka eduze: Ukuhlela ukwenziwa kanye neziphakamiso zokwenziwa kabusha 🧰

I-AI ekhiqizayo nayo ingena ngokunyenya emisebenzini yamakhemikhali, hhayi nje ekucabangeni ngama-molecule.

Abahleli be-Retrosynthesis bangenza okulandelayo:

• phakamisa imizila eya endaweni ethile eqondiwe

• phakamisa izinto zokuqala ezitholakala kwezentengiselwano

• izindlela zokuhlela ngezinyathelo noma okubonwayo ukuthi kungenzeka

• siza osokhemisi ukuthi basuse ngokushesha imibono "emihle kodwa engenakwenzeka" i -AiZynthFinder 2020 Coley 2018 (CASP)

Lokhu kungonga isikhathi sangempela, ikakhulukazi uma uhlola izakhiwo eziningi ezingaba khona. Noma kunjalo, abantu babaluleke kakhulu lapha ngoba:

• izinguquko zokutholakala kwe-reagent

• ukukhathazeka ngokuphepha kanye nobukhulu kuyiqiniso

• ezinye izinyathelo zibukeka kahle ephepheni kodwa zehluleka ngokuphindaphindiwe

Isingathekiso esingaphelele, kodwa ngizosisebenzisa noma kunjalo: i-retrosynthesis I-AI ifana ne-GPS elungile kakhulu, ngaphandle kokuthi ngezinye izikhathi ikuhambisa echibini futhi iphikelela ukuthi iyindlela emfushane. 🚗🌊 Coley 2017 (i-retrosynthesis esizwa yikhompyutha)

Idatha, amamodeli amaningi, kanye neqiniso elibi lamalebhu 🧾🧪

I-AI ekhiqizayo iyayithanda idatha. Amalebhu akhiqiza idatha. Ephepheni, lokho kuzwakala kulula.

Cha.

Idatha yangempela yelabhorethri yile:

• okungaphelele

• umsindo

• egcwele imiphumela yeqembu uLeek et al. 2010 (imiphumela yeqembu) iNature 2024 (imiphumela yeqembu ekuthathweni kwezithombe zamaseli)

• kusakazeke kuwo wonke amafomethi

• ubusiswe ngemithetho yokuqamba amagama “yokudala”

Izinhlelo zokukhiqiza eziningi zingahlanganisa:

• imiphumela yokuhlolwa

• izakhiwo zamakhemikhali

• izithombe (i-microscopy, i-histology)

• ama-omics (ama-transcriptomics, ama-proteomics)

• umbhalo (amaphrothokholi, ama-ELN, imibiko) i-npj Digital Medicine 2025 (izindlela eziningi ku-biotech) Ukuhlaziywa Kwesithombe Sezokwelapha 2025 (i-AI yezindlela eziningi kwezokwelapha)

Uma isebenza, iyamangalisa. Ungathola amaphethini angabonakali bese uphakamisa izivivinyo ezingaphuthelwa uchwepheshe oyedwa.

Uma yehluleka, yehluleka buthule. Ayivali umnyango. Imane ikushukumisela esiphethweni esingalungile esiqinisekile. Yingakho ukubusa, ukuqinisekiswa, kanye nokubuyekezwa kwesizinda kungeyona into ongayikhetha. Umhlahlandlela wodokotela (ukungaqondi kahle) npj Digital Medicine 2025 (ukungaqondi kahle + uhlaka lokuphepha)

Izingozi, imikhawulo, kanye nesigaba esithi “ungakhohliswa umphumela ocacile” ⚠️

Uma ukhumbula into eyodwa kuphela, khumbula lokhu: i-AI ekhiqizayo iyakholisa. Ingazwakala ilungile ngenkathi ingalungile. Umhlahlandlela wodokotela (ukungaqondi kahle)

Izingozi Eziyinhloko:

• Izindlela ezingaqondakali: i-biology enengqondo engeyona isiqondiso sangempela sodokotela (imibono engaqondakali)

• Ukuvuza kwedatha: ukudala into esondele kakhulu kuma-compound aziwayo i-USENIX 2021 (ukuqeqeshwa kokukhishwa kwedatha) i-Vogt 2023 (izinkinga ezintsha/ezihlukile)

• Ukwenza ngcono kakhulu: ukuxosha amaphuzu abikezelwe angahumushi i-in vitro Nature 2023 (isibuyekezo sokutholwa kwe-ligand)

