Optimization of first-line treatment selection in advanced pancreatic adenocarcinoma using artificial intelligence Authors: Sebastian Cole, Paul Campitelli, Brian Grieb, Igor Astsaturov, Daniel Von Hoff, Harshabad Singh, Michael J. Pishvaian, Tanios S. Bekaii-Saab, Peter J. Hosein, Philip A. Philip, Anthony Helmstetter, Todd Maney, Jennifer R. Ribeiro, James Hamrick, Daniel Magee, David Spetzler Published: 2026 ABSTRACT Improved clinical outcomes are reported for patients with advanced pancreatic adenocarcinoma (PDAC) treated with first-line FOLFIRINOX/NALIRIFOX, but elderly patients with comorbidities are more often treated with gemcitabine/nab-paclitaxel (gem/nab-p). There is currently no comprehensive method to optimize first-line treatment selection between these regimens. We developed a propensity score matched, transcriptomic-based AI model using 2202 molecularly-profiled PDAC specimens to provide clinically relevant treatment recommendations and prognostic information. In a testing dataset of patients predicted to have superior outcomes on first-line FOLFIRINOX, time-to-next-treatment (TTNT) and overall survival (OS) were significantly longer for patients treated with FOLFIRINOX first (HRs = 0.55 and 0.48, respectively, p < 0.001). Patients recommended for gem/nab-p treatment had similar outcomes on either treatment, but a subset had improved outcomes on gem/nab-p. Approximately half of patients had received the opposite therapy from the model recommendation. Applied clinically, this model could improve treatment decision-making in advanced PDAC in the first-line setting. Introduction Pancreatic adenocarcinoma (PDAC) remains a highly fatal disease with poor clinical outcomes, but proper selection of treatment regimens for advanced disease can extend overall survival. The current first-line standard-of-care for advanced/metastatic PDAC includes FOLFIRINOX (5-FU, leucovorin, irinotecan, and oxaliplatin), NALIRIFOX (includes liposomal irinotecan) or gemcitabine/nanoparticle albumin–bound (nab)-pax (gem/nab-p)1–4. The American Society for Clinical Oncology (ASCO) recommends first-line FOLFIRINOX for patients with a good performance status and favorable comorbidity profile5, and National Comprehensive Cancer Network guidelines exclude FOLFIRINOX from the preferred regimens for patients with intermediate performance status6. In practice, this means that older patients are more often given gem/nab-p instead of FOLFIRINOX7–10. While the pivotal clinical trials supporting the FDA approvals of FOLFIRINOX and gem/nab-p (PRODIGE-4/ACCORD-11 and MPACT, respectively) compared each regimen to gemcitabine alone1,2, large retrospective studies of real-world data, meta-analyses, and more recent prospective trials have offered direct comparisons of the two therapies. Some of these studies have demonstrated superior outcomes with FOLFIRINOX3,7,9,11, although conflicting results have been reported such as the phase II Pass-01 trial which reported similar efficacy of modified FOLFIRINOX and gem/nab-p12–16. In the phase III NAPOLI-3 trial, a prospective head-to-head comparison of NALIRIFOX and gem/nab-p demonstrated greater overall survival with the triplet therapy4. Overall, there is a lack of definitive evidence supporting the superiority of one regimen over the other. Additionally, FOLFIRINOX/NALIRIFOX have demonstrated different toxicity profiles than gem/nab-p, and studies investigating the use of triplet therapies in older patients are limited3,4,8,12,17. Differential outcomes to each therapy have furthermore been linked to homologous recombination repair deficiency (HRD), which supports preferential selection of platinum-based therapy18, and PurIST molecular subtypes. It has been reported that the classical subtype derives greater benefit from FOLFIRINOX, and the basal subtype—which confers a worse overall prognosis—responds similarly to either regimen19,20. However, the Pass-01 trial also reported divergent findings with regards to PurIST subtypes, with significantly longer survival for patients with the classical subtype when treated with gem/nab-p compared to FOLFIRINOX16. Transcriptomic and artificial intelligence (AI)-based signatures have furthermore shown promise in predicting sensitivity to fluoropyrimidine-based and gemcitabine-based regimens in PDAC, demonstrating the potential of such an approach21–25. Nonetheless, there is no comprehensive tumor-centric approach to guide first-line treatment strategy between FOLFIRINOX and gem/nab-p in advanced PDAC. To address this need, we aimed to develop an AI model trained on transcriptomic feature sets from a large cohort of molecularly-profiled PDAC to identify patients for whom FOLFIRINOX treatment should be prioritized in the first-line while simultaneously identifying those who would be good candidates for gem/nab-p. Implementation of this model during routine molecular tumor testing may aid first-line treatment decision-making for patients