Back to the fu­ture: From tar­get-se­lec­tive drug de­sign to AI-en­abled polyphar­ma­col­o­gy

CNS drug dis­cov­ery is be­ing thrown back to the 50s, when the drugs’ phe­no­typ­ic pro­files alone led to the first block­buster drugs for men­tal dis­or­ders. Psy­choGen­ics’ SmartCube® plat­form al­lows us to use AI on a jour­ney back to phe­no­typ­ic drug dis­cov­ery.

AI in drug dis­cov­ery and de­vel­op­ment

AI ap­pli­ca­tions range from tar­get dis­cov­ery, tar­get-based drug de­sign, tox­i­c­i­ty pre­dic­tions, man­u­fac­tur­ing, and clin­i­cal tri­als, to name a few. Some AI-based tar­get-dri­ven drug dis­cov­ery ef­forts fo­cus­ing on known pro­teins have proven suc­cess­ful, sup­port­ed by the ex­is­tence of good qual­i­ty da­ta [1, 2]. For com­plex CNS pathol­o­gy, how­ev­er, the tar­get is ei­ther un­known or there is no unique tar­get that can pro­vide the nec­es­sary re­lief. At Psy­choGen­ics, we tack­le CNS com­plex­i­ty head on, pre­dict­ing ef­fi­ca­cy for nov­el polyphar­ma­co­log­i­cal, un­known mech­a­nism of ac­tion, or tar­get-dri­ven com­pounds, us­ing a tar­get-ag­nos­tic  Al-en­abled drug dis­cov­ery plat­form.

CNS tar­get-se­lec­tive drug de­sign fails. Com­plex liv­ing sys­tems func­tion is sup­port­ed by the in­ter­ac­tion of mul­ti­ple over­lap­ping and in­ter­act­ing path­ways, which achieve home­osta­sis through re­dun­dan­cy and feed­back loops. Thus, mod­u­lat­ing a sin­gle tar­get can trig­ger com­pen­sato­ry mech­a­nisms by oth­er path­ways, ren­der­ing the treat­ment in­ef­fec­tive. An un­jus­ti­fied de­sire to cre­ate mol­e­cules with the high­est sin­gle-tar­get se­lec­tiv­i­ty, com­pound­ed with the hu­man ten­den­cy to sim­pli­fy com­plex sys­tems, de­sire to un­der­stand in­di­vid­ual dis­ease el­e­ments, and ig­no­rance of phys­i­ol­o­gy and CNS func­tion has led to fail­ures in the clin­ic[3].  In­deed, of 1,144 FDA-ap­proved small mol­e­cule drugs, on­ly 123 (~11%) were dis­cov­ered by tar­get-based drug dis­cov­ery and 1,021 (~89%) by phe­no­type-based drug dis­cov­ery ap­proach­es[4]. CNS dis­or­ders present with mul­ti­ple risk fac­tors and are un­like­ly to arise from a sin­gle el­e­ment, and there­fore phe­no­typ­ic drug dis­cov­ery for polyphar­ma­co­log­i­cal com­pounds can lead to nov­el and ef­fec­tive treat­ments. Here we elab­o­rate on the way AI-en­abled plat­forms can be used for tar­get-ag­nos­tic de no­vo drug dis­cov­ery.

AI-en­abled phe­no­typ­ic drug dis­cov­ery

Where­as drugs for men­tal dis­or­ders have been used around the world for cen­turies (e.g. re­ser­pine and psilocin), West­ern psy­chophar­ma­col­o­gy was rev­o­lu­tion­ized in the 50s as sci­en­tists ob­served that non-psy­chi­atric drugs had ben­e­fi­cial ef­fects on psy­chi­atric symp­toms[5]. The dis­cov­ery of chlor­pro­mazine, and soon there­after imipramine, opened the road to the syn­the­sis of nov­el com­pounds and phar­ma­co­log­i­cal treat­ments for men­tal ill­ness­es[6]. These drugs are an ex­am­ple of com­plex polyphar­ma­col­o­gy, as they act at many tar­gets in the CNS (dopamin­er­gic, sero­ton­er­gic, etc), dis­cov­ered in a tar­get-ag­nos­tic phe­no­typ­ic man­ner.