• Ukubandlulula: idatha yokuqeqesha eqondiswe kuzinhlobo ezithile ze-chemo noma izinhloso i -Vogt 2023 (ukuhlolwa kwemodeli + ukubandlulula/okusha)

• Okusha okungamanga: ama-molecule “amasha” empeleni ayizinhlobo ezincane Vogt 2023

• Izikhala zokuchaza: kunzima ukuzithethelela izinqumo ezibhekiswe kubabambiqhaza Izimiso zokuqinisekisa ze-OECD QSAR

• Ukukhathazeka ngezokuphepha kanye ne-IP: imininingwane yohlelo ebucayi ezixwayisweni 😬 USENIX 2021 (ukuqeqeshwa kokukhishwa kwedatha)

Izinyathelo zokunciphisa ezisiza ekusebenzeni:

• gcina abantu besesimeni sokukhetha

• izicelo zelogi kanye nemiphumela yokulandelela

• qinisekisa ngezindlela eziqondile (izivivinyo, amamodeli ahlukile)

• sebenzisa imikhawulo futhi uhlunge ngokuzenzakalelayo

• phatha imiphumela njengemibono, hhayi amaphilisi eqiniso isiqondiso se-OECD QSAR

I-AI Ekhiqizayo iyithuluzi lamandla. Amathuluzi kagesi awakwenzi ube ngumbazi… avele enze amaphutha ngokushesha uma ungazi ukuthi wenzani.

Amaqembu asebenzisa kanjani i-AI ekhiqizayo ngaphandle kwesiphithiphithi 🧩🛠️

Amaqembu avame ukufuna ukusebenzisa lokhu ngaphandle kokuguqula inhlangano ibe umbukiso wesayensi. Indlela yokwamukelwa ewusizo ibukeka kanje:

• Qala nge-bottleneck eyodwa (ukwanda kwe-hit, ukukhiqizwa kwe-analog, i-triage yezincwadi) I-Nature 2023 (isibuyekezo sokutholwa kwe-ligand)

• Yakha iluphu yokuhlola eqinile (izihlungi + ukudokha + ukuhlolwa kwempahla + ukubuyekezwa kwe-chem) HLAZIYA 4 I-AiZynthFinder 2020

• Linganisa imiphumela (isikhathi esilondoloziwe, izinga lokushaya, ukunciphisa ukwehla) I-Waring 2015 (ukwehla)

• Hlanganisa namathuluzi akhona (i-ELN, irejista ehlanganisiwe, izizindalwazi zokuhlola) Isisetshenziswa se-ELN sase-Edinburgh

• Dala imithetho yokusebenzisa (okungacelwa, okuhlala kungaxhunyiwe ku-inthanethi, izinyathelo zokubuyekeza) USENIX 2021 (ingozi yokukhipha idatha)

• Qeqesha abantu ngobumnene (ngokungathi sína, amaphutha amaningi avela ekusetshenzisweni kabi, hhayi kumodeli) Umhlahlandlela wodokotela (ukungaboni kahle)

Futhi, ungalithathi kancane isiko. Uma osokhemisi bezwa sengathi i-AI iyabacindezela, bazoyishaya indiva. Uma kubasindisa isikhathi futhi kuhlonipha ubuchwepheshe babo, bazokwamukela ngokushesha. Abantu bahlekisa kanjalo 🙂.

Iyini indima ye-AI yokukhiqiza ekutholakaleni kwezidakamizwa uma usondeza ngaphandle? 🔭

Uma sekukhululiwe, indima akuyona “ukufaka esikhundleni sososayensi.” “Ukwandisa umkhawulokudonsa wesayensi.” I-Nature 2023 (isibuyekezo sokutholakala kwe-ligand)

Kuyasiza amaqembu:

• hlola imibono eyengeziwe ngesonto

• phakamisa izakhiwo ezengeziwe ezifanelekayo ngomjikelezo ngamunye

• beka phambili izivivinyo ngokuhlakanipha okukhulu

• cindezela izihibe zokuphindaphinda phakathi komklamo nokuhlolwa

• yabelana ngolwazi kuzo zonke izikhungo zokuthengisa izidakamizwa ngo-2025 (ama-LLM ekutholakaleni kwezidakamizwa)

Futhi mhlawumbe ingxenye engahlonishwa kakhulu: ikusiza ukuthi ungachithi ubuhlakani bomuntu obubizayo emisebenzini ephindaphindwayo. Abantu kufanele bacabange ngendlela yokusebenza, isu, kanye nokuhumusha - hhayi ukuchitha izinsuku bekhiqiza uhlu oluhlukile ngesandla. I-Nature 2023 (isibuyekezo sokutholwa kwe-ligand)

Ngakho-ke yebo, indima ye-AI ekhiqizayo ku-Drug Discovery iyi-accelerator, i-generator, isihlungi, futhi ngezinye izikhathi idala izinkinga. Kodwa iwusizo.

