An AI-Assisted Workflow for Reconstruction, Extension, and Calibration of Quantitative Systems Pharmacology Models

Q u a n titativ e systems ph arm ac ology (Q SP) models provide me chanisti c insi ght into dr ug response b ut are lim ited by labor -in tensive, exper t-d r iv en wor kfl ows . We developed an AI-assisted QSP (AI- Q SP ) fr amework that integrat es large lan guage models (LLMs) wi t h SBML-based modeling to enable autom a ted reconstr uc t ion, ex t ens ion , a n d cali b ration of m ec hanistic models.

The fr a m ew o rk was applied to a published CAR- T QSP model. The mod el was reconstr uc t ed in S B ML and ex t ended via LLM-guided pr omp t s to incor pora t e key resi s t ance mechanism s : T- cell ex h aus t i o n, PD-1/ PD- L 1 che ckpoint regulat i o n, and tum or antigen es c ape. M odel dev elop men t fo llowed an iter at ive ex pert-i n- the -loop wor kflow. The resulting model (21 r eac t ions , 9 spe ci es ) wa s calibrat ed to sy n thetic bench mark data using 19-par ameter optim iza t ion. Model cr edibili t y was as se ssed u s ing ASM E V & V 40 and ICH M15 pr inci p l es, in cluding gl o bal sen sit iv it y and p rof ile -likeliho od analys e s.

The cal ibr ated model reprod u c ed benchmar k d ynam i c s w it h high accurac y (mean log-RM SE = 0.132) . S en sitivity analy s i s identified CAR-T ki lling and bystand er cytot oxi city a s dominant dr ivers of tumor response. Prof i le-l ikelihood ana lysi s showed 71 % o f parameters were pr a ct i cally identifiable, with re m ai n i n g p a r amete r s pr iorit ised for f u tur e d a t a-d riven ref i n em ent.

AI-as s i sted QSP modeling enables r eproducible, scalable model reconstr uc t ion and evoluti on while ma intain ing mechanistic transparen cy and re gulator y a lignmen t. This fr a m ework provides a f oundat ion f o r ac c eler ating model- informed dr ug de v elop ment in cell and gene ther apies . Keywor ds. C AR-T th e r apy; quantitat i v e sy s t em s ph a r macolog y; A I-assi sted modeling; SBML; T-cell ex haustion; ant igen escape; ch ec kpoint inhibition; profi le-l ikeli h ood identif iabil ity; global s en sitivity analys i s; model-info r med dru g develop m ent. 1. Introduction Chimeric antigen receptor T-cell (CAR-T) therapies have transformed the treatment of hematologic malignancies, yet clinical responses remain heterogeneous: while durable remissions are observed in a substantial proportion of patients, treatment failure is frequently linked to limited in vivo CAR-T persistence, T-cell exhaustion, tumor antigen escape, and cytokine-related toxicities [1–3]. Understanding and was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint predicting such nonlinear, patient-specific dynamics require mechanistic frameworks that couple biological realism with quantitative inference [8,9,31]. Quantitative systems pharmacology (QSP) models have emerged as a central methodology for integrating multiscale immune–tumor processes into predictive simulations [10,21,31]. In the context of adoptive cell therapy, QSP approaches enable exploration of the interplay between CAR-T proliferation, exhaustion, cytokine release, and tumor burden under varying therapeutic regimens [10,15,16]. However, traditional QSP development is labor-intensive, relying on manual calibration and limited reproducibility across platforms [21,31]. To addres s these challenges, we developed an AI- Q SP (Artificial Intelligence–assisted Quantitative Systems Pharmacology) pr otot y pe designed to extend tradit i o nal QSP workflows b y i n tegrating aut omated mod el gener ati on, paramet e r opt i miz at i o n, a n d model verification wi thin an A I-as si s t ed modeling envi r onm ent . Unli ke cla ssica l mo del ing pipelines th at r ely en tire ly on manual mode l cons truct i o n a n d cali b ration, th e AI-QSP framewor k co mbines large language models (LLM s) with establis h ed s ystems pharm ac olog y too l s to as s i st in: • conversion of litera t ure m odels into SBML f or m at , • generat i o n of s t ructur al m od el updat es ba s ed on t ex t ual bio logic al knowle dge, • autom ated para met er cali b ration, an d • iter a tive m od el v al idat i o n. All m odel componen ts ar e i mplemen ted using Systems Biology M arkup Language (SB ML) Level 3 Version 2 , enabling full repr o ducibili t y a n d i n terope r abil ity wi t h s t andard simulation tools su c h a s Tell u rium [26], H e ta [27 ] , a n d DBSolve O pt imum [28] . This hybrid framework couples mechanistic transparency with computational intelligence, allowing continuous model refinement as new data become available [8,9]. In this study, the AI-QSP prototype was applied to CAR-T therapy modeling, capturing effector and exhausted CAR-T populations, antigen-positive and -negative tumor subclones, and cytokine feedback (IL-6, IL-10, IFN-γ ). Our objective was to evaluate the feasibility, strengths, and limitations of AI-assisted automated model evolution. Model behavior under the Triple Combination scenario was then generated and quantitatively fitted to synthetic benchmark data. The framework is evaluated here through a controlled test case: reconstruction of an existing CAR-T QSP model, AI-guided structural extension to include resistance mechanisms, and automated calibration against synthetic benchmark data derived from the original model. was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint The fr a m ew o rk is desi gned to align wi t h FDA, EM A, and ICH guidance on mod el - in formed developmen t [ 4,5,33], pro v iding an SBM L -encoded , aud i t able m odeling pi peline c on s istent with regulat ory expec t ations for tr an sparenc y and r eprod uc ibility in comput a t ional pharm ac ology. The goal of this study is not to develop a new CAR-T model, but to evaluate whether an AI-assisted modeling framework can semi-automated reconstruct, extend, and recalibrate an existing QSP model while preserving its quantitative behavior. Specifically, we assess (i) the capability of LLMs to introduce biologically motivated structural extensions from textual descriptions; (ii) the ability of the iterative expert-in- the-loop workflow to converge on a technically correct SBML implementation; and (iii) the capacity of automated calibration to recover reference-model dynamics under a challenging multi-mechanism scenario. 2. Methods 2.1 AI-QSP workflow AI- Q SP f ramework overview The overall architectur e o f the AI- as si s t ed quantitat iv e sys t ems ph a r ma cology ( AI-QSP) fr a m ework is illustrat ed i n Figure 1 . The framewor k impl em ents a n it e rativ e workflow th at integr a t es ar tificia l i nt el li gence–base d knowl ed g e int erpr etat ion, mechanistic modeling in Sy s t ems Biolog y Mar kup L an guage (S BM L ) , and au tomated pa r amete r opt i mization. Start ing fr om a l i t era tur e-d er i ved mo del des cr ipt i o n, the sy stem f i r s t reconstru cts a mechanis t i c core model in SBML for ma t. The r ec onstr ucted model is t hen ver ified th rough simulation to ensure con si sten c y with the r efe rence for mu lation. Subsequently, the A I s y stem pro pos es structu ral model extensi o ns based on textual bi ological kno w ledge des cribing additional mechanis m s r e levant t o the th e rapeu tic s ystem. These prop o s ed mod i f ic ations ar e evaluated th r ough expert rev iew, and t he f eedback is u s ed t o refine p r omp ts and regener ate imp ro v ed model i m plem e ntat i o ns . The wor k f low it erates between AI-d r iven model generat i on and exper t valida ti on unt il a t ec hn i cally c on s i stent mo del str ucture is obtained. The f i n a l m odel i s cali b rated against ben chmark dataset s u s ing a utom ated paramet er estima tion and subsequently validated by repr oducing t he dynamics of t he ref erence mode l. This it erative p roces s enables sy stematic reconstru c t i o n, ex t ens ion, and vali d ation of mechanistic p harmacology models while maintaini ng t ransparency a n d repr oducibili ty thro ugh s t a n dardized S B ML representat ions . Description of workflow steps The number ed ele ments in Figure 1 c orrespond t o the sequentia l s t eps of the AI-QSP workflow. In Step 1 , a liter atur e-d eriv ed model description is inte rp reted b y th e A I system, which ex t ract s r elev an t bio logic al ent i t ies , pr ocesse s , and mod el ing instr uct ions . In St ep 2, these elements are tr ans lated into a mechanis t i c mod e l encoded in SBM L fo rmat, pr oducing was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint the r ec on s t ructed core mod e l (Out put O1). In Step 3 , t he S B ML model is si mulated and compared with t he behav ior of t he r e fer enc e mo del to ver ify the corr ectnes s of t he reconstr uc t ion (Output s O2– O 3) . In Step 4 , the AI s y s t em proposes s tr uctu ral updat es t o the model based on textual d escriptions of additi onal bio logical mechanis m s. Thes e up dates ar e examined during Step 5, w her e a domain ex p ert evaluates the generat ed model s t ructur e and identi f i es technical i n c on sisten cies or b i o l o g ically i m plaus ible mechanisms ( Out put O4). The expert f ee d bac k is then used in Step 6 to ref ine the prom p ts guiding the AI mode l generat i o n proces s ( Output O 5) . In St ep 7 , the r e f ined prompts ar e us ed t o regenerat e an updat e d m odel impl ementa t ion inc o r por a t ing the pro p os ed m ec hanism s ( O ut put O6). Step 8 repr es ent s t he i n tegration of the AI - ge n erated mod e l exten s ions into the SBM L fr a m ework, produ c