Priv­i­leged struc­tures for polyphar­ma­col­o­gy. Through tri­al-and-er­ror, nat­ur­al se­lec­tion leads to vari­a­tions in the char­ac­ter­is­tics of re­cep­tors and neu­ro­trans­mit­ters, giv­ing on­to to the cre­ation of priv­i­leged scaf­folds, ba­sic core struc­tures com­mon to a se­ries of bi­o­log­i­cal­ly ac­tive mol­e­cules. As an ex­am­ple, take the in­dole scaf­fold, present in nat­ur­al prod­ucts and CNS mol­e­cules (e.g., re­ser­pine, physostig­mine,  psilo­cyn, trypt­a­mine, sero­tonin, mela­tonin, harmine, and DMT) and count­less phar­ma­co­log­i­cal com­pounds (e.g., molin­done, suma­trip­tan, on­dansetron, pin­dolol, ziprasi­done, risperi­done, and LSD)[7, 8]. Priv­i­leged scaf­folds serve, there­fore, as the ba­sis of polyphar­ma­co­log­i­cal se­ries of com­pounds with dif­fer­ent func­tions, thanks to, some­times, sub­tle changes or ad­di­tions to the core struc­ture.

AI-en­abled phe­no­typ­ic ap­proach. De­sign­ing com­pounds in sil­i­co sole­ly based on struc­ture and lig­and da­ta is dif­fi­cult[9], but be­comes a near im­pos­si­ble feat when polyphar­ma­col­o­gy is the goal, as the op­ti­miza­tion of PK/PD re­la­tion­ships at sev­er­al tar­gets at once can­not be achieved. This can be solved by a com­pre­hen­sive in vi­vo screen sen­si­tive to polyphar­ma­col­o­gy, en­abling rapid pre­dic­tions about ther­a­peu­tic ef­fi­ca­cy. This strat­e­gy is cap­tured in a few avail­able phe­no­typ­ic as­says, of­ten us­ing ze­brafish, C. el­e­gans, or oth­er “sim­ple” or­gan­isms. Psy­choGen­ics us­es the more com­plex be­hav­iors of mice in the SmartCube® plat­form, to cap­ture drugs’ ther­a­peu­tic pro­files for de­pres­sion, psy­chosis, and oth­er in­di­ca­tions.

Psy­choGen­ics’ SmartCube® be­hav­ioral screen­ing plat­form[10] is based on ma­chine learn­ing clas­si­fi­ca­tion of com­pounds’ be­hav­ioral pro­files in mice cap­tured through 3D com­put­er vi­sion. Ma­chine learn­ing clas­si­fiers are trained us­ing mar­ket­ed drugs to de­fine ther­a­peu­tic pro­files (an­tipsy­chotics, anx­i­olyt­ics, anal­gesics, mood sta­bi­liz­ers, treat­ment-re­sis­tant de­pres­sion an­ti­de­pres­sants, and oth­ers), in a tar­get-ag­nos­tic man­ner. The plat­form can be used to char­ac­ter­ize a nov­el com­pound, per­form struc­ture-ac­tiv­i­ty re­la­tion­ship, test a can­di­date in an an­i­mal mod­el of dis­ease, or com­pare po­ten­cy of dif­fer­ent enan­tiomers, to name a few ap­pli­ca­tions.