Isifinyezo sokuvala 🧾✅

I-AI ekhiqizayo isiba yikhono eliyinhloko ekutholakaleni kwezidakamizwa zanamuhla ngoba ingakhiqiza ama-molecule, imibono, ukulandelana, kanye nemizila ngokushesha kunabantu - futhi ingasiza amaqembu ukuthi akhethe ukuhlolwa okungcono. Isibuyekezo se-Frontiers in Drug Discovery 2024 Nature 2023 (isibuyekezo sokutholakala kwe-ligand)

Izinhlamvu ezifingqiwe:

• Kungcono kakhulu ekutholeni kusenesikhathi kanye nokulungisa ama-lead loops ⚙️ HLAZIYA 4

• Isekela ama-molecule amancane kanye ne-biologics GENTRL (Nature Biotech 2019) I-ProteinMPNN (Science 2022)

• Kukhulisa umkhiqizo ngokwandisa i-idealine funnel Nature 2023 (isibuyekezo sokutholwa kwe-ligand)

• Kudinga imingcele, ukuqinisekiswa, kanye nabantu ukuze kugwenywe ukungabi nangqondo okuqinisekile Izimiso ze-OECD QSAR Umhlahlandlela wabahlengikazi (ukungaboni kahle)

• Ukuphumelela okukhulu kuvela ekuhlanganisweni komsebenzi, hhayi ekukhangiseni okuvela ku -Nature 2023 (isibuyekezo sokutholwa kwe-ligand)

Uma uyiphatha njengomuntu osebenzisana naye - hhayi njengomuntu okhuluma izilimi - ingaqhubekisela phambili izinhlelo. Futhi uma uyiphatha njengomuntu okhuluma izilimi ... kahle, ungase ugcine ulandela leyo GPS echibini futhi. 

Isibonelo sangempela: Ukwakha umsebenzi wokukhiqiza ama-molecule okuvimbela kuqala 🧪

Isimo

Ithimba elincane le-biotech eliqanjiwe kodwa elingokoqobo lisebenza endaweni yokuhlasela izifo ezivuvukalayo. Selivele linemiphumela engu-42 eqinisekisiwe ebuthakathaka ekuhlolweni, kodwa iningi lalo alincibiliki kahle, kanti abambalwa baseduze kakhulu nendawo yobunikazi yabancintisani.

Esikhundleni sokucela imodeli ekhiqizayo ukuthi “ithole ama-molecule angcono” - okuyisimemo sokuthola izinto ezingenangqondo ezinhle - ithimba lakha umsebenzi oqinile wokwandisa ama-hit.

Umgomo ulula: ukukhiqiza isethi ebanzi yama-analogue, uwahlunge kakhulu, bese uthumela kuphela abantu abavikelekile kakhulu ekubuyekezweni kwekhemistri yezokwelapha.

Lokho okudingwa umsizi

Ithimba linikeza uhlelo:

iphrofayela eqondiwe kanye nolwazi lwe-ligand olwaziwayo

izakhiwo ezingu-42 eziqinisekisiwe zokushayiswa

imikhawulo yempahla yesisindo sama-molecule, i-logP, i-TPSA, ukuncibilika, kanye nokususwa okubikezelwe

izisekelo ezivinjiwe kanye nemingcele yokufana yokugwema i-IP

Izihlungi ze-PAINS kanye neqembu elisabelayo i-Baell & Holloway 2010

Ukubikezela kwe-ADMET kuhlola i-ADMEtlab 2.0

ukuhlolwa kokwenzeka kwe-retrosynthesis AiZynthFinder 2020

imithetho yokubuyekezwa kwabantu yokukhetha kokugcina

Into ebalulekile: imodeli ayivunyelwe ukuthuthukisa amandla ayo ngokwayo. Kufanele ilinganisele amandla, ubusha, ukuthuthuka, kanye nokwenzeka kokwenziwa.