ing an updat ed mec h a n i sti c model. In Step 9 , t he result ing model is cali b rated using automat ed paramet er es t i m ation to r ep roduce benchmar k dat as et s (Output O 7) . F inally , in Step 10, th e c alibrat ed model is validated by comparing its si m ulated dynamic s with r eference model beha vior and ben c hm a r k obs er vations ( O u t put O8). Together , these s tep s for m a c losed iter a t iv e loo p tha t enables progr e ssiv e r efinem ent and vali d ation of mechanistic QSP models us ing A I-assi s t ed model devel opmen t. Figure 1. Iterative AI-QSP workflow for automated reconstruction, extension, and calibration of mechanistic pharmacology models. was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint The framework integrates large language model–based knowledge interpretation with SBML-based mechanistic modeling and numerical optimization. A literature model description is first interpreted and converted into an SBML representation of the core model (O1). The reconstructed model is verified through simulation to ensure consistency with the reference implementation (O2–O3). AI-guided model extension introduces additional biological mechanisms based on textual knowledge prompts. Expert review and correction ensure biological plausibility and technical consistency (O4). Updated prompts are generated iteratively to refine model structure (O5–O6). The resulting model is calibrated using automated parameter optimization (O7) and subsequently validated by comparing simulated dynamics with benchmark data (O8). The workflow forms a closed loop enabling progressive improvement of mechanistic QSP models through AI-assisted model evolution. The AI- Q SP prototype integr a t es larg e language models (LL M s ) with P yt hon-based modeling tools to suppor t automated Q SP mo del d ev elop m ent. The fr a mework oper at es th rough structu red prom pt workflows that guide t he AI s ystem in inter pre t ing b i o l ogi cal knowledge, generat i n g m odel updates, and prod ucing exe c u table model code. The AI s ystem int eracts wi t h modelin g tools exec ut ed loc ally within a federat ed comput a t ional env i r onmen t to ensure pro tection o f pr opr ietar y a n d c on fi dential data. Hu man ex perts va lidat e each step of t he w o rkflow to ensure biol ogical pla u s ibility and technical c or rectne s s . The main functional r o les of t he AI-Q SP sy stem include: 1. Conversion of literature models into SBML f ormat 2. Identification of missing biological mechanisms fr om textu a l descriptions 3. Generation of structural model updates 4. Automat ed code generation and modif icat ion 5. Assistance in model calibration workflows Hu man ex pert valida tion r em ai n s an es s ent i al componen t of the pr oces s t o ensure model corr ec t ness. The AI-QSP fr am ewor k can be e x ec u t ed locally within a f eder at ed en v i r on men t t o ens ur e pr ot ec t ion o f pr o pr iet ar y and c onfide n t ial dat a. To validate the AI-QSP prototype, we designed a closed-loop tuning and validation workflow (Figure 1). This framework integrates a mechanistic QSP model, encoded in SBML, with an algorithmic parameter optimization engine and automated data- driven feedback loops. The workflow's primary test, as outlined in this study, was its ability to autonomously update core QSP model to include new processes, species, regulations and phenomena then calibrate the complex CAR-T QSP model against a pre-defined, was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint ground-truth benchmark. Each cycle begins with the SBML base model and an imported dataset. An algorithmic parameter optimization strategy explores the parameter space to minimize the error between simulated and observed data. The framework then automatically generates visual and quantitative reports, such as overlay plots (Figure 6) and goodness-of-fit statistics, before the process repeats for subsequent scenarios. 2.2 AI implement ation The AI-QSP framework was implemented in Python using a modular architecture combining large language models with mechanistic modeling tools.LLM componentOpenAI GPT-based language models were used to interpret biological text descriptions and generate candidate model updates.Execution environmentAll AI interactions were executed through a Python interface that converts LLM- generated instructions into structured SBML model modifications.Model compilationGenerated model structures were validated using Tellurium.Expert validationAll AI-generated model modifications were inspected by a QSP expert prior to integration into the model. 2.3 Mechanistic Core Model The refer ence model us ed f o r validat i on of t he AI-QSP framewor k was a pr evious ly published quantit a t i ve sy s t ems p har macology (QSP) m odel describing th e d y n a m ics of chimeric antigen recepto r T-cell ( CAR -T) therapy. The model captur es ke y inter a ct i o ns between t umo r cells , CA R-T cell po pulations, a n d cyto k ine si gnali n g p roce sse s t hat regulate ther apeut ic response , i m mune activa t ion, and tr e a t ment r es istance me cha nis ms [10]. To ensure repr oducibility and inter o p erabili ty, the mo del was implement e d u s ing Syst ems Biology Markup Language ( SBML) Level 3 Version 2 , a widely adopted sta ndar d for repr es ent ing and ex changin g mechan is t ic biological models [ 25]. SBML enc oding al lows models to be exec u ted and analyz ed acros s m ultiple s imu l at ion env i r onm e nt s commo nly used in sy stems pharma cology and c o m putationa l biol ogy. In this s t ud y, s i mulation s and structu ral v alidat i o n o f the mod e l we r e p erfo r med usi n g several SBML-compat i b l e m odel ing platfo r ms, inc lud i n g Tellurium [26] , Heta [2 7], and DB So l ve Opt i mum [28] . Th es e environm e nts suppo rt det e r min istic o rdinary diff e rentia l equati on ( O D E) sim ula t i on, paramet er estima tion, and m ode l i n s pection. The SBM L r epresentation also enables compatibility with addi t ional m odel in g framewor ks and numer ic al engine s bas ed on libSBML and libRoadRunner, facilitating repr oducible s imulat i o n and c r os s- platf or m verification of t h e model impl ement a tion [29,30] . The mechanis t i c str u c t ure of t he c or e model includes several interacting bi ological entities repr es ent ing k ey componen ts o f the CAR -T therapeutic s y s t em. CAR-T cell populations • Effector CAR- T cells ( CARTE) was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint • Mem ory CAR- T cells ( CARTM ) The core model includes effector CA R -T cells (CA RTE) and memor y CAR-T c ells (C ARTM) , repr es ent ing ac t iv at ed c yto toxic and long-lived me mo ry populations respectively . The exhaus t ed CAR- T stat e ( CARTE_EX), w hich arise s dur ing prolonged antigen stimulat i o n, i s intr oduced as a new s p ecie s by t he AI-as s isted mod e l extension desc r ibed in Sec t ion 2.4. Tumor cell populations • Ant ig en -positiv e t umo r c ells ( B_pos ) The core model t rack s antigen-po sitive tum or cell s (B_pos) only. Ant igen-n egative tumor cell s (B_neg), representing t umor cells t hat have l o s t t he target ed s ur face ant igen throu g h antigen escape [2,3], ar e int roduced by the A I-as si s t ed model extensi o n (S ec t ion 2.4). Cytokine signaling molecules • Inter leukin-6 ( IL-6) • Inter leukin-10 ( IL-10 ) • Inter feron-γ ( IFN -γ ) These c ytokines repr es ent k e y inflammat ory mediato r s as s o c iated with C AR-T expansion, immune activati on, and sy s t emic c ytokine response s du ring ther a p y . Therapeutic agent • Ant i -PD- 1 c hec kpoint inhibit o r ( aPD1) Ant i -PD- 1 c hec kpoint inhibit o r (aPD 1) is no t present in the cor e m od el. It is int roduced as an explic it t herapeu t ic specie s by the AI- assi s t ed model ex t e nsi o n (Sec t ion 2.4), where it modulat es CA R-T functional exhausti on by attenu a t ing PD-1 inhibit o ry s ig naling, ther eby enhancing CAR- T p e r s i stence and cyt otoxic ac t i vity [1] . The model is formula t ed as a sy stem of coup l ed ordinary differential equations ( ODEs) des c r ibing the tempo ra l ev olution of cell populations and cy t okine c on c ent rations. A single well-mixed compartment is as s u med , approximating t he com bined tumor and peripher al blood envi ron ment. Cel lular proli fera t ion, diff eren t iation, exhaustion, tumor growt h, antigen esc ape , and CA R-T-med i at ed cyt otoxic it y ar e described using mech anis t ic ally inter pre tab l e kineti c rat e l aws con s i stent with established QS P m odeling practice s [31] . was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint Figure 2. Mechanistic structure of the core CAR-T QSP model. The model describes the dynamics of antigen-positive tumor cells (Bpos), effector CAR-T cells (CARTE), and memory CAR-T cells (CARTM). CARTE proliferation is driven by antigen stimulation from Bpos. CARTE cells can differentiate into memory CAR-T cells (CARTM). Effector CAR-T cells mediate tumor killing of Bpos. Both CARTE and CARTM undergo natural death processes. Tumor cells proliferate logistically. Three split-dose intravenous CAR-T administrations were implemented to reproduce the dosing schedule described in the reference study. The total CAR-T infusion dose was divided into three fractions administered on consecutive days: 10% of the total dose on day 0, 30% on day 1, an d 60% on day 2 . This staggered infusion strategy reflects commonly used clinical protocols designed to mitigate acute cy tokine release syndrome while maintaining