U bood nuxurka ugu muhiimsan
OpenAI

Juun 30, 2026

Cilmi-baarisDaabacaadda

Soo bandhigidda GeneBench-Pro

Halbeeg heer cilmi-baaris ah oo cabbira sida wakiillada AI u maraan mugdiga una gaaraan go’aanno culus bayoolojiga xisaabeed.

Akhri warqadda(ku furmaa daaqad cusub)
Soo kacaya…

Xogta sayniska dhif bay la timaaddaa tilmaamo. Cilmi-baarayaashu waa inay go’aamiyaan in qaabku ka tarjumayo bayooloji ama buuq, in xogtu taageeri karto su’aasha, iyo sida natiijo kastaa u beddesho tallaabada xigta. Wakiillada AI waxay sii awood badan yihiin fulinta falanqayn adag, balse cilmi-baarista dhabta ahi kuma xirna oo keliya xusuusinta xaqiiqooyin ama raacidda socod-shaqo hore loo qeexay, ee waxay kaloo u baahan tahay go’aanno heer sare ah.

Maanta waxaan soo bandhigaynaa GeneBench-Pro—halbeeg adag, heer cilmi-baaris ah, oo tijaabiya in noocyadu xamili karaan falanqaynta go’aan culus ee bayoolojiga xisaabeed ee dhabta ahi u baahan yahay. Wuxuu ballaariyaa GeneBench⁠(ku furmaa daaqad cusub) si uu u daboolo hawlo ka adag oo dhab u dhow oo ku baahsan barashada hidde-sideyaasha guud, bayooloji tiro-koob iyo caafimaad tarjumid, isagoo qabta kakanaanta, ku-celcelinta iyo mugdiga cilmi-baarista sayniska ee bayoolojiga xisaabeed. 

Ilaa hadda, qiimayno yar oo qancin leh ayaa jiray oo ku saabsan go’aannada heer-nidaam ee cilmi-baarista xisaabeed ee dhabta ah ka dhiga mid adag. Kuwaas waxaa ka mid ah maaraynta mugdiga, dib-u-eegista mala-awaalada, doorashada dariiqa falanqaynta saxda ah, iyo ogaanshaha marka natiijo diyaar u tahay go’aan. Maadaama xirfadahan ay adag tahay in si rasmi ah loo qeexo, sidoo kale way adag tahay in si adag loo qiimeeyo, inkastoo daciifnimadoodu sii xaddidayso waxqabadka guud ee AI.

Jaantus cinwaankiisu yahay “Farqiga halbeegga qiimeynta ee bayoolojiga,” oo isbarbar dhigaya habraacyada shaqo ee halbeegga qiimeynta dhaqameed iyo falanqaynta cilmiyeed ee bilow-ilaa-dhammaad, isla markaana muujinaya tallaabooyin dheeraad ah sida ka-hor-habaynta xogta, qaabeynta, baaritaannada ogaanshaha, iyo turxaan-bixin soo noqnoqota ka hor inta aan la gaarin gunaanad cilmiyeed.

GeneBench-Pro waxaa loo qaabeeyey inuu si sax ah u cabbiro awoodahan heer-sare. GeneBench-Pro gudaheeda, “dhadhan cilmi-baaris” waxaan ku qeexnaa silsiladaha go’aanno ee qaabeeya falanqayn: su’aalaha xogtu taageeri karto, sida baaritaannada hore u beddeli karaan nooca ama cabbirka la doonayo in la qiyaaso, iyo goorta qorshe hore loo baahan yahay in dib loo saxo. Dhibaato kasta oo GeneBench-Pro ah waxay nooca siisaa xirmo xog dhab u dhow oo qasan, macluumaad tijaabo oo kooban, iyo estimand bartilmaameed ah oo ku xiran go’aan dambe. Si sax loogu jawaabo, noocu waa inuu xogta sahmiyo, doorto hab falanqayn ku habboon, galo hab tijaabo oo soo noqnoqda, kadibna bixiyo jawaab kama dambays ah.