Isibonelo semiyalelo

Khiqiza imibono ye-analogue engu-150 ngokusekelwe kulezi zakhiwo eziqinisekisiwe zokushaya. Gcina isisindo sama-molecule siphakathi kuka-300 no-480, i-logP ebikezelwe iphakathi kuka-1.5 no-4.0, i-TPSA ingaphansi kuka-110, futhi ugweme ama-scaffolds avinjiwe abhalwe kufayela le-IP. Beka phambili izakhiwo ezingenazo izexwayiso ze-PAINS, amaqembu asebenzayo asobala, kanye nomzila wokwenziwa onengqondo wezinyathelo ezinhlanu noma ngaphansi. Ku-molecule ngayinye, chaza ukuguqulwa okuyinhloko, ukuthuthukiswa kwesakhiwo okuhlosiwe, ingozi eyinhloko, nokuthi ingabe i-compound kufanele yenqatshwe, ibuyekezwe, noma ibekwe phambili.

Indlela yokuyihlola

Ithimba alithembi umphumela wokuqala. Basebenzisa iluphu encane yokuhlola:

Hlola ukuthi ama-molecule akhiqiziwe ayayithobela yini imikhawulo yempahla

Susa izinto ezicishe zifane nezifanekiso kanye nezakhiwo eziseduze kakhulu nezinhlanganisela ezaziwayo

Sebenzisa izihlungi ze-PAINS, iqembu elisabelayo, kanye nezihlungi eziyisisekelo ze-chemistry yezokwelapha

Sebenzisa imodeli yesibili yesakhiwo ukuze uqhathanise izibikezelo ze-ADMET

Cela osokhemisi ababili ukuthi babhale ngokuzimela amaphuzu angu-30 aphezulu

Thumela uhlu olufushane oluthole amaphuzu amaningi kuphela engxoxweni yokuhlanganisa

Umbuzo wokuhlola obalulekile uthi: “Besingasacabanga ngale molekyuli ukube i-AI ayizange iyiphakamise?”

Uma impendulo ingucha, ithimba libuza ukuthi kungani. Ngezinye izikhathi lokho kwembula umqondo omusha omuhle. Ngezinye izikhathi kwembula ukucabanga okufiswayo okuqhutshwa yimodeli.

Umphumela

Umphumela obonisa kuphela - hhayi ucwaningo lwenkampani yangempela.

Ngokusekelwe ekubekeni isikhathi imisebenzi emithathu yokwandisa amasampula, ukuhamba komsebenzi ngesandla kuthathe cishe amahora ama-5 ukudala nokuhlola imibono engama-analogue engama-60. Ukuhamba komsebenzi kwe-GenAI kokuqala okukhawulelwe kukhiqize abantu bokuqala abayi-150 cishe ngemizuzu engama-55.

Ngemva kokuhlunga, abantu abangu-27 kuphela abasindile esikrinini esigcwele. Kulabo, osokhemisi bamake abangu-9 njengabafanele ukubuyekezwa okujulile, abangu-12 “njengabathakazelisayo kodwa abayingozi”, kanye nabangu-6 njengabenqatshwayo lapho kubuyekezwa.

Lokho kusho ukuthi umphumela obalulekile wawungewona “ama-molecule amasha angu-150”. Umphumela obalulekile wawungu-9 ongase ubuyekezwe ngaphansi kwehora eli-1, kanye nomzila ocacile wokuhlola obonisa ukuthi yimiphi imingcele edlulile noma ehlulekile kumuntu ngamunye.

Ithimba lingakuqinisekisa lokhu ngokulandelela:

isikhathi esichithwa ngomjikelezo ngamunye womklamo

inani lezakhiwo ezikhiqizwe

iphesenti lisusiwe yizihlungi

izinga lokwamukelwa kwamakhemikhali

inani labantu abakhethiwe ukuze bahlanganise

inombolo kamuva iqinisekiswe ukuthi iyasebenza ekuhlolweni

Yini engase ihambe kabi

Imodeli ingase ithuthukise izihlungi esikhundleni sokuphakamisa ikhemistri ezwakala kahle ngempela.