therapeutic cell expansion. In the model, each dose fraction is introduced as an external input to the effector CAR-T cell population ( CARTE). All model variables, parameters, and rate constants were defined with explicit physical units to ensure dimensional consistency and reproducibility of simulations. was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint 2.4 AI-QSP–Based Automatic Update of the AIcore QSP Model To extend the r ec o ns t ructed AIcor e Q SP model, t he AI-QSP framewor k was us ed t o identify additional b iological mec h a n i sms relevant to C AR -T th erapy that wer e n o t explic it l y repr es ent ed in th e ini tial m odel. This s t ep w a s designed t o ev aluat e th e capabil it y o f th e fr a m ework to in ter pre t bio logic al kno wledge from t ex t ua l des cript ions and t ranslate this knowledge into mechanistic m odel ex t ens ions. Large language m odels (LLM s) are part ic ular ly suitable for th is task becau se t hey can extr a ct mechanistic r elationships and biologi cal entities from un s tr uctured scientif ic text and convert t hem in to stru c t ured rep rese ntations s u i t able for comput at ional modeling. In the AI-QSP fr amework, LL M s were t herefor e us ed as a kno w ledge-int erpr eta ti on layer t hat tr ans lates biological des cription s into candidate m odel com ponents, inc lud i n g n e w species, reactions, and regulato ry interactions . The pro mpts provided to t he AI syste m fo c used on ident i f y ing, int erpr e tin g, and implemen t ing m ec han i sms a sso ci at ed wi t h well - known resi stance pr oc e s ses observed in CAR -T therapy. In par ticular, the fr am ework wa s instr u c t ed to incorpor ate biological knowledge related to t he fol lowing phenomena: • T-cell exhaustion , rep rese n ting the function al decline of activated C A R-T cel ls during pro l o nged a n tigen s t imulat ion. • PD-1/PD - L1 checkpoint signaling , wh ic h suppr e sse s T-cell ac t iv it y t hrough inhibit o ry recepto r–ligand int eractions at th e i m munological syn aps e. • Tumor antigen escape , w h ereby tumor c ells los e expres sion of the tar gete d antigen and ther eby ev ade CA R-T- m edia t ed cy t otoxicity. These mec h a n ism s are known to red uc e CA R-T pro l if era tion and cytot ox ic efficacy and ar e ther efo re important determinants of tr e at men t r es pon s e and resistan ce. Based on the textu a l p r ompts, the A I- QS P s y s t em generated pr oposed upd ates to th e SBM L model stru c t ure. These updates i n c lu ded t he in tro duc t ion of new state var iables, add i t ional reactions des c r ibing biologi cal tr ans ition s , and modif i ed r ate laws c aptur ing regulator y inter ac t ions . The pr oposed model ex tensi o ns were sub sequently integrat ed into the SBM L repr es ent ation and e valuated th rough simulation and exper t r ev iew as part of the ite r ative workflow des cribed in Figure 1 . 2.5 Example prompt used in the AI-QSP framework The following s im pli f ied promp t ill u s trates the instruction s used to guide the A I-as si s t ed model updat e process: Given an SBML model describing CAR-T therapy dynamics with the species CARTE, CARTM, B_pos, cytokines, and tumor growth processes: 1. Identify important biological mechanisms in CAR-T therapy that are not currently represented in the model. 2. Focus on the following mechanisms: was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint - T-cell exhaustion - PD-1 / PD-L1 checkpoint regulation - tumor antigen escape 3. Propose model extensions by: - introducing new state variables if required - defining new reactions or transitions - modifying rate laws if necessary 4. Describe the proposed model updates in SBML-compatible form including species, parameters, and kinetic expressions. The generat ed model m odif ications were sub s equen tly e valuated b y several L LM s and domain exper t s t o ens ur e biologic al plausibil it y and t ec h nic al corr ec t ne ss befor e b e i n g incorpor ated int o t he upda ted mo de l i m plem entati o n. 2.6 Simulation Scenario To evaluate the ab ility of t he A I -QSP framewor k t o repr oduce t he b ehavior of t he r efer enc e model, simulations were performed u sing a single integrated m echanistic scenario that incorpor ates the principal bio logic al mechanism s inf l u enc ing CA R-T ther apy d y nam i c s. Thi s s c enar io i n c ludes CAR- T c ell exhaus t i on, c h e c kpoint i n hibition thr ough ant i-PD -1 ther apy, tum or ant igen esc ap e, and by stander -mediat ed cytot ox icity. In the mo del i m ple mentati o n, these m ec hanism s w er e repr es ent ed by as signing non- z er o values t o the rate constant s governing each me chanis m ( set t o z er o in bas eline si n g le- mechanism s cenario s), ther eby switc hing eac h p a t hwa y o n w it hin the SB ML reaction networ k. S pecifically, t he ex hau s tion t ransition fr o m ef fector CAR-T cells t o the exhaus t ed phenot ype w as enabled by s et ting the exhaus t ion rate par ame ter t o a pos it iv e value /g1863_/g1857/g1876/g1860 /g3408 0 . Tumor antigen esc ape wa s int roduc ed thr ough conversion of antigen-p os itive tum or cell s ( B_pos ) t o ant igen-negati ve cells ( B_neg) via t he antigen-loss r a t e /g1863_/g1864/g1867/g1871/g1871 > 0 . Chec kpoint blockade wa s repr es en ted by the presence of circulating anti-PD -1 ( aPD1), which reduce s t he effective ex hausti on r ate th r ough the r egula t o ry relat i ons hip def ined in Eq. (3). In addit ion, a bystander -medi ated killing mechanis m wa s inc lud ed throu g h activation of th e b y stander ki lling rat e /g1863_/g1854/g1877/g1871/g1872 /g3408 0 , enabling par tial e l im i n a tion o f anti gen- negative tumor cel ls. This combined mechanistic configuration (“Triple Combination” scenario) ( n o t e : t h e t e r m ref ers t o the th ree p rincipal resis t a n ce axes — exhau s t i o n, c he c kpoint r egulation, and antigen esc ape — a s fr amed in the r e ference model [10] ; by stand er kil ling fun c t ions as an auxili ar y c ytot oxic mechanism) wa s s elec t ed bec ause it simultaneously c a pt ures the major biological pr ocesse s kno wn to inf luen ce C AR-T th e r apeutic outcomes, inclu ding im mune exhaus t ion, immune chec k p oint regulation , t umor escape, and s e condar y c ytot ox ic mechanism s . U s ing th is integr a t ed s c enario pr ov ides a stringent test of the A I-Q SP fr a m ework’s ability t o recons t ruct an d calibrate a mechanistic m odel c apa ble of reprodu c ing complex CAR-T t reatm en t dynamics . was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint 2.7 Ground-Truth Benchmark Generation To evaluate the calibra tion perfor ma nce of the AI-QSP fr a m ework, a s ynth etic dat as et was generat ed us ing simulations of the r e c on str ucted AIcore QSP model. This d atas et s er v ed as a ground-truth benchmark for t es t in g w h ether t he updated AI- generated m odel c ould repr oduce the dynamics of the refer e nce model. The model was simulated ove r a 28- day post-infusion window (nine observation time points, corresponding to the measurement sch edule of the reference stu dy) , c o n s istent with the time s cale repor ted in the o riginal stud y [ 2], where long-ter m kine tics of C AR-T cells and c ytokine respon ses were analyz e d. Synth etic o b s er v at i o ns were gener ated for key model var iables inc luding total tumor burden (Total_Tumor = B_pos + B_neg) , circulating effector CAR -T cells (CARTE_PB) , and cyt ok ine concentrat ions ( IFN-γ, IL-6, IL-10 ). Ob s er v at ion tim e point s were selec t ed to corr es pon d to the m easurem ent s c hedu l e repo rted in th e refe r e n c e study [2] , c aptu ring both t he early ex pansion phase of CAR- T c ells and the late r c o ntraction and per s istence pha ses . The se time poin ts span t he entir e si m ulation int er v al and al low eval u a ti on of th e m odel’s ability to r ep rodu c e bot h short-te rm treatment dynamics and long-term system behavior . To emulate bio logical v ar iabil i t y a n d meas ur ement uncer tainty typicall y o bserved in clinic al datasets , multiplicative log- normal noise wa s app l ied t o the simulated val ues . The noi se distribut ion was defined using a coefficient of variation (CV) of 20% , r epre senting moder ate experimen t al var i ab ili t y comm only r eport ed in immuno logical meas ur ements. The resulting dataset pr ovi d e s a self -consistent synthetic benchmark deri ved directly fro m the r efe rence model. This benchmark enables qu a n titativ e ass e s s ment of t he A I-Q SP cali b ration wor k f l o w by t es ting whet her the upd a ted mode l impl ement a ti on can repr oduce the 28- day post-infusion dynamics of tumor burden, CAR-T expansion and contraction, and cytokine responses with comparable accura c y t o the original m odel. 2.8 Automated Parameter Optimization and Evaluation Model calibration was performed by the AI-QSP framework's optimization engine. The objective was to minimize the log-transformed root mean square error (log- RMSE) between the model simulations and the synthetic benchmark data. A random-restart global search strategy, implemented in Python with Tellurium and NumPy, was used to sample 200 candidate parameter vectors within plausible biological bounds. The full calibration comprised 19 parameters: seven kinetic parameters governing CAR-T and tumor dynamics (listed below), plus six cytokine production and stimulation parameters ( α and β for IL-6, IL-10, and IFN-γ ), three cytokine degradation rate constants (k_deg, IL6, k_deg, IL10, k_deg, IFNγ ), two antigen-recognition parameters (K_DX and K_kill), and one baseline degradation constant (k_deg,base). The seven primary kinetic parameters were: • CARTE expansion rate constant k_prolif was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint • Exhaustion rate k_exh • Cytotoxic killing efficiency KBC • Antigen-loss rate k_loss • Bystander-killing rate k_byst • Effector death scaling factor dE_scale • Anti-PD-1 efficacy Emax_PD1 Model fidelity was assessed by the final log-RMSE score across six calibration variables and by visual inspection of overlay plots (Figure 6). The Mechanistic Reaction-Based model (21 SBML reactions, 19 calibrated parameters) achieved a mean log-RMSE of 0.132, with exhausted CAR-T cells (log-RMSE = 0.085) and IL-6 (log-RMSE = 0.067) meeting the < 0.10 accuracy threshold. All SBML models, fitting scripts, parameter files, and benchmark data are available to ensure full reproducibility. 