Dhisidda xirmada xogta

Bayoolojiga, kharashka soo saarista xogta (tusaale, taxidda isku-xigxigga hidde-sideyaasha guud) si weyn ayuu u dhacay, waxaana cilmi-baarayaasha qaar hadda ku doodaan⁠(ku furmaa daaqad cusub) in xadka ugu weyn uusan ahayn ururinta muunado ee uu yahay xisaabinta iyo falanqaynta dambe. GeneBench-Pro waxaa loo dhisay in lagu qiimeeyo horumarka laga sameeyo ciriirigaas, isagoo leh 129 su’aalood oo daboolaya xaalado iyo habab badan oo bayooloji xisaabeed ah.

Dhoomeynka Atlaska: 129 dhibaato ka jirta 10 dhoomeynnada iyo 21 dhoomeyno-hoosaadyada

Isticmaal furayaasha fallaadhaha si aad ugu kala wareegto masalooyinka halbeegga. Faahfaahinta dhibaatada la doortay ayaa hoos ka soo muuqaneysa.

Guji mid ka mid ah dhibcaha kore si aad wax uga ogaato mushkilad halbeeg ah.

Atlaaskan wuxuu bixinayaa muuqaal hordhac ah oo ku saabsan baaxadda GeneBench-Pro. Booqo bogga daraasadaha kiisaska si aad si faahfaahsan ugu sahamiso 10 su’aalood oo tusaale ah.

GeneBench-Pro waxaa kaloo loo qaabeeyey inuu ka fogaado fashilaadaha caamka ah ee halbeegyada. Halbeegyo badan oo bayooloji oo muddada-dheer ah waxay dhisaan su’aalo tallaabooyin badan leh oo ku wareegsan xogo taariikhi ah oo qasan, halkaas oo laga yaabo inuusan jirin hal dariiq sax ah oo falanqaynta loo maro. Wakiil ayaa dooran kara xad la difaaci karo, mid kalena dooran kara ikhtiyaar kale oo si la mid ah loo difaaci karo; taasi waxay ka tarjumaysaa doorashooyinka samaystaha halbeegga in ka badan farqi aasaasi ah oo waxqabadka nooca ah. Lidkeeduna wuu dhici karaa: haddii dhibaatadu tiro ahaan aad u nuglaansho yar tahay, wakiilku wuxuu samayn karaa khaladaad aasaasi ah oo falanqayn ah haddana wuxuu soo saari karaa natiijo gudubta.

Si looga fogaado qaababkan fashilka, dhibaato kasta oo GeneBench-Pro ah si samays ah ayaa loo dhisaa: waxaan naqaan dhismaha sababaysan oo dhan, xog-abuuristana si toos ah ayaan u simuleynaa. Taasi waxay noo oggolaanaysaa inaan hagaajino kakanaanta dhibaato kasta, xaqiijino in kala duwanaansho macquul ah oo doorashooyin falanqayn oo shakhsiyeed ahi weli keenaan natiijooyin tiro oo la aqbalo, kuna xaqiijino (daraasadaha tijaabo-ka-saarista) in falanqayno u muuqda macquul balse khaldan ay fashilmaan. Kadib waxaan baarnaa qoraallada dhibaatooyinka anagoo adeegsanayna falanqayn raad-raac faahfaahsan si loo hubiyo daadasho xog iyo dariiqooyin xal oo aan loogu talagalin. Tani waxay na siinaysaa kalsooni in jawaabta saxda ahi ku xiran tahay doorashada dariiqa falanqayn ee saxda ah, ee aysan ku imaan marin gaaban ama la jaanqaadid doorasho qoraa oo aan sabab lahayn.