with advanced PDAC. Results Clinical and demographic characteristics Among the entire PDAC cohort (N = 2202) (Fig. 1), patients treated with gem/nab-p were older (mean 71 years) than those treated with FOLFIRINOX (mean 63 years) and had higher frailty scores, reflecting real-world clinician preferences for treatment selection. Men were slightly more likely than women to receive FOLFIRINOX (~55% versus ~45%) (Table 1). Comparing clinical outcomes between FOLFIRINOX and gem/nab-p treatment groups, TTNT and OS were both longer for FOLFIRINOX compared to gem/nab-p (HRs = 0.79 and 0.72; p < 0.0001 for both outcomes), and median OS for each group was comparable to OS durations reported in clinical trials (11.8 months and 8.9 months for FOLFIRINOX and gem/nab-p, respectively)1,2. Clinical outcomes were comparable in training and testing cohorts (Supplementary Fig. S1).Fig. 1STROBE diagram shows identification of cohort from available pancreatic adenocarcinoma (PDAC) specimens in the Caris Life Sciences database.The final cohort included 2202 PDAC specimens from patients treated with first-line gemcitabine/nab-paclitaxel (n = 870) and FOLFIRINOX (n = 1332). Propensity score matching (PSM) was conducted (Supplementary Fig. S2) prior to splitting the cohort into training and testing cohorts to control confounding factors. Only PSM-matched samples were used for training, but all samples were used for out of fold predictions.Table 1Clinical and demographic characteristics of study cohortFOLFIRINOX, total cohort(n = 1332)gem/nab-p, total cohort(n = 870)FOLFIRINOX, training(n = 932)gem/nab-p, training(n = 609)FOLFIRINOX, testing(n = 400)gem/nab-p, testing(n = 261)Mean Age (range)63.4(23–89)71.4 (30- > 89)63.3(28–85)71.1 (36- > 89)63.8(23–89)71.9 (30- > 89)Sex, No. (%)Male727 (54.6)418 (48)510 (54.7)292 (47.9)217 (54.3)126 (48.3)Female605 (45.4)452 (52)422 (45.3)317 (52.1)183 (45.8)135 (51.7)Specimen Site Type, No. (%)Prim.628 (47.1)373 (42.9)451 (48.4)271 (44.5)177 (44.3)102 (39.1)Met.704 (52.9)497 (57.1)481 (51.6)338 (55.5)223 (55.6)159 (60.9)KRAS Mutation Status, No. (%)Mut.1,229 (92.3)779 (89.5)866 (92.9)541 (88.8)363 (90.8)238 (91.2)WT103 (7.7)91 (10.5)66 (7.1)68 (11.2)37 (9.3)23 (8.8)Mean Frailty Score (range)8.33 (0.0-73.7)11.62 (0.0-97.8)8.49 (0.0-73.7)11.29 (0.0-97.8)7.97 (0.0-45.3)12.39 (0.0-70.9) STROBE diagram shows identification of cohort from available pancreatic adenocarcinoma (PDAC) specimens in the Caris Life Sciences database. The final cohort included 2202 PDAC specimens from patients treated with first-line gemcitabine/nab-paclitaxel (n = 870) and FOLFIRINOX (n = 1332). Propensity score matching (PSM) was conducted (Supplementary Fig. S2) prior to splitting the cohort into training and testing cohorts to control confounding factors. Only PSM-matched samples were used for training, but all samples were used for out of fold predictions. Table 1Clinical and demographic characteristics of study cohortFOLFIRINOX, total cohort(n = 1332)gem/nab-p, total cohort(n = 870)FOLFIRINOX, training(n = 932)gem/nab-p, training(n = 609)FOLFIRINOX, testing(n = 400)gem/nab-p, testing(n = 261)Mean Age (range)63.4(23–89)71.4 (30- > 89)63.3(28–85)71.1 (36- > 89)63.8(23–89)71.9 (30- > 89)Sex, No. (%)Male727 (54.6)418 (48)510 (54.7)292 (47.9)217 (54.3)126 (48.3)Female605 (45.4)452 (52)422 (45.3)317 (52.1)183 (45.8)135 (51.7)Specimen Site Type, No. (%)Prim.628 (47.1)373 (42.9)451 (48.4)271 (44.5)177 (44.3)102 (39.1)Met.704 (52.9)497 (57.1)481 (51.6)338 (55.5)223 (55.6)159 (60.9)KRAS Mutation Status, No. (%)Mut.1,229 (92.3)779 (89.5)866 (92.9)541 (88.8)363 (90.8)238 (91.2)WT103 (7.7)91 (10.5)66 (7.1)68 (11.2)37 (9.3)23 (8.8)Mean Frailty Score (range)8.33 (0.0-73.7)11.62 (0.0-97.8)8.49 (0.0-73.7)11.29 (0.0-97.8)7.97 (0.0-45.3)12.39 (0.0-70.9) Clinical and demographic characteristics of study cohort Using a propensity score matching (PSM) approach to analyze the effect of clinical confounders on outcomes revealed a modest effect of age and a medium effect of frailty, both of which were reduced to a small effect upon PSM (Supplementary Fig. S2). As anticipated, gem/nab-p-treated patients were more likely to have higher frailty scores, which also increased with age (Supplementary Fig. S3). Post-PSM median TTNT and OS were identical to the pre-PSM outcomes (TTNT: 7.1 months vs. 6.0 months and OS: 11.8 months vs. 8.9 months for FOLFIRINOX vs. gem/nab-p, respectively; p < 0.0001 for both outcomes). These results demonstrated FOLFIRINOX benefit over gem/nab-p even after adjustment for age, frailty, and sex (Supplementary Fig. S4), although other clinical variables not investigated could have affected this finding. PSM samples were used for model training to reduce confounders being introduced into the model. Model recommendations for FOLFIRINOX and Gem/nab-p The final model classified patients into two treatment recommendation groups: “recommend first-line FOLFIRINOX” and “recommend first-line gem/nab-p”. Each group was