3. Results A was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint B Figure 3. Structural representation of AI-generated and expert-corrected CAR-T QSP models. (A) First AI-generated model update (AIupdate1) incorporating mechanisms of T-cell exhaustion, PD-1 checkpoint regulation, and tumor antigen escape. (B) Expert-corrected model implementation (corrected_AIupdate1) after resolution of SBML structural inconsistencies, including explicit reaction definitions and parameter annotations. was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint Figure 4. SBML representation of the AI-generated CAR-T QSP model update (“AIupdate1”). Selected fragments of the Systems Biology Markup Language (SBML) implementation generated by the AI-QSP framework are shown to illustrate the structure of the automatically produced model. (A) Definition of model compartments, species, and parameters introduced during the AI-assisted update, including exhausted CAR-T cells (CARTE_EX) and antigen- negative tumor cells (B_neg). (B) Example reaction declarations implementing key biological processes such as CAR-T proliferation, exhaustion transitions, and tumor antigen escape. (C) Corresponding kinetic rate-law expressions defining the mathematical representation of the biological interactions. was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint Figure 5. Expert-corrected SBML implementation of the CAR-T QSP model (“corrected_AIupdate1”). Representative fragments of the SBML code after expert validation and correction. (A) Revised species and parameter declarations ensuring dimensional consistency and explicit model annotations. (B) Corrected reaction definitions implementing CAR-T proliferation, exhaustion transitions, and tumor killing processes. (C) Corresponding kinetic rate laws defining the final mathematical formulation used for simulation and calibration. 3.1. Analysis of structure of updated QSP model generated by AI-QSP prototype We have compared structure of initial AIcore QSP model with AI-QSP generated “AIupdate2” version of the model. This analysis includes comparison of variables, processes and expressions for rate laws of the models. 3.1.1. Variables The AIupdate2 model differs from the AIcore model in both the set of state variables and the implemented biological processes. Indeed, there are 3 groups of variables in these model versions: - Group 1: Included in AIcore version only - Group 2: Included in “AIupdate2” version only - Group 3: Included in both versions was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint Group 1 includes variables describing CART cells in tissue. Group 2 includes variables describing exhausted state of CART cells in blood and antigen negative tumor cells. Group 3 includes variables describing effector and memory CART cells in blood, antigen positive tumor cells and cytokine concentrations. 3.1.2. Processes “AIupdate2” and AIcore versions of the model differ each other in processes. Indeed, “AIupdate2” version does not include processes describing migration of CART cells in tissue but AIcore does not include transition between effector and exhausted states of CART cells and transition between antigen positive and negative tumor cells. 3.1.3. Expression for rate laws Difference in expressions of rate laws between “AIupdate2” and AIcore versions of the model are summarized in Table 1. Main differences are summarized in Table 1. The AIcore model includes tissue-compartment migration reactions for both effector and memory CAR-T cells that are absent from AIupdate2; conversely, AIupdate2 introduces exhausted CAR-T state transitions (Rxn_CARTE_Exhaustion), antigen- negative tumor cell dynamics (Rxn_Bneg_growth, Rxn_Bneg_Bystander_Killing, Rxn_Antigen_Escape), and a bystander killing mechanism, none of which are present in AIcore. Empty cells in the AIcore column indicate processes unique to AIupdate2 [10]. T able 1. C omparison of r a t e law s ID “AIupdate2” QSP model Expert choice Comment Do s e C A RTE in jec t ed Rxn_CARTE _pr o lif k_pr o lif_eff * CART E * B_p o s /(K_pr o lif + B _ p os) * ( 1 - CA R T E / CA R T E_ max ) V *k _p ro l i f _e f f * C AR T E * B _ p os /(K_pr o l if + B_pos) * (1 - C A RTE / CA R T E _ m a x ) Sa tur a tio n wi th a n t ige n an d resour c es limit ati on wer e t a k e n i n t o a c c o u nt Rxn_CARTM_Activ a tion k _ a ct * f _ A g * C A R T M V *k _a c t * f _A g * C A R T M Rxn_CARTE _Exhaus tion k _ exh _e f f * C A R T E V *k _ e x h _e f f * C A R T E Exhaus tion i s imp lem ent ed R x n _ C A R T E_ d e at h k _ d e at h _E * C A R T E V *k _d ea t h _ E * C A R T E Rxn_CARTE _Mem Di ff k_me m * CA R T E V * k _ m e m * CARTE In c o nt e x t o f t he m o d e l d i f fe ren t i at ion E /barb2rightM shoul d n o t n e g ative ly d e p e nd on an ti g en Rxn_CARTM_dea th k_dea th _ M * CARTM V * k _ d e ath_M * C A RTM Prolife r ati on o f CARTM C A RTM c ells a r e able to p r ol i f e r at e slowl y Rxn_CARTX_dea th k_dea th _ X * CARTX V * k _ d e ath_X * C A RTX Mi g r ati on o f C A RTE t o ti s s u e f ro m b lo od Tissue compartm e nt a nd C A RT ce l l migr a t i o n is i m port a nt was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint Mi g r ati on o f C A RTE to blo od fr o m t i ss u e Tissue compartm e nt a nd C A RT ce l l migr a t i o n is i m port a nt Mi g r ati on o f CARTM t o tissue fr o m bloo d Tissue compartm e nt a nd C A RT ce l l migr a t i o n is i m port a nt Mi g r ati on o f CARTM t o blo od fr o m ti ss u e Tissue compartm e nt a nd C A RT ce l l migr a t i o n is i m port a nt Rxn_B p os_g r o w th r_ tu m o r * B_p os * ( 1 - B_p os / K_tumor) V * r _tumor * B_po s * (1 - B_ po s / K_ t um or ) resour c es limit ati on wer e t a k e n i n t o a c c o u nt Rxn_B p os_k i lli n g k_kill * CA R T E * B_po s / (K _ kil l + B_pos) V * k _kill * CARTE * B _ p os / (K _kil l + B_p o s) Ki l li n g is s a t ur a b l e f u nc t i o n o f an ti g en Dea th of t umor ce l ls V * k _kill * CARTE * B _ p os / (K _kil l + B_p o s) I t i s b ette r t o s ep ara te o f d e ath of t u m o r ce l ls a n d kil li n g R x n _ Bn eg _g ro w t h r_ tu m o r * B_neg * (1 - B_ne g / K_tumor) V * r _tumor * B_neg * (1 - B_ ne g / K_ t um o r ) An t i g e n ne g a t iv e t um or c e l ls are imp lem ent ed R x n _ Bn eg _B ysta n d er _ K illing k_bys t ander * bys t ander_ factor * CA R T E * B_neg /(K_ k i ll + B _ n e g) V* k_ by s t a n de r * by s t a n de r _ f a c t or * C A RT E * B _ n e g / ( K_ k il l + B_ ne g ) By s t a n de r e f f e c t i s i m ple ment ed Rxn_Antige n _ Esc ap e k_esc *B_ p os V * k _esc*B_po s A n tige n esc ape is i m ple ment ed Rxn_IL6_S y n _ C ARTE p_IL6_E * C A RTE V * p_IL6_E * CA R T E Rxn_IL6_S y n _ B pos p_IL6_T * B _ p os V * p_IL6_T * B_pos Rxn_IL6_Deg k de g _IL6 * IL 6 V * k deg_IL6 * I L 6 IL6 pr od uct i o n by endo g en o us T ce l ls R x n _ I L 10 _Sy n _ C A R TE p _ I L 10 _E * C A R TE V *p _ I L 1 0 _E * C AR T E R x n _ I L 10 _Sy n _ B p o s p _ I L 10 _T * B _ p o s V *p _ I L 1 0 _T * B _ po s Rxn_IL1 0 _ D e g k de g _IL10 * IL1 0 V * k deg_IL10 * IL 1 0 I L 10 p rod u c t ion b y endo g en o us T ce l ls R x n _ I F N g _ Syn p _ I F N g_ E * C A R T E V *p _ I F N g _E * C A R TE Rxn_IFNg_ D eg k de g _IF N g * IFNg V * k deg_IFN g * IF Ng IFNg pr od u c tio n by endo g en o us T ce l ls 3.2 Application of the AI-QSP workflow to automatic update of the AIcore CAR-T model Us ing the r ec on s tr ucted AIcore Q S P mod el as input and applying the pr o mpt wor kflow des c r ibed in S ect i o n 2.4, the AI-QSP fram ew o rk generated a fir s t u pdated version of the CAR -T model ( AI update1) , i n c or porat ing t he requested biol ogical mechanis m s related t o T- cell exhaustion, PD -1/ PD-L1 che c kpoi nt regulation, and tumor anti gen esca pe. The corr es pond i n g m odel scheme and SBML implem ent ati on are shown in Figure 3A and Figure 4 , respe ctively . Inspe c t ion of the SBM L code of t he AIupdate1 model r ev ealed th a t the f r a mework wa s able to int r oduce the int end ed biol ogica l mechanism s, b ut the initial impl ementation contained sev er al t ec h nic al incon si sten cies. T h e s e included incomplete annot a t io n of compar tments, spe c ies, and paramet ers; a b s en ce of explic it stoichiomet ric process definiti o ns for s om e reactions; ambiguou s mapping between r a t e ex pr es sions and biologic al pr oc es ses; and an inconsistenc y in the CAR TE-to-C ARTM t ransition, which w a s repr es ent ed i n t he equat ion for was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint CARTM but not consistent ly ac counted f or in the equation fo r CARTE. Thes e o bs er v at ions indicate that t he AI-QSP workflow could prop os e biologically r elev ant m od el extensi o ns , but that ex per t validat ion re m a ined n e ce s s ar y t o ens u re technical cor rectne s s and SBM L c o ns i s t e nc y. All identi f i ed technic al deficiencies in the AIupdat e1 model were manually c o rrected and reimp l e m ented in H et a fo rm a t , fo llo wed by c o nv er s ion int o SBML us ing t he Heta compil er. This yielded th e corrected_AIupdat e1 model, s ho w n in Figure 3B and Figure 5 . The corr ec t ed model s er ved a s the expert-validated reference for the nex t i tera t i o n of the AI- QSP w o r kflow. Based on the observed deficiencies , the pr ompt s and workf l o w instr uc t ions dr ivi n g t he AI- QSP framewor k wer e r efined. A f ter this ref inement step, th e f ramework gener a t ed a AIupdate2 mo del who s e structu re and code w er e c o n s istent with t he man ually c or rected implemen t atio n. This r es u l t demonstrat es t hat the i ter at ive ex pert-in - the -l oop wor kflow impr ov ed t he technical qu a lit y of th e AI-gener ated m echanistic model and enabled convergence toward an expert -vali d a ted updated model s tr ucture. 