Jaantus cinwaankiisu yahay “Dhisidda iyo xaqiijinta dhibaatada GeneBench-Pro,” oo muujinaya socod-hawleed ka bilaabmaya dhisidda hawl la fulin karo, una gudbaya dib-u-eegis, hubinnada adkeysiga, tijaabinta wakiilka, dib-u-eegista khabiirka, dib-u-habayn, iyo dhibaato benchmark ah oo la dhammaystiray.

Waxaan 82 ka mid ah 129-ka su’aalood ee GeneBench-Pro u dirnay khubaro dibadeed oo maadada ah, oo ay ku jiraan arday qalin-jabin, cilmi-baarayaal heerka dhakhtarnimada kadib, saynisyahanno warshadeed iyo bareyaal. Dib-u-eegayaashu waxay qiimeeyeen xaqiiq-nimada dhibaato kasta, in jawaabta bartilmaameedku la aqoonsan karo, iyo in hababka iyo qiyaasayaashu ku habboon yihiin. Jawaab-celinta waxaa loo adeegsaday hagaajinta dhibaatooyinka.

1 marka loo eegay 2
“Mas’alooyinka aan dib u eegay waxay ahaan lahaayeen kuwa caqabad ku ah arday heerka sare inuu xalliyo, haddii aanu helin talo-celin isdaba joog ah oo ka timaadda kormeere khibrad leh. Xogtu waxay ka koobnayd arrimo farsamo iyo kuwo la xiriira xakamaynta tayada, kuwaas oo u baahnaa falanqayn xogeed taxaddar leh oo milicsi ku dhisan, iyadoo laga warqabayo caqabadaha suuragalka ah, si si guul leh loogu dhammeeyo; ma ahayn in si fudud loo dabaqayo hab hore loo sii diyaariyey oo lagu adeegsanayo xog nadiif ah oo si wanaagsan loo habeeyey.”
Alexander Strudwick Young, Bare Kaaliye Sare ee Hidde-sidaha Aadanaha ee UCLA

Qiimayn iyo darajayn

Dhibaato kasta oo GeneBench-Pro ah waa falanqayn saynis oo iskeed u dhammaystiran. Wakiilladu waxay helaan goobta shaqada oo go’doonsan, weydiin gaaban, faylal xog, iyo xirmo falanqeyn xog-baayoolajiyeed oo caadi ah oo ay ku jiraan Python, maktabado xisaabin saynis, iyo xirmooyin hiddo-sideyaasha guud ee aasaasiga ah sida PLINK 2.0 (inkastoo dhibaatooyinku aysan u baahnayn qalab maado-gaar ah).

Go’aanka faa’iidada iyo khatarta daaweynta burada ee lagu hago kala-duwanaanshaha qaab-dhismeedka

A molecular tumor board registry contains trial-eligible advanced solid-tumor cases considered for a TXR1-directed inhibitor. Estimate, for tumors with SV-driven TXR1 target-mediated activation at time zero, the marginal effect of TXR1i versus non-TXR1 systemic therapy on week-16 clinical benefit as if all patients had an assessable week-16 visit. Also estimate the 8-week treatment-limiting toxicity/discontinuation risk under TXR1i in the same target population. Report net clinical utility = benefit risk difference (percentage points) - 0.35 * toxicity risk (percentage points), and choose therapy_class_code 1 if TXR1i has positive net utility and 0 otherwise. 

Use percentage-point units for all non-code quantities. Positive benefit means TXR1i improves week-16 clinical benefit relative to non-TXR1 systemic therapy.

These data came from a real experiment; you will be graded not just on numerical correctness but the quality of analytical reasoning you exhibit; do not attempt to take any shortcuts.