Umuntu ozongenela ukhetho angabonakala emuhle kakhulu ku-ADMET ebikezelwe kodwa ahluleke ngokushesha ekuhlolweni kwangempela. Izimiso zokuqinisekiswa kwe-OECD QSAR

Iziphakamiso zokwenziwa kabusha zingase zibonakale zinengqondo ngenkathi zithembele kuma-reagent angatholakali, izimo ezingavamile, noma i-chemistry engaphephile.

Isihlungi esisha singasusa ama-compounds abalulekile ngamandla amakhulu, noma sivumele ama-molecule asondele kakhulu ku-IP eyaziwayo.

Iphutha elikhulu kakhulu ukuphatha uhlu olubekwe ohlwini njengeqiniso. Luwuhlu lwemibono olubekwe phambili kuphela.

Ukudla okuwusizo

Ukusetshenziswa okungcono kakhulu kwe-AI yokukhiqiza ekutholakaleni kwezidakamizwa akusikho "ukucindezela inkinobho, thola izidakamizwa". Kuyimboni yemibono elawulwayo: khiqiza kabanzi, hlunga ngonya, ubhale phansi zonke izinqumo, bese uvumela ososayensi benze isinqumo sokugcina.

Imibuzo Evame Ukubuzwa

Iyini indima ye-AI yokukhiqiza ekutholakaleni kwezidakamizwa?

I-AI ekhiqizayo ngokuyinhloko yandisa i-funnel yomqondo ekutholakaleni kwasekuqaleni kanye nokwenza ngcono i-lead ngokuphakamisa ama-molecule afanelekayo, ukulandelana kwamaprotheni, imizila yokwenziwa, kanye nezimvo zezinto eziphilayo. Inani lincane "lokufaka esikhundleni sezilingo" kodwa "likhetha ukuhlolwa okungcono" ngokukhiqiza izinketho eziningi bese lihlunga kanzima. Isebenza kahle kakhulu njenge-accelerator ngaphakathi komsebenzi oqeqeshiwe, hhayi njengomenzi wezinqumo ozimele.

I-AI ekhiqizayo isebenza kanjani kangcono kulo lonke ipayipi lokutholwa kwezidakamizwa?

Kuvame ukuletha inani elikhulu lapho isikhala se-hypothesis sikhulu khona futhi ukuphindaphinda kubiza kakhulu, njengokuhlonza ama-hit, ukwakheka okusha, kanye nokwenza ngcono ama-lead. Amaqembu ayisebenzisela futhi i-ADMET triage, iziphakamiso ze-retrosynthesis, kanye nokusekelwa kwezincwadi noma kwe-hypothesis. Izinzuzo ezinkulu zivame ukuvela ekuhlanganiseni isizukulwane nezihlungi, amamaki, kanye nokubuyekezwa kwabantu kunokulindela ukuthi imodeli eyodwa "ihlakaniphe."

Uzibeka kanjani imingcele ukuze amamodeli akhiqizayo angakhiqizi ama-molecule angenamsebenzi?

Indlela esebenzayo ukuchaza imikhawulo ngaphambi kokwenziwa: ububanzi bezakhiwo (njengokuncibilika noma imigomo ye-logP), imithetho ye-scaffold noma yesakhiwo, izici zesayithi lokubopha, kanye nemikhawulo emisha. Bese usebenzisa izihlungi ze-chemistry yezokwelapha (kufaka phakathi amaqembu e-PAINS/reactive) kanye nokuhlolwa kokwenziwa. Isizukulwane sokuqala se-Constraint sisiza kakhulu ngomklamo wama-molecule wesitayela sokusabalalisa kanye nezinhlaka ezifana ne-REINVENT 4, lapho imigomo yezinhloso eziningi ingafakwa khona ikhodi.

Amaqembu kufanele aqinisekise kanjani imiphumela ye-GenAI ukuze agweme ukubona izinto ezingekho kanye nokuzethemba ngokweqile?

Phatha konke okukhiphayo njengombono, hhayi isiphetho, bese uqinisekisa ngezivivinyo namamodeli ahambisanayo. Hlanganisa ukwenziwa kokuhlunga okunamandla, ukufaka idokhi noma ukufaka amaphuzu lapho kufanele khona, kanye nokuhlolwa kwesizinda sokusebenza kwezibikezeli zesitayela se-QSAR. Yenza ukungaqiniseki kubonakale lapho kungenzeka, ngoba amamodeli angaba nephutha ngokuqiniseka kumakhemikhali angaphandle kokusabalalisa noma izimangalo zebhayoloji ezintengantengayo. Ukubuyekezwa komuntu ngaphakathi kwe-loop kusalokhu kuyisici esiyinhloko sokuphepha.