3.3 Structural analysis of the updated CAR-T QSP model C o m pa ri s o n of t h e f i n a l AIupdate2 mod el with t h e original AIcore model s ho w ed t hat the AI-QSP workflow int roduced biological ly meaningful str uc t ural change s co nsistent wi t h the requ es ted m ec hanisms . The k ey struc tu ral ex t ens ion s were th e expl icit represe n tation of an exhaus t ed CAR- T stat e, incor porat i o n of antigen-negativ e tum o r cells , and regulat ory pro c es s e s a ssociated with che c kpo i n t modulati on and ant igen es cape. The s e addit ions expanded the mechanistic scope of th e m odel beyond the or ig inal core r e presentation and enabled si m ulation of mo r e real istic C AR-T tr eatm ent d y n a m ic s under r es i stance c o nditions. At t he variabl e lev e l, t he updated mo d el differed fr om the original m odel b y int roducing new s t ate variables r epresenting exhaus t ed CAR- T cells and ant igen-negative tumor c ells, while retain ing th e shared var i ables descr i b i n g ef fector and memory CAR- T c ells, antigen- positive tumor c ells, and c yt o k ine con centr ations. At th e p r ocess level , t he updated mod el intr oduced transitions between ef f ec t or and exhaus t ed CAR- T stat es and bet w een ant ige n - positive and antigen-negative tum o r states. In c o ntrast, some featur es of the original core model, such as tis s ue m i gr ation pro c es s e s, were not r etained in the same f or m in th e updat e d im pl ementa tion . Comp a r is on o f the rate-law str ucture fu rther s howed t hat the upd a t ed mo del incorpor ated mechanistic t erms for exhaustion, an tigen esca p e, and b ystander -mediate d killing, while preserving the gener al str ucture of p r olifera t iv e and cytot oxi c pr oc es ses fr om the orig inal model. Togeth er, these observations i n dic at e t hat the A I-QS P wor kflow s u c ce ssfully tr ans f ormed th e original c or e mode l i nto an extended m ec hanistic repr es e nt ation ca p able of describing ke y r es i stan ce mechanisms relevant to C A R-T t herapy. was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint 3.4 Calibration of the updated model under the Triple Combination scenario The ability of AI-QSP to calibrate a QS P model was explor e d in accordan ce with workflow presented at F ig 1. Initially, updated ver s ion of QSP model AI update2 wa s ap pli ed f o r cali b ration against synt hetic datas et . However, we have fai le d t o cal ibr ate the version of t he model, pot en tially, due to some s t ructur al li mit a tions of imp lemen t ed mec h a n i sms . Then, we have allowed AI - Q SP t o a d a p tivel y change right h a n d sides of the model in such a way to allow a p propr ia te c alibr ation agains t c ho sen s y n thetic dataset s. A s a re sult we have c o me to AIupdate3 ver sion of t he Q SP model. This versi o n is des c r ibed in S u ppleme ntar y m a t erials S1. The c alibr ated performance of the u pdated mode l AIupdate3 wa s eva luat ed us ing t he Triple Combination scenario , w h i ch s imultaneously inc luded C AR-T exhaustion, PD- 1 checkpo i n t block ade , ant ig en es cape, and by s tand er-mediated k illing. This int egrated s c enar io w as cho s en as the most stringent test of the A I-QS P f ramework becau se it combines the principal biologic al mechanisms expe cted t o s hape CA R-T ther apeutic response and re sis t a n ce. Figure 6 illus t rates the fit ted tum or and C AR -T tr ajector ies , while Table 2,3 s um marizes the par ameter estimat es obt ai n ed fro m t he au toma t ed opti mization pro c edur e. U sing synt hetic b enc hm ar k data generated f rom the AIcore model o v er a 28- day (nine ti m e-po int) time h orizon, th e A I -QSP framewor k per form ed au tomated paramete r optim i zation of th e updated model AIupdate3 (21 r eaction s , R1–R 21) u s ing a 19-par ameter multi - phas e (comp ris ing 7 pr imary kinetic parameter s listed in Table 2,3, plus 12 c ytokine sub sy stem param eters) L- BFGS - B s t ra tegy. The fitted model r epr oduced th e benchmark dynamic s fo r tot a l t u mor bu rden, circulating effector CA R-T cells , and c yto k ine concentrat ions with high f i d e lit y acro ss the 28-day po st-infu s ion window. Visual c om parison of mod el t rajectories and digitised be nchmark ob servations [32] showed c lose agreement , with the A Iupdate3mo del a chieving a mean log-RMSE of 0.132 acro s s s ix s i mult aneous ly fitt ed va r iabl es . Per -variable log-RMS E values ar e repor ted in the Goodness - of-Fit row of Table 2, 3. T w o var i ables — exhau sted CA R-T cel ls (log-RM SE = 0.0 85) and IL - 6 ( log-RMSE = 0.067) — met th e p re-speci f ied a ccur ac y thr es hold of log- RMSE < 0.10. The optim i sed AI update3 m odel c apt ured the character is t ic behavior expe cted und er the Triple Com binati on configurat ion. Tumor bur den declined in response t o CAR- T expansion mediated by mass-a c t ion k illing (R10) a n d bystand er cyt otoxici t y on antigen-esc ape cells (R14) . L onger -ter m dynamics ref l ected t he competing eff ects of exhaustion (R4, PD -1 modulat ed), chec kpoint res c ue (P D1_ relie f ass i gnmen t rule), and ant ig en e s cape (R11). The CAR -T trajector y reprod uc ed t he ex pec t ed ex pansion–contr ac t ion patter n, and cyt oki n e dynamic s (IL-6, IL-10 , IF N-γ; rea ctions R 15–R20) rem a ined consis t ent with the timing and magnitud e of th e benchmark immu n e response . Table 2 — Model species, parameters, initial conditions, and units (SBML representation) Category Symbol / ID Description Initial Value Unit Species was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint Category Symbol / ID Description Initial Value Unit Tumor B_pos Antigen-positive tumor cells 1 × 10/i4 cells Tumor B_neg Antigen-negative (escape) tumor cells 1 × 10/i4 cells CAR-T Cells CARTE Effector CAR-T cells 1 × 10/i4 cells / µL CAR-T Cells CARTM Memory CAR-T cells 1 × 10/i4 cells / µL CAR-T Cells CARTE_EX Exhausted CAR- T cells 1 × 10 /i4 cells / µL Cytokines IL-6 Interleukin-6 1.0 relative units Cytokines IL-10 Interleukin-10 0.5 relative units Cytokines IFN- γ Interferon- γ 1.5 relative units Drug aPD1 Anti-PD-1 checkpoint inhibitor 0 (variable) µg / mL Kinetic Parameters Tumor Growth r_Bpos Growth rate of B_pos cells Fitted day ⁻ ¹ Tumor Growth r_Bneg Growth rate of B_neg cells Fitted day ⁻ ¹ was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint Category Symbol / ID Description Initial Value Unit Tumor Killing k_kill CAR-T-mediated killing of B_pos Fitted day ⁻ ¹ Tumor Killing k_byst Bystander killing of B_neg by CARTE Fitted day ⁻ ¹ Tumor Escape k_loss Antigen loss rate (B_pos → B_neg) Fitted day ⁻ ¹ CAR-T Prolif. k_prolif CARTE expansion rate constant Fitted day⁻ ¹ CAR-T Prolif. $K_{prolif}$ Half-saturation for tumor-driven proliferation Fixed cells CAR-T Death k_death Natural CARTE death rate Fitted day ⁻ ¹ CAR-T Exhaust. k_exh Rate of CARTE → CARTE_EX conversion Fitted day⁻ ¹ CAR-T Memory k_mem CARTE → CARTM conversion rate Fixed day ⁻ ¹ Checkpoint E_max_PD 1 Max fractional reduction of k_{exh}$ by aPD1` Fitted dimensionless Checkpoint IC_50 aPD1 concentration for Fixed µg / mL was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint Category Symbol / ID Description Initial Value Unit 50% effect Simulation Settings t_end Simulation duration 28 days dt Integration time step 0.01 days Table 3 — Best-fit parameter estimates — Mechanistic Reaction-Based model (19- parameter) Parameter Description Lower Bound Upper Bound Best-Fit Value Unit k_prolif CARTE expansion rate constant 0.1 2.0 0.95 day ⁻ ¹ k_exh Effector exhausted conversion rate 1.0 $\times $ 10 ⁻ /i4 1.0 $\times $ 10⁻ ² 8.6 $\times $ 10⁻ ³ day⁻ ¹ KBC k_kill CAR-T– mediated tumor killing rate 1.0 $\times $ 10 ⁻ ³ 10.0 1.0 $\times $ 10⁻ ³ (cells·day)⁻ ¹ k_loss Antigen- loss rate (B_pos 1.0 $\times $ 10 ⁻ /i4 1.0 $\times $ 10⁻ ¹ 2.9 $\times $ 10⁻ /i4 day⁻ ¹ was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint Parameter Description Lower Bound Upper Bound Best-Fit Value Unit $\rightarrow $ B_neg) k_byst Bystander killing rate on B_neg 1.0 $\times $ 10 ⁻ ³ 10.0 1.7 $\times $ 10 ⁻ ³ day⁻ ¹ dE_scale Effector death scaling factor 0.2 3.0 0.33 dimensionles s Emax_PD1 Max fractional reduction of $k_{exh}$ by aPD1 0.3 0.99 0.35 dimensionles s Goodness- of-Fit Goodness- of-Fit (per variable): Total Tumor 0.224; Effector CAR-T 0.140; Exhausted CAR-T 0.085 /i4 ; IL- 6 0.067/i4 ; IL-10 0.116; IFN- γ 0.159. Mean log- RMSE (6 variables, Mechanistic ) — — 0.132 log /i4/i4 units was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint Figure 6 — Mechanistic Reaction-Based model calibration fit: six variables, Triple Combination scenario AIupdate3 model (21 SBML reactions R1–R21, 19-parameter calibration, MA v11 best-fit) trajectories for six simultaneously calibrated variables under the Triple Combination scenario (exhaustion R4, PD-1 relief rule, antigen escape R11, bystander killing R14). Panels: (A) Total Tumor (B/i1 +B/i1 , reactions R8–R14), (B) effector CAR-T (R1– R5), (C) exhausted CAR-T (R4, R7), (D) IL-6 (R15–R16), (E) IL-10 (R17–R18), (F) IFN- γ (R19–R20). Filled circles: digitised benchmark data (Kimmel et al., 2021 [32]). Solid lines: AIupdate3 best-fit trajectories. Per-panel log-RMSE values are shown in panel legends; /i1 denotes variables meeting the log-RMSE < 0.10 accuracy threshold. Mean log-RMSE = 0.132 across six variables. All axes: log /i1/i1 scale. 