Screen­ing com­pounds and li­braries us­ing SmartCube®

We have screened thou­sands of com­pounds from com­mer­cial or part­nered li­braries, in­clud­ing 489 with in­dole and oxin­dole cores. Where­as a di­verse Lip­in­sky-com­pli­ant small mol­e­cules li­brary has an ac­tiv­i­ty rate be­tween 30 and 50%, choos­ing these scaf­folds and fo­cus­ing on com­pounds with low mol­e­c­u­lar weight brings the ac­tiv­i­ty rate high­er than 80%. His­tor­i­cal­ly, we have shown that de­sign of com­pounds around a priv­i­leged struc­ture com­bined with our rapid phe­no­typ­ic ther­a­peu­tic pre­dic­tion can de­liv­er a polyphar­ma­co­log­i­cal de­vel­op­ment can­di­date in few­er than 300 com­pounds. This ap­proach to screen­ing and lead op­ti­miza­tion de­liv­ers a can­di­date at a frac­tion of the cost and time as­so­ci­at­ed with tar­get­based dis­cov­ery.

From SmartCube® to the clin­ic. Our most ad­vanced com­pound, de­vel­oped in col­lab­o­ra­tion with Sum­it­o­mo Phar­ma, is Ulotaront, an an­tipsy­chot­ic with ag­o­nis­tic ac­tion at the 5HT1A and TAAR1 re­cep­tors, D2 spar­ing[11] dis­cov­ered em­pir­i­cal­ly us­ing SmartCube®. Ulotaront is in sev­er­al Phase II and III clin­i­cal tri­als, af­ter promis­ing re­sults in Phase II tri­als lead­ing to a Break­through Ther­a­py Des­ig­na­tion from the US FDA for the treat­ment of schiz­o­phre­nia. Specif­i­cal­ly, ulotaront was ef­fec­tive in coun­ter­ing pos­i­tive and neg­a­tive symp­toms of schiz­o­phre­nia with place­bo-like side ef­fects, all pre­dict­ed pre­clin­i­cal­ly.

New op­por­tu­ni­ties part­ner­ing with Psy­choGen­ics

To­day’s in­no­v­a­tive en­vi­ron­ment presents nu­mer­ous part­ner­ing op­por­tu­ni­ties where Psy­choGen­ics’ SmartCube and oth­er AI-en­abled plat­forms can be uti­lized to iden­ti­fy com­pounds with po­ten­tial to treat neu­ropsy­chi­atric dis­or­ders. SmartCube® can of course al­so be used to screen tar­get-based de­sign­er drugs, and, not on­ly small mol­e­cules, but al­so small pep­tides. Us­ing SmartCube® for Psylin, a joint ven­ture with Amylin Phar­ma­ceu­ti­cals, we screened a li­brary of small pep­tides lead­ing to the dis­cov­ery of sev­er­al pep­tides, in­clud­ing a GLP-1 ag­o­nist with a mixed anx­i­olyt­ic and an­ti­de­pres­sant po­ten­tial in ad­di­tion to cog­ni­tive en­hanc­ing ac­tiv­i­ty and the ex­pect­ed weight-loss ef­fect. ​ New po­ten­tial li­braries in­clude PRO­TACs, mol­e­c­u­lar glues, com­bi­na­tions, ion chan­nels, mem­brane sta­bi­liz­ers​, in­trin­si­cal­ly dis­or­dered pro­teins, bio­mol­e­c­u­lar con­den­sates, Mas-re­lat­ed GPRCs, and mi­to­chon­dria​l, glym­phat­ic sys­tem​, neu­roin­flam­ma­to­ry, and gut-brain ax­is tar­gets​.

Psy­choGen­ics’ pro­pos­al. The com­plex in­ter­play among mul­ti­ple neu­ro­trans­mit­ter sys­tems, par­tic­u­lar­ly in the con­text of the patho­phys­i­ol­o­gy of CNS dis­eases where the un­der­ly­ing bi­o­log­i­cal mech­a­nisms are of­ten poor­ly un­der­stood, presents a for­mi­da­ble chal­lenge for to the de­vel­op­ment of new drugs for psy­chi­atric dis­or­ders. The most ef­fec­tive drugs used in the clin­ic typ­i­cal­ly ex­hib­it polyphar­ma­col­o­gy, mod­u­lat­ing mul­ti­ple tar­gets si­mul­ta­ne­ous­ly. Psy­choGen­ics plat­forms can as­sist in sev­er­al as­pects of the drug dis­cov­ery process, from com­pound screen­ing, SAR, and lead op­ti­miza­tion to fa­cil­i­tate the dis­cov­ery of nov­el polyphar­ma­co­log­i­cal treat­ments.