Ungakuvimbela kanjani ukuvuza kwedatha, ingozi ye-IP, kanye nemiphumela "egcinwe ngekhanda"?

Sebenzisa ukubusa kanye nezilawuli zokufinyelela ukuze imininingwane yohlelo ebucayi ingafakwa kalula ezicelweni, bese ubhala phansi izixwayiso/imiphumela ukuze kuhlolwe kahle. Qinisekisa ukuthi abantu abazongenela ukhetho abasondelene kakhulu nama-compounds aziwayo noma izifunda ezivikelwe. Gcina imithetho ecacile mayelana nokuthi iyiphi idatha evunyelwe ezinhlelweni zangaphandle, futhi ukhethe izindawo ezilawulwayo zomsebenzi wokuzwela okuphezulu. Ukubuyekezwa kwabantu kusiza ekubambeni iziphakamiso "ezijwayelekile kakhulu" kusenesikhathi.

Isetshenziswa kanjani i-AI yokukhiqiza ukuze kwenziwe ngcono i-lead kanye nokulungisa amapharamitha amaningi?

Ekuthuthukiseni i-lead, i-AI yokukhiqiza iwusizo ngoba ingaphakamisa izixazululo eziningi zokushintshana esikhundleni sokuphishekela inhlanganisela eyodwa "ephelele". Izindlela zokusebenza ezivamile zifaka phakathi isiphakamiso se-analog, ukuskena okuqondiswayo okufakwa esikhundleni, kanye nokushintshashintsha kwe-scaffold lapho imikhawulo ye-potency, i-tox, noma i-IP ivimba inqubekela phambili. Izibikezeli zempahla zingaba buthakathaka, ngakho-ke amaqembu avame ukulinganisa abantu abazongenela ukhetho ngamamodeli amaningi bese eqinisekisa izinketho ezinhle kakhulu ngokuhlola.

Ingabe i-AI yokukhiqiza ingasiza nge-biologics kanye nobunjiniyela bamaprotheni?

Yebo - amaqembu ayisebenzisela ukukhiqiza ukulandelana kwama-antibody, imibono yokuvuthwa kwe-affinity, ukuthuthukiswa kokuzinza, kanye nokuhlola i-enzyme noma i-peptide. Ukukhiqiza amaprotheni/ukulandelana kungabonakala kunengqondo ngaphandle kokuthi kuthuthukiswe, ngakho-ke kubalulekile ukusebenzisa izihlungi zokuthuthukiswa, i-immunogenicity, kanye ne-manufacturability. Amathuluzi esakhiwo afana ne-AlphaFold angasekela ukucabanga, kodwa "isakhiwo esinokwenzeka" asisona ubufakazi bokuveza, ukusebenza, noma ukuphepha. Ama-loop e-Wet-lab ahlala ebalulekile.

I-AI yokukhiqiza isekela kanjani ukuhlela ukwenziwa kanye nokuhlelwa kabusha?

Abahleli be-Retrosynthesis bangaphakamisa imizila, izinto zokuqala, kanye nokuhlelwa kwemizila ukuze kusheshiswe ukucabanga futhi kuqedwe ngokushesha imizila engenakwenzeka. Amathuluzi nezindlela ezifana nokuhlela kwesitayela se-AiZynthFinder zisebenza kahle kakhulu uma zihambisana nokuhlolwa kokwenzeka kwangempela okuvela kosokhemisi. Ukutholakala, ukuphepha, imikhawulo yokukhulisa, kanye "nokusabela kwamaphepha" okwehluleka ekusebenzeni kusadinga ukwahlulela komuntu. Uma kusetshenziswa ngale ndlela, konga isikhathi ngaphandle kokwenza sengathi ikhemistri ixazululiwe.

