Can AI replace animal testing?

Arti­fi­cial intel­li­gence (AI) algorithms have proved to be highly effect­ive when it comes to sim­u­lat­ing human voices or the pro­duc­tion of images. But are they also cap­able of sim­u­lat­ing liv­ing beings well enough to make it pos­sible to dis­pense with anim­al test­ing? The ques­tion has emerged in the face of grow­ing con­cern around anim­al wel­fare. “We no longer think of anim­als as machines, and our soci­ety believes that humans are respons­ible for pro­tect­ing them. In addi­tion advances in eth­o­logy and anim­al psy­cho­logy, and the emer­gence of con­cepts such as anim­al cul­ture. All these factors are call­ing anim­al test­ing into ques­tion,” explains Jean-Michel Besni­er, philo­soph­er of sci­ence and pro­fess­or emer­it­us of philo­sophy at Sor­bonne Uni­versité. This soci­et­al con­cern coin­cides with a grow­ing aware­ness of anoth­er issue. “It is not that easy to draw con­clu­sions about humans from mice… What is the point of mak­ing anim­als suf­fer for res­ults that are open to question?”

Anim­als are used in research for sev­er­al pur­poses: to study and under­stand bio­lo­gic­al phe­nom­ena as part of fun­da­ment­al research, and to check the safety and effic­acy of a product or drug as part of reg­u­lat­ory and tox­ic­o­lo­gic­al research. This issue has not been over­looked by the sci­entif­ic com­munity. The 2010 European Dir­ect­ive (2010/63/EU) provides a frame­work for anim­al test­ing through the three Rs: Replace­ment, Reduc­tion and Refine­ment, which aims to reduce the suf­fer­ing inflicted.

Can arti­fi­cial intel­li­gence sys­tems replace anim­al test­ing? This is the aim of so-called “digit­al twin” sys­tems. These are sim­u­la­tion pro­grammes that, for example, mim­ic the bio­chem­ic­al or bio­phys­ic­al prop­er­ties of human tis­sue. “This resource is increas­ingly used in the med­ic­al field, par­tic­u­larly in sur­gery, to give prac­ti­tion­ers the oppor­tun­ity to train before the pro­ced­ure, on a sim­u­la­tion of an organ resem­bling the patient. This avatar reduces the risks involved in sur­gery,” con­tin­ues Jean-Michel Besnier.

Tox­ic­o­logy is explor­ing anoth­er altern­at­ive aven­ue, that of organoids. These are three-dimen­sion­al cul­tures that are inten­ded to rep­res­ent an organ. “They are bio­lo­gic­al objects resem­bling an organ and pro­duced in the labor­at­ory from stem cells. How­ever, it is not a real organ, and herein lie the lim­it­a­tions. There is no guar­an­tee that the response of an organoid will be identic­al to that of a real organ,” explains Nic­olas Dav­id, a devel­op­ment­al bio­lo­gist at the Optics and Bios­ciences Labor­at­ory (IP Par­is), which is devel­op­ing this approach(1)¹. Sim­il­ar sys­tems are also envis­aged for per­son­al­ised medi­cine, to test the response of a patient’s cells before an anti-can­cer pre­scrip­tion, for example.

Fun­da­ment­al research is cer­tainly the field in which it is most dif­fi­cult to replace anim­al test­ing. Recently, a team from Wash­ing­ton Uni­ver­sity School of Medi­cine in St Louis² presen­ted a machine learn­ing algorithm cap­able of pre­dict­ing how a net­work of genes and the reg­u­la­tion of their expres­sion inter­act to con­struct the iden­tity of a cell dur­ing devel­op­ment. The sys­tem pre­dicts what drives a cell to become a muscle, skin or nerve cell, depend­ing on the genet­ic levers activ­ated. Called Cel­lOr­acle, it com­piles dec­ades of research from around the world, draw­ing on pub­lic data­bases that cata­logue known genet­ic inter­ac­tions. For example, it can be asked what effect the dis­ap­pear­ance of a gene will have in one of the mod­el organ­isms that the soft­ware integ­rates. This saves research­ers the trouble of design­ing anim­als that carry the genet­ic anom­aly. “It’s an in-silico sim­u­la­tion of knock-outs,” explains Nic­olas Dav­id, “anim­als whose gen­ome con­tains a muta­tion pre­vent­ing the expres­sion of a cer­tain gene.”

Cer­tain molecu­lar bio­logy tech­niques have been repro­duced so many times that they can be pro­cessed by a machine learn­ing algorithm. “But if we are able to sim­u­late a sys­tem, then we have under­stood it,” com­ments Nic­olas Dav­id. Such sys­tems save explor­at­ory work or point to unex­ploited aven­ues of research, but their res­ults are not infal­lible. Before embark­ing on an aven­ue and enter­ing a phase of applied research, they need to be verified.

The sys­tems also need to be veri­fied using oth­er math­em­at­ic­al approaches. “There are now bio­s­tat­ist­ic­al sys­tems for anti­cip­at­ing the min­im­um num­ber of anim­als that need to be used in research to answer a giv­en ques­tion. This approach is deployed at the Insti­tut Pas­teur and helps us to reduce the volume of anim­al exper­i­ments through optim­isa­tion,” explains Jean-Bap­tiste Mas­son, a stat­ist­ic­al phys­i­cist at the Insti­tut Pas­teur in Paris.

In recent years, new replace­ment approaches have emerged. Nic­olas Dav­id explains: “Not all anim­als are recog­nised as sen­tient. We can try to use sub­sti­tute spe­cies.” This may involve study­ing the earli­est stages of devel­op­ment, before the organ­ism is covered by the law on the pro­tec­tion of anim­als in sci­entif­ic research³. “In mam­mals, this depends on the total gest­a­tion peri­od. They are only pro­tec­ted after two thirds of nor­mal devel­op­ment. The situ­ation is more com­plic­ated when it comes to non-mam­mals. The law refers to an autonom­ous lar­val form. In fish, autonomy is inter­preted as the abil­ity to feed by itself, and there­fore at the moment when the mouth opens,” explains bio­lo­gist Nic­olas David.

Finally, some teams are turn­ing to inver­teb­rate anim­al mod­els, with one obvi­ous con­straint: the fur­ther one moves away from the human spe­cies in the evol­u­tion­ary tree, the great­er the risk that the con­clu­sions will not be applic­able to clin­ic­al research. This trend is also being chal­lenged by the sci­entif­ic com­munity itself. Last April, 287 philo­soph­ers, eth­i­cists, eth­o­lo­gists and neuro­bi­o­lo­gists spe­cial­ising in anim­al con­scious­ness signed the New York Declar­a­tion on Anim­al Con­scious­ness⁴. This states that it is “irre­spons­ible” to ignore the pos­sib­il­ity that all ver­teb­rates and sev­er­al inver­teb­rate spe­cies (such as ceph­alo­pods, insects and crus­ta­ceans of the crab and shrimp fam­ily) have a “con­scious exper­i­ence” in the light of a grow­ing body of sci­entif­ic research. The authors refer to ground­break­ing work on bees⁵, octopuses⁶ and two spe­cies of snake⁷. This declar­a­tion there­fore con­sti­tutes a pre­cau­tion­ary prin­ciple in the face of the pos­sib­il­ity that spe­cies used for sci­entif­ic exper­i­ments may become con­scious. All the more reas­on to resume the search for alternatives.

Agnès Vernet