To learn more about Psy­choGen­ics’ com­mit­ment to drug dis­cov­ery in­no­va­tion vis­it https://www.psy­chogen­ics.com.

Ref­er­ences

1. Ke­shavarzi Ar­sha­di, A., et al., Ar­ti­fi­cial in­tel­li­gence for COVID-19 drug dis­cov­ery and vac­cine de­vel­op­ment. Fron­tiers in Ar­ti­fi­cial In­tel­li­gence, 2020. 3: p. 65.
2. Chávez-Hernán­dez, A.L., E. López-López, and J.L. Med­i­na-Fran­co, Yin-yang in drug dis­cov­ery: re­think­ing de no­vo de­sign and de­vel­op­ment of pre­dic­tive mod­els. Fron­tiers in Drug Dis­cov­ery, 2023. 3: p. 1222655.
3. Ste­fan, S.M. and M. Rafe­hi, Med­i­c­i­nal polyphar­ma­col­o­gy: Ex­plo­ration and ex­ploita­tion of the polyphar­ma­colome in mod­ern drug de­vel­op­ment. Drug Dev Res, 2024. 85(1): p. e22125.
4. Sadri, A., Is tar­get-based drug dis­cov­ery ef­fi­cient? Dis­cov­ery and “off-tar­get” mech­a­nisms of all drugs. Jour­nal of Med­i­c­i­nal Chem­istry, 2023. 66(18): p. 12651-12677.
5. Lopez-Munoz, F., et al., His­to­ry of the dis­cov­ery and clin­i­cal in­tro­duc­tion of chlor­pro­mazine. Ann Clin Psy­chi­a­try, 2005. 17(3): p. 113-35.
6. Chow, W.S. and S. Priebe, Un­der­stand­ing psy­chi­atric in­sti­tu­tion­al­iza­tion: a con­cep­tu­al re­view. BMC Psy­chi­a­try, 2013. 13: p. 169.
7. de Sa Alves, F.R., E.J. Bar­reiro, and C.A. Manssour Fra­ga, From na­ture to drug dis­cov­ery: the in­dole scaf­fold as a ‘priv­i­leged struc­ture’. Mi­ni re­views in med­i­c­i­nal chem­istry, 2009. 9(7): p. 782-793.
8. Stem­pel, E. and T. Gaich, Cy­clo­hep­ta [b] in­doles: a priv­i­leged struc­ture mo­tif in nat­ur­al prod­ucts and drug de­sign. Ac­counts of chem­i­cal re­search, 2016. 49(11): p. 2390-2402.
9. Ste­fan, S.M. and M. Rafe­hi, Med­i­c­i­nal polyphar­ma­col­o­gy: Ex­plo­ration and ex­ploita­tion of the polyphar­ma­colome in mod­ern drug de­vel­op­ment. Drug De­vel­op­ment Re­search, 2024. 85(1): p. e22125.
10. Brun­ner, D., Gond­halekar, V., Leahy, E., LaRose, D., Ross, W.P., Method for pre­dict­ing treat­ment class­es us­ing an­i­mal be­hav­ior in­for­mat­ics, USP­TO, Ed­i­tor. 2009, Carnegie Mel­lon, Uni­ver­si­ty Psy­chogen­ics Inc: U.S.
11. Dedic, N., et al., SEP-363856, a Nov­el Psy­chotrop­ic Agent with a Unique, Non-D(2) Re­cep­tor Mech­a­nism of Ac­tion. J Phar­ma­col Exp Ther, 2019. 371(1): p. 1-14.