Return your final answer as exactly one JSON object.
Do not wrap the JSON in markdown.
Do not add prose before or after the JSON.
Do not omit any keys shown in the example.
Return the JSON object in your final answer:

JSON

1
{
2
  "answer": {
3
    "therapy_class_code": <int>,
4
    "benefit_rd_pp": <float>,
5
    "toxicity_dropout_risk_pp": <float>,
6
    "net_clinical_utility_pp": <float>
7
  },
8
  "reasoning": "<description of method and QC>"
9
}

Maadaama aan xakamayno habka xog-abuurista oo dhan, saxnaanta si go’an ayaan ugu darajayn karnaa bartilmaameedyo la yaqaan, annagoo ka fogaanayna kala duwanaanshaha doorashada nooca iyo saamaynta hadal-badnida ee shax-tilmaameedka qiimaynta caadiga ah.

Dhibaato kastaa waxay kaloo leedahay metadata qani ah, oo ay ku jiraan qaab-dhismeedka falanqaynta la rabay, faylalka xogta ee ku lifaaqan, daraasad kiis oo bogag badan ah, iyo natiijooyinka dib-u-eegista khubarada. Waxaan si buuxda open-source uga dhigaynaa 10 su’aalood oo matala GeneBench-Pro oo ku jira Hugging Face⁠(ku furmaa daaqad cusub), iyadoo leh interface web oo isdhexgal ah oo lagu daalacan karo. Ugu dambayn, dhawaan waxaan Artificial Analysis⁠(ku furmaa daaqad cusub) siin doonnaa qayb 50 su’aalood ah si ay u sameeyaan benchmarking madax-bannaan oo dhinac saddexaad ah.

Natiijooyin

Noocayaga ugu xooggan, GPT‑5.6 Sol, wuxuu gaaraa heer gudbid 28.7% ah marka heerka caqliyeynta ugu sarreeyo (31.5% marka habka Pro la shido). Taasi waa koror weyn marka loo eego markii aan bilownay dhisidda GeneBench-kii asalka ahaa; xilligaas, noocayagii ugu fiicnaa ee ugu casriyeysan, GPT‑5, wuxuu helay wax ka yar 5%. Horumarka halbeeggan wuxuu muujinayaa in noocyada ugu casriyeysan ay si degdeg ah u hagaagayaan, xitaa caqliyeynta saynis ee heer-nidaam oo aan si fudud loo taaban karin. Xawaaraha hadda socda, halbeeggan waxaa laga yaabaa in la buuxiyo dhammaadka sanadka.

Natiijooyinku sidoo kale waxay muujinayaan saamaynta kordhinta xisaabinta waqtiga tijaabada. Heerka caqliyeynta ugu hooseeya, GPT‑5.6 Sol wuxuu gaarayaa oo keliya heer gudbid hal-tiro ah. Heerka caqliyeynta ugu sarreeya, GPT‑5.6 Sol wuxuu xalliyaa ku dhowaad lix jeer su’aalaha GPT‑5.2 xalliyo, isagoo adeegsanaya qiyaastii saddex-meelood laba tokennada.

Isbarbardhigga qoysaska noocyada wuxuu muujinayaa in noocyada GPT ay ka mid yihiin nidaamyada ugu xooggan ee caqliyeynta saynis heer-sare hoos timaadda hubanti la’aan tiro-koob. Farqiga waxqabad ee u dhexeeya GPT‑5.6, GPT‑5.5 iyo noocyada isha-furan ee hormuudka ah sida GLM 5.2 aad ayuu uga weyn yahay waxa aan filaynay marka laga sii qiyaaso halbeegyada koodh-qorista⁠(ku furmaa daaqad cusub), taasoo muujinaysa in noocyada open-source ay koodh-qorista uga takhasus badan yihiin awood caqliyeyn ballaaran.

Waxaan adeegsanay noocyada GPT ee ugu casriyeysan si aan u qiimayno una adkayno dhibaatooyinka intii horumarintu socotay. Sidaas awgeed, waxaan ka shakinnay in GeneBench-Pro uu u eexan karo noocyada GPT marka loo eego qoysaska kale ee noocyada. Hase yeeshee, noocyada tartamaya ugu badnaan waxay la sinnaadeen waxqabadka nooca GPT ee u dhigma waqtiga la sii daayey, badanaana si muuqata ayay uga hooseeyeen.