3.5 Quantitative goodness-of-fit Quantitative agreement between the calibrated model and the synthetic benchmark data was assessed using the log-transformed root mean square error ( log-RMSE ). The f in al calibrated solution achieved a mean log-RMSE of 0.132 (AIupdate3 model) . Per-variable log-RMSE values ranged from 0.067 (IL-6, /iN1) to 0.224 (total tumor burden), as detailed in Table 3. Two variables did not meet the pre-specified accuracy threshold: IFN-γ (log-RMSE = 0.159) and total tumour burden (log-RMSE = 0.224). The relatively higher error for total tumour burden reflects a structural trade-off between the antigen-escape rate (R11) and bystander killing (R14) that limits simultaneous optimisation of both early decline and late- phase dynamics. The IFN-γ fit (log-RMSE = 0.159) is constrained by the semi-decoupled cytokine source term structure in the current model; a more flexible IFN-γ production rate law is identified as a direction for AIupdate4. These are documented as directions for future model extension. Taken together, these results demonstrate that the AI-QSP framework can: (1) reconstruct and extend a published CAR-T QSP model, (2) iteratively improve the technical correctness of AI-generated model structure through was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint expert f e edbac k, and (3) calibrate t he upda ted mo de l t o re pr oduce the dynamic s of t he original refer ence m ode l under a challenging integrat ed mec h a n i stic scenar i o . Th i s support s th e f eas ibility of AI- assi sted model evo lutio n as a potent i al wor k f low f or mechanistic ph a r mac ology m odeling. 3.6 A Sobol’ variance-based GSA was perfor m ed on all 19 c alibr a ted param et ers us ing a Saltelli quas i-r ando m sample of N = 1024 b ase vec t ors (40 ,960 tot al m odel ev aluat i o ns ; Supplement ary S3). Fi r s t -ord er (S 1 ) an d t otal -or der (S T ) S ob ol’ indic es were computed u s ing the J a n s en (1999 ) es t i m ator with 95% boot s t rap confidence inter v als (n = 5 00 replicates). The GSA f ulfi lls th e m ode l-rob us t nes s requir em ent o f A SME V&V 40 b y qua ntifying out put sen s itivity to ind i v idual par a meters and their i n teractions, and s u pport s pr i n c ipled mod e l redu c t i o n by iden tifying five n on -inf lu ent ial par am eter s (S T < 0.05 acros s all out puts) that were fixed at b est-f i t va lues for the s ubs equ e n t prof i le - l ikelih ood analysis. Key findings : tum or bur d en wa s most sens it i ve to k kill , k byst , a n d K DX ; c yto k ine ou tputs were dominated by their r espec t iv e pr oduction (α ) and d egradation ( k deg ) parame ter s . Full numer ic al results, heat-m ap v isualiz at ions , and r eproducibility mater ial s are pr ov i d ed i n Sup p lementary S3. 3.7 Following GSA-guided m odel redu ction, pr ofile -likeliho od ident ifiab ili t y was as se ssed for the 14 active paramet e rs us in g sin g le- pass marginal prof i les eval u at ed acros s each paramet er’s full bound r a n g e ( S up ple mentar y S4). Two objective-fu nc t ion thresholds were applied: ΔJ = 0.010 ( i n ner confidence inter v al ) and Δ J = 0.020 (outer c on fid ence i n terval), yielding three iden t ifiab ili ty c la s s es . T en of 14 paramet e rs w er e c las s if i ed as Cla ss A (ident i fiable ), wi t h fin ite boun d ed c o nfidence i n terva ls; four par amete rs were Class C (no n- identifi able) : k byst , K DX , α IL6 , and k prolif . Thes e r epresent stru ctu ral li m i tations of si n g le- datas et cali b ration and are designated as priorit y t argets for B a yesian hierarchical r e-estimation fr om c linical data in Stage 2. An id enti f i abili t y r ate of 71% (10/14) for a 21- r eaction model cali b rated fr om a s ingle benchmar k d at as et is c o ns i stent with p ublis hed expectations for QSP models of c om parable c o mplex it y. C onfidence interval statistic s and per -parame t er classific at ion s ar e p rovided in Suppl e mentar y S4 (Table S4.1) and visualise d in Figure 7. was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint Figure 7. Profile-Likelihood Confidence Intervals — 14 Active Parameters. D ark bar: ΔJ = 0.010 ( i n ner CI); light bar : ΔJ = 0.020 (ou ter CI); d ot: best-fit v alue ( log s ca le) . B l u e: Class A — Identifiab le (10/14); r ed: Class C — Non-ident if iable ( 4/ 14) . P a r ameters k_by st, K_D X, α_IL6, and k_prolif were classified as C la s s C , representing pr iority tar gets f or Bayesi an r e- estimation from clinic al data. Sour c e dat a: S up plementary S4 (Table S4.1). 4. Discussion This stud y pr e s ent s a pr ototype AI-assisted quantitative systems pharmacology (AI-QSP) fr amework designed to suppor t auto mat ed reconstruction , ext ens ion, and calibration o f mechanistic ph a r ma cology m odels. T he pr i m ary objective o f the work was t o eva luat e whether s uch a fram ew o rk can r eli abl y tr ansform a liter a ture - deriv ed mod el into an updat e d m ec han i stic repr es ent ation and repr oduce the dynamics o f the refer ence s y stem thr ough autom a t ed c alibr a tion. To test th i s capabili t y, the workflow was applied to a pu bli shed CA R-T cel l t herapy Q SP model. The AI- Q SP fr amework first r e c o n s tr ucted the lite r atur e model into a standardized SBML repr es en tation ( A Ic or e). It t hen generated str uc t ural model upda t es based on biological ly mot iv at ed pr omp ts descri bing ke y res i stance mechani s m s , in cluding T-cel l exhaus t ion, che ckpoint r egulation through PD -1/PD -L1 s i gnaling, and tumo r antigen esc ape. These mec h a n ism s represent widely recognized drivers of tr ea t ment r es istance in CAR- T ther apy a n d theref o re pro v ide a bio lo gically m e an i n g f ul tes t case for aut omated mod e l evoluti on. The results demonstr a t e that t he AI- QSP wor kflow can su c cessfully gener ate mechanis t ic model extensions an d i t erat ively impr ove th e technic al corr ec t ne ss of the res ult ing S B ML model through an expert - in-th e- lo op val idat ion p roces s . The fir s t A I-generat ed model updat e captu red the in t ended b iologi ca l pr oce sses but c ont ained s everal tec hnical inconsistencie s t ypi c al of aut omatical ly gener ated code. Aft e r r efine m ent of th e p r ompt s and corr ec t ion of these is sue s, t he s econd AI -generated mo de l v er s ion r ep roduced t he expert- v a lidat ed imple mentation. Thi s iter ative i mprove ment highli ght s an impo rtan t pro perty of t he proposed fr amework: the abili ty to comb i n e aut oma ted model g en eration with expert va lidati on t o pr ogr es sively converge towar d a t echnic ally correct and biological ly con sistent mechanis t i c m odel. The cal ibr ation phase of the study revealed an impo r tant p r ac t ic al insight: t he fir s t structu rally exten ded model, AIupdate2, c ou l d n ot be s u c ce ssfully cal ibr ated agains t the s y nt hetic benchmark data. Calibrat io n of AIupd ate2 c o nv er ged to p oor s ol ut ions, likely due to str uc t ural limi tation s in the r ight-h and side rate expr es sions int roduced dur ing t he AI- guided extension s t ep — in par tic u l ar , parameter in teraction s between t he ex h a u s tion tr ans it i o n (Rxn_CA RTE_Exhau stion) and t he PD-1 r el i ef a ssignment ru l e creat ed near- redu ndant contr i b ut ions that r ende re d t he objec t iv e f unc t ion landsc ape ill- condition e d . T h i s failure is itsel f a n inf or mat iv e out com e: it demonstr ates t hat AI- ge n erated str uc t ural extensions , even when technic ally cor rect at the SBM L l evel, may not be i mmed i at e ly cali b ratable and requ i r e fu rt her adap t i ve r e finem ent. In r es pon s e, the A I- QS P f ramework was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint was r e-enga ged to adaptively modi fy the r ate ex press ion s of the pr obl ema tic proces s e s, pro duc ing A Iupdate3. This v er s ion inc orp orates revis ed kinet ic for mula tion s for the exhaus t ion and by s t ander killi n g pat hways (det a iled in Supplemen t ary S1) that im prov e the conditioning of t he optim isation land s ca p e. A Iupdate3 was then s u cce ssfully cali b rated under the “Tripl e Combinat i o n s cenar io” — simultaneou s ly in corpor ating ex haustion , checkpo i n t block ade , ant ig en es cape, and by s tand er-mediated c yt otoxic it y. This integrat ed s c enar io represents the most str ingent test of t