Izinkomba

1. Isibuyekezo sokutholakala kwe - Nature - Ligand (2023) - nature.com

2. I-Nature Biotechnology - GENTRL (2019) - nature.com

3. Imvelo - AlphaFold (2021) - nature.com

4. Imvelo - Ukusabalala kwe-RF (2023) - nature.com

5. I-Nature Biotechnology - I-ProteinGenerator (2024) - nature.com

6. Ukuxhumana Kwemvelo - Imiphumela yeqembu ekuthathweni kwezithombe zamaseli (2024) - nature.com

7. Imithi Yedijithali ye-npj - Ukungaboni kahle + uhlaka lokuphepha (2025) - nature.com

8. I-npj Digital Medicine - I-Multimodal ku-biotech (2025) - nature.com

9. Isayensi - IphrotheniMPNN (2022) - science.org

10. Amaphethini Eseli - Ama-LLM ekutholakaleni kwezidakamizwa (2025) - cell.com

11. I-ScienceDirect (Elsevier) - Amamodeli akhiqizayo ekwakhiweni kwezidakamizwa ezintsha (2024) - sciencedirect.com

12. I-ScienceDirect (Elsevier) - Vogt (2023): ukukhathazeka ngezinto ezintsha/ezihlukile - sciencedirect.com

13. Ukuhlaziywa Kwezithombe Zezokwelapha (ScienceDirect) - I-Multimodal AI kwezokwelapha (2025) - sciencedirect.com

14. I-PubMed Central - Umhlahlandlela wodokotela (ingozi yokubona izinto ezingekho) - nih.gov

15. Ama-Akhawunti Ocwaningo Lwekhemikhali (Izincwadi ze-ACS) - Isikhala Sekhemikhali (2015) - acs.org

16. I-PubMed Central - Irwin & Shoichet (2009): isikali sesikhala samakhemikhali - nih.gov

17. Imingcele Ekutholakaleni Kwezidakamizwa (i-PubMed Central) - Isibuyekezo (2024) - nih.gov

18. Ijenali Yolwazi Lwekhemikhali Nokumodela (Izincwadi Ze-ACS) - Amamodeli Okusabalalisa Ekwakhiweni Kwezidakamizwa Ezintsha (2024) - acs.org

19. I-PubMed Central - REINVENT 4 (uhlaka oluvulekile) - nih.gov

20. I-PubMed Central - I-ADMETlab 2.0 (i-ADMET yokuqala ibalulekile) - nih.gov

21. I-OECD - Izimiso Zokuqinisekisa Ngezinhloso Zokulawula Zamamodeli e-(Q)SAR - oecd.org

22. I-OECD - Idokhumenti yesiqondiso ngokuqinisekiswa kwamamodeli e-(Q)SAR - oecd.org

23. Ama-Akhawunti Ocwaningo Lwamakhemikhali (Izincwadi ze-ACS) - Ukuhlela ukwenziwa kwezinto ngosizo lwekhompyutha / I-CASP (Coley, 2018) - acs.org

24. Isayensi Ephakathi ye-ACS (Izincwadi ze-ACS) - Ukwenziwa kabusha kwe-retrosynthesis okusizwa yikhompyutha (Coley, 2017) - acs.org

25. I-PubMed Central - AiZynthFinder (2020) - nih.gov

26. I-PubMed - Lipinski: Umthetho womongo wesi-5 - nih.gov

27. Ijenali Yezokwelapha (Izincwadi ze-ACS) - Baell & Holloway (2010): PAINS - acs.org

28. I-PubMed - Waring (2015): ukuhoxa - nih.gov

29. I-PubMed - Rives (2021): amamodeli olimi lweprotheni - nih.gov

30. I-PubMed Central - Leek et al. (2010): imiphumela ye-batch - nih.gov

31. I-PubMed Central - Isibuyekezo Sokusabalalisa (2025) - nih.gov

32. I-FDA - E14 kanye ne-S7B: ukuhlolwa kwezokwelapha kanye nokungakwenzi ukwelashwa kwe-QT/QTc interval prolongation kanye ne-proarrhythmic potential (Q&A) - fda.gov

33. I-European Medicines Agency - Isifinyezo se-ICH isiqondiso se-E14/S7B - europa.eu

34. USENIX - Carlini et al. (2021): ukukhipha idatha yoqeqesho kumamodeli olimi - usenix.org

35. IYunivesithi yase-Edinburgh – Izinsizakalo Zocwaningo Lwedijithali - Incwajana yelebhu ye-elekthronikhi (i-ELN) - ed.ac.uk

36. I-ScienceDirect (Elsevier) - Weaver (2008): Isizinda sokusebenza se-QSAR - sciencedirect.com