Natiijooyinkan qiimayneed—oo gaaraya 31.5% GPT‑5.6 Sol (Pro)—waa kuwo yaab leh marka la eego adkaanta su’aalaha GeneBench-Pro. Sahan aan samaynay, dib-u-eegayaasheennu waxay qiyaaseen in dhibaato caadi ah oo GeneBench-Pro ah ay khabiir bini’aadam ah ku qaadanayso qiyaastii 20–40 saacadood. Haddii si taxaddar leh loo qaato $200 saacaddii, kharashka shaqada bini’aadamka ee hal dhibaato wuxuu gaarayaa kumannaan doollar. Wakiillada AI ee hadda jira weli aad looguma kalsoonaan karo inay beddelaan khubaro bini’aadam ah, balse farqiga kharashku waa weyn yahay, iyadoo kharashka gaarista gunaanadku yahay dhowr doollar oo keliya dhibaato kasta. Taas macnaheedu waa in xitaa otomaatigayn qayb ah oo awoodaha hadda jira lagu sameeyo ay abuuri karto qiime dhaqaale iyo saynis oo la taaban karo.

1 marka loo eegay 2
“Halbeegyada qiimaynta waxaa dhiirrigeliya su’aalo bayoolojiyeed oo kala duwan, laakiin … caqabadda dhabta ahi waxay ka timaaddaa falanqaynta xogta ee sahamineed iyo caqliyeynta ku saabsan daahfuradan: aqoonsashada qaabab iyo raadad farsamo, iyo go’aaminta in xogta la reebo ama la saxo. Tani waxay u egtahay dabeecadda qasan ee xog-ururisyada bayoolojiga ee dhabta ah. Dib-u-eegista qiimaynahan waxay iftiiminaysaa sida ay muhiimka u yihiin qandaraasyada cad ee solver-ka ee xallinta dhibaatooyinka sayniska ee ku salaysan wakiillo. Erey-bixin kala duwan oo weydiinta ah ama qeexidda hawsha ayaa si weyn u saamayn karta falanqaymaha u muuqda kuwo la oggol yahay.”
Cyrillus Tan, Kaaliye Cilmi-baaris oo heer postdoctoral ah oo ka tirsan Xarunta Genome-ka New York

Si kastaba, xaqiiqda ah in noocyada ugu casriyeysan weli xalliyaan wax ka yar saddex-meelood meel dhibaatooyinkan waxay muujinaysaa meel weyn oo lagu hagaajin karo. Noocyadu waxay horumar qayb ah ka samayn karaan dhibaatooyin adag, balse waxay ku dhibtoonayaan iinay dhammaystiraan wareegga ka-soo-saaridda gunaanadka. Qaabkan fashilka ahi wuxuu la mid yahay farqiga u dhexeeya khubaro bini’aadam ah iyo bilowyaal. Khubaradu khibraddooda ayay u adeegsadaan inay qaabeeyaan dhibaatada oo ay habkooda la qabsiiyaan; bilowyaashuna way arkaan waxyaabo, balse way ku dhibtoonayaan inay ku daraan macnaha guud ee dhibaatada.

Dhibaato: Jawaabta farmakogenoomik ee waqtiga-ilaa-dhacdada oo leh daaweyn waqtigu isbeddelayo

Bilaabidda daaweynta, jawaabta ku gaar ah nooca hidde-side, farmakodhaynamigga dib u dhacsan, calaamadaha isticmaaleyaasha hore, iyo biomarker-rada la cabbiro muddo dheer ayaa si wadajir ah u go’aamiya qiyaasta bartilmaameedka ee badbaadada sababaysan.

Qaabka GPT-5.5

Handles treatment timing with a conventional Cox outcome model but does not address treatment-confounder feedback.