he framewor k because it ac t iv at es all four resis t an c e m ec hani sms simult a n eous ly. The autom ated paramet er op timi sation r ec o v er ed the benchmar k dy n amics with cl o s e a greement ac r oss tumo ur bur den, CAR - T cell kinetics, and c ytokine respon ses (mean log-R MSE = 0.13 2; s um log- RMSE = 0.792 a c r os s si x out puts), demonstr ating that t he ite rat iv e A I-Q SP r efinement cycle can navigate initi al ca libr ation failures and converge on a quantitati vel y con sis t e n t s o l u tion. From a met hodo logical p erspec t iv e, t he k ey contr ibution o f t his wor k is the demon s tr ation that AI-assisted model evolution can be integrated into a reproducible systems pharmacology workflow . Tr aditional QS P m odel dev elopm en t rel ies heavi l y on manu a l inter pre tat i on of bi ological lit e ratur e and iterati ve model m od ifica t ion by d omain experts. Whil e t his pr oce s s produce s hig h ly de tailed m o dels , it is l ab or-in t ens ive and often d i fficult to repr oduce ac r oss modeling teams. Th e AI-QSP fram ew o rk address e s this li m itati on by intr oducing a n aut omated knowledge-int erp r e t at ion layer capable of t r a n s lating text ual biological description s into candidat e m odel c om ponents. W h en c om bined wi t h standardized SBML repr es ent ations and automat ed c alibra t ion t ools, this ap proach provides a scalable s t rategy for maintaining and updating complex mechanis t i c mod els as new biological knowledge become s av ailable. The fr a m ew o rk als o aligns with cur rent r egulatory expectations for M odel - Inf ormed D rug Development (MID D) wor kflows . Regulat ory agenci e s s u c h a s the U.S . Food and D rug Adm i n i str at ion and the Eur opean Me d i cines Agen cy em pha s ize transparen c y, repr oducibili ty, and traceability in co mputation a l m ode ls used for d rug development. B y encoding all m odels in SBM L and mai nt ai n i n g a str uc t ured r e cor d of mode l updat es , cali b ration datasets, and paramet er e stimates, the AI- Q SP wor k f l o w f a cilitates the generat i o n of aud itable model ing pipelines that could support regulator y-grade applica t ions in the future. 4.1 Limitations and future work Several li mitat ions of the pr es en t s t u dy s hou l d b e ackno wledged. First, the validation experimen t s were per formed us ing synthetic benchmark data gener ated fr om the ref erence mode l. This desi gn provides a cont rolled t est of t he c om putat ional wor kflow becau s e t he true sy s t em dynamics ar e known. Ho w ever , it does no t directl y a ssess the biological pr edictive p er for mance o f t he updat ed mode l. App lication of t he framewor k to independent ex pe r ime ntal o r c linical datas et s will be neces sary to evaluate its abil it y t o generat e b i o l o g ically predictive mode ls. Second, although th e A I-QS P f ramework can prop os e stru c t ural model upd at es , th e cur rent implemen t atio n still requ ires expert v alidation t o ens ur e technica l corr ectnes s and biological plausibility. This expert -in-the-loop desig n is intent iona l an d ref l ects the curr ent was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint state of mechanistic modeling, where domain exper t is e rem ai n s ess ent ial . Fut ure work may explore i mpr oved a utomated v alidat i o n procedures, including rule-based con s isten c y check s, SBML va lidat ion t ools, and form a l mo de l-ve rif ication appr oac hes. Finally , t he present study fo cu ses on a relativ ely compact CAR- T QS P model. Ap ply ing t he fr a m ework to larger m u ltisc ale s ystems pharm ac olog y models involv ing mult iple tis sues , c ell populat i o ns , and signali n g pat hwa ys wil l pr ovid e a m ore compr e h ens ive e valuation o f its s c alability. The prop os ed A I-Q SP fr amework complement s r ec ent effort s t o integrate art i f ic ial intel ligence wi t h mechanis t ic phar mac o l o gy mo deli n g and m a y a c celer ate model - infor med drug developm en t workflo ws for emer ging cell and gene ther apie s. 4.2 Regulatory Compliance and Model Credibility Assessment The present stud y wa s d e veloped with regulator y t ransparenc y as a primar y de sign objective. The c o mplete regulato ry a nd complianc e do c um entation is dis t ributed across four supplementar y m aterials: S up plemen tary S1 pr ov ides the fu l l mat h ematica l s pecification of the r eac t ion-based S B ML model (21 r eactions, 9 s p e c ies , 19-paramet er cali b ration); Supplement ary S2 p resents t he s t ru ctur ed Model Cr edib ili t y As se s s ment , com pleted ICH M15 A ssessment Table, and M IDD su bmissi o n roadmap; Supplemen tary S3 repor ts the Gl o bal Sensitivity Analysis (Sobol’, N = 1024 ); and Supplementar y S4 det ai ls the Pr ofile- Lik elihood identif iabili ty analy s is for 1 4 active parameter s . Toget her, these document s con s t itute t he ev identiary pack a ge r equired fo r regulator y evaluat ion of th e model at t he curr ent explorato r y dev elop men t stage. The ICH M 15 G uideline on General Princ iples for M odel-Inf o rmed Drug Development [33] pro v ides the overa r ching h a r monized framewor k f or MI DD ev idence asses sment a cr oss FDA, EMA , a n d other IC H memb er regions. U nder M15, model evaluat ion r equ ir ement s ar e cali b rated to M ode l Risk — defined as the combination of M ode l Influence (t he i ntended weight of mod el ou tcomes in decisio n-making) and Consequen ce of Wrong Deci si o n. F o r the present pr e-clinic al pr oto type, M odel Ris k i s a sses sed a s Low-to-M edium: mod el in f luenc e is low, as simulation out puts s upp l em e nt rat her than r eplac e clinica l evidence at this developmen t stage, and th e con sequence of a wrong deci s ion is limited gi ven the explorato ry c on text. Model Imp a ct is rated as Medium , reflecting the no vel AI- a ssi s ted model evo lution app roach that departs from convent i o nal m anual QSP workflows. The completed M 15 Ass e s s men t Table, i n c luding pr e-spec ified t ec hn i cal criter i a and ou tcome of the M IDD ev idence ass e s s men t, is pr ovi d e d in Supp l e m entary S2. The tech nical cr i t eria defined a pri o ri wer e: mean log-RM S E ≤ 0 .15 acro ss al l fit ted outputs; s t ru ctur al identifi abil it y of ≥ 70% of active para m eters by profile- likelihood; and GSA -confir med paramet er relevance (to t al-o rder Sob ol’ index S T ≥ 0.05 f or a t least one out p ut). Al l th re e criter ia wer e s atisfied by the M A v11 c alibr a t ion s olut ion ( mean log-RM SE = 0 .132; 10 / 14 paramet ers Class A identif iable; 14/19 paramet ers w it h S T ≥ 0.05; s ee Supp lement ary S3). Mod el cred i b i l it y was a sse ssed following the risk-infor med ve rif ication and validation fr a m ework of ASM E V & V 40-2018, a s det ail ed in Supp l em en tary S2. Ver i f ic at i o n c onf irmed that the SBML m odel (L evel 3 Version 2) corr ec t ly implem e n ts t he mathe ma t ic al equat i o ns document ed i n Sup plemen tary S1, wi th mass - bala n c e and stoichiometr y v alidated pro g r ammatically u s ing libSBML and Tellurium. V a lida t ion c om prised thre e compo nents: (i) was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint goodnes s - of-fit ; ( ii) global sensitivity analysis e stablishing the relative in flu ence of all paramet ers ac r oss all o utput s (Supplemen tary S3); and (iii) profil e-l ikeli h oo d identi fiabil ity analy s is c har a c t erizi n g pr ac t ic al par am eter uncert ain ty for t he 14 ac t iv e p ar ameters (Supplement ary S4). The unified para m eter credibi lity tab l e int egrating all three ev idence streams i s pr ovided in S up plem entary S2 (Table S2.4). A stru c t ured ro a d map fo r pr ogress ion f rom t h e c ur rent p re -clinic al pr oto ty p e t o a regulato ry-submission-r eady MIDD p ackage i s pr ovided in Supplemen tar y S2. The r oadmap identifi es thr ee dev el o pmen t stages a ligned with ICH M15 and ASM E V&V 40. Stage 1 (curr ent work, complet e): pre-clinical p roof o f concept wit h synthet ic benc hmar k data, SBML model encoding, GS A , and profile-likel ihood iden t if iabil ity — all Stag e 1 deli ver ab les are document ed in Supple mentar y S1–S4. Stage 2: integr ation of patien t -le vel clinical P K/PD and biomar k er data with Bayesi an hier archica l par ameter estimat i o n to r esolve the four Clas s C non- i d entifiable pa ramet e rs; prepar a t ion of a Model A naly sis P l an (M AP) per ICH M15 Section 4.1. Stage 3: exter nal validation against independent c lin ical cohor ts, full Mod el A na ly si s R eport ( MAR) per ICH M15 Section 4.2, and prep a r ation of r egulatory submission document ation in Comm o n Tec hnical Do cument form a t . 5. Conclusions In summary , th i s work demon str ates the feasi b i li t y of i n tegrating ar t ificia l i nt ell igence with mechanistic ph a r ma cology m odeling t hrough an AI-QSP workflow capable of reconstr uc t ing , ex t ending, and calibrat ing Q SP models in a repro duc ible c omput ationa l environm e nt. By combining large language m odel–bas ed kno wledge interpretat ion with SBML-based modeling too l s and auto m ated opt im i zation, t he framewor k provides a structu red appro a ch for iterativ e model evolu t ion whil e p res er ving m ec hanistic tr ans par ency. The se c apabilities position AI-QSP as a pr omisi n g enabling technolog y for s c alable, repr oducible model ing work flows in s y s t ems ph a r macolog y and m odel- informed dru g developm e n t.