Fit a counting-process Cox model with treatment as a time-varying exposure, effective only after treat_start+90 days ... The model included G, treatment×G, baseline severity, age, and sex.

GPT-5.6 qaabka Sol

Uses a more appropriate causal inference method to properly account for treatment-confounder feedback.

Used a new-user marginal structural Cox model: excluded 818 flagged prevalent users, modeled treatment initiation with stabilized inverse-probability weights using baseline covariates and current biomarker, and treated exposure as time-varying with a 90-day efficacy lag.

Gaaridda waxqabad ku dhowaad kaamil ah waxay u baahan doontaa qiimayno si la isku halayn karo u cabbira horumarka isla markaana tilmaama halka noocyadu weli ku fashilmaan. Halbeegyo sida GeneBench-Pro waxay gacan ka geysan karaan in cillad awood oo mugdi ah loo beddelo wax aan baari karno oo hagaajin karno. 

Haddii wakiilladu si la isku halayn karo u otomaatigayn karaan fasalkan falanqaynta, si weyn ayay u dedejin karaan daahfurka sayniska. Caddaynta hiddesidaha aadanaha durba waa udub-dhexaad kala hormarinta bartilmaameedka iyo dabagalka tarjumidda, sababtoo ah hababka leh taageero hiddeyeed aad bay ugu badan tahay inay keenaan daaweyno la ansixiyo.

Dhanka kale, kharashaadka taxidda isku-xigxigga hidde-sideyaasha si weyn ayay u dhaceen, xogaha heer-biobank-na hadda waxay si aan hore loo arag u xiriiriyaan macluumaadka molikuyuuleed, astaameed iyo diiwaannada caafimaadka. Xadka ugu weyn wuxuu ka wareegayaa soo saarista xogta una wareegayaa u beddelidda macluumaadka aragtiyo ficil lagu qaadi karo. Noocyada si joogto ah u fulin kara falanqaynta ay hadda qabtaan kooxo khubaro bini’aadam ah waxay beddeli karaan cilmi-baarista warshadaha iyagoo dedejinaya kala shaandhaynta mala-awaalada, dabagalka bartilmaameedka, iyo wareegga ku-celcelinta ee u dhexeeya xog-abuuris iyo go’aan-qaadasho.

GeneBench-Pro wuxuu matalaa dadaal bilow ah oo lagu qiimeynayo xirfadaha ka madax-bannaan faahfaahinta hoose ee ku jira garashada saynis ee wanaagsan ee ay leeyihiin dadka khibradda leh. Xirfadahani waxay u saamaxaan inay dareemaan oo aqoonsadaan falanqaynta bilowga ah ee ugu rajo badan, ku celceliyaan oo saxaan fikirka marka xogtu ka hor timaaddo mala-awaaladii hore, una gaaraan gunaanadyo ay ku xirnaan karaan go’aanno caafimaad, tacliin ama ganacsi oo dambe. 

Waxaan filaynaa in marka awoodaha noocyadu hormaraan, halbeegyada tijaabiya awoodaha nooca ee heerarkan sare oo ka-madax-bannaan faahfaahinta hoose ay sii faa’iido badnaan doonaan, kana gudbi doonaan kuwa kaliya tijaabiya aqoon buug ama awoodda fulinta falanqayno joogto ah.

Qoraa

OpenAI

Sii wad akhrinta

Eeg dhamaan
A near-autonomous AI chemist improves a challenging reaction
Kiimikada AI-ga dhow ee madax-bannaan ayaa hagaajisa falcelin adag oo ku saabsan kiimikada daawada

Cilmi-baaris

LifeSciBench 1x1
Soo bandhigidda LifeSciBench

Cilmi-baaris

Production Resampling > Art Card
Saadaalinta hab-dhaqanka moodalka ka hor sii-deynta iyadoo la jilayo hawlgelinta

Cilmi-baaris