1. Santomasso, B.D. et al. T-cell exhaustion in CAR T-cell therapy. Nature Reviews Clinical Oncology 20, 455–473 (2023). 2. Haradhvala, N.J. et al. Antigen escape as a mechanism of resistance to CAR T-cell therapy. Nature Medicine 28, 1570–1580 (2022). 3. Liu, E. et al. Addition of anti-PD-1 checkpoint blockade augments CAR T-cell efficacy in preclinical models. Clinical Cancer Research 29, 3012–3023 (2023). 4. U.S. Food and Drug Administration (FDA). Model-Informed Drug Development Pilot Program. https://www.fda.gov/drugs/development-approval-process- drugs/model-informed-drug-development-pilot-program (accessed 2025). was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint 5. European Medicines Agency (EMA). Guideline on the qualification and reporting of physiologically based pharmacokinetic (PBPK) modeling and simulation. EMA/CHMP/458101/2016 (2018). 6. Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). 7. Varadi, M. et al. AlphaFold Protein Structure Database: massively expanding structural coverage of protein-sequence space. Nucleic Acids Research 50, D439–D444 (2022). 8. Vicini, P., van der Graaf, P.H. Systems pharmacology meets AI: advancing model-informed precision medicine. Nature Reviews Drug Discovery 22, 267– 284 (2023). 9. Polasek, T.M., Rostami-Hodjegan, A. Regulatory perspective on model- informed drug development (MIDD): bridging QSP and clinical translation. CPT: Pharmacometrics & Systems Pharmacology 12, 452–463 (2023). 10. Hardiansyah, D. & Ng, C. M. Quantitative Systems Pharmacology Model of Chimeric Antigen Receptor T-Cell Therapy. Clin. Transl. Sci. 12(4), 343–349 (2019). 11. Wherry, E. J. & Kurachi, M. Molecular and cellular insights into T cell exhaustion. Nat. Rev. Immunol. 15, 486–499 (2015). 12. Ribas, A. & Wolchok, J. D. Cancer immunotherapy using checkpoint blockade. Science 359, 1350–1355 (2018). 13. Majzner, R. G. & Mackall, C. L. Tumor Antigen Escape from CAR T-cell Therapy. Cancer Discov. 8, 1219–1226 (2018). 14. He, J. et al. The bystander killing effect of CAR-T therapy: a brief review. J. Cancer 13, 3234–3240 (2022). 15. Santurio, D. S. et al. Mathematical modeling unveils the timeline of CAR-T cell therapy and macrophage-mediated cytokine release syndrome. PLOS Comput. Biol. 20, e1011867 (2024). 16. Singh, P ., Sinko, P . J. & Lyou, Y . Development of a quantitative relationship between CAR-affinity, antigen abundance, tumor cell depletion and CAR-T cell expansion using a multiscale systems PK-PD model. mAbs 11, 1269– 1280 (2019). 17. Sun, D. et al. CAR-T cell therapy: efficacy in the management of cancers. International Immunopharmacology (2024). 18. Ferreri, C. J. & Bhutani, M. Mechanisms and management of CAR T toxicity. Front. Oncol. 14, 1396490 (2024). was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint 19. Abou-el-Enein, M. et al. Advances in CAR T cell therapy: antigen selection, modifications, and current trials for solid tumors. Front. Immunol. 15, 1489827 (2024). 20. Lu, Y . & Zhao, F. Strategies to overcome tumor relapse caused by antigen escape after CAR T therapy. Mol. Cancer 24, 126 (2025). 21. Wang, H. et al. Quantitative Systems Pharmacology Modeling in Immuno- Oncology: Hypothesis Testing, Dose Optimization, and Efficacy Prediction. Handb. Exp. Pharmacol. (2024). doi: 10.1007/164_2024_735. 22. Melenhorst, J. J. et al. Decade-long leukaemia remissions with persistence of CD4+ CAR T cells. Nature 602, 503–509 (2022). 23. Xiao, X. et al. Mechanisms of cytokine release syndrome and neurotoxicity of CAR T-cell therapy and associated prevention and management strategies. J. Exp. Clin. Cancer Res. 40, 367 (2021). 24. Melenhorst JJ, Chen GM, Wang M, et al. Decade-long leukaemia remissions with persistence of CD4+ CAR T cells. Nature. 2022;602:503–509. doi:10.1038/s41586-021-04390-6. (2023). 25. Hucka M, Finney A, Sauro HM, et al. The systems biology markup language (SBML): a medium for representation and exchange of biochemical network models. Bioinformatics. 2003;19(4):524–531. doi:10.1093/bioinformatics/btg015. 26. Choi K, Medley JK, König M, et al. Tellurium: An extensible Python-based modeling environment for systems and synthetic biology. Biosystems. 2018;171:74–79. doi:10.1016/j.biosystems.2018.07.006. 27. Metelkin E, Borisov I, Chudaev M. Heta Compiler: A software tool for the development of large-scale models of quantitative pharmacology. J Open Source Softw. 2022;7(74):4400. doi:10.21105/joss.04400. 28. Gizzatkulov NM, Goryanin II, Metelkin EA, et al. DBSolve Optimum: a software package for kinetic modeling which allows dynamic visualization of simulation results. BMC Syst Biol. 2010;4:109. doi:10.1186/1752-0509-4-109. 29. Somogyi ET, Bouteiller JM, Glazier JA, et al. libRoadRunner: a high performance SBML simulation and analysis library. Bioinformatics. 2015;31(20):3315–3321. doi:10.1093/bioinformatics/btv363. 30. Bordbar A, Monk JM, King ZA, Palsson BO. Constraint-based models predict metabolic and associated cellular functions. Nat Rev Genet. 2014;15(2):107–20. doi:10.1038/nrg3643. Nat Rev Genet. 31. Sorger PK, Allerheiligen SRB, Abernethy DR, et al. Quantitative and Systems Pharmacology in the Post-genomic Era: New Approaches to Discovering Drugs and Understanding Therapeutic Mechanisms. NIH White Paper. 2011. was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint Available at: https://www.nigms.nih.gov/training/documents/syspharmpaper.pdf. 32. Kimmel GJ, Loc ke FL, Altro c k PM . The r oles of T cell c o mpetition and stochastic extinction even ts in c h i m eric a n tigen receptor T cell ther apy. Proc R Soc B. 2021;28 8(1 947):2 0210 229. doi:10. 10 98/rspb .2021 .0229 . 33. ICH M15 Har monised G uideline: General Principles f or Mo del -Infor m ed Drug Development. Inter nat ional Council f or H arm o nis ation of Tec hnical Require m ents for Pharm ac eut ic al s for H uman U se. Adopted 29 Januar y 2026. A v ailable at: htt ps : / /www.i ch.org/pa ge/ m 15. was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (whichthis version posted April 7, 2026. ; https://doi.org/10.64898/2026.04.05.716273doi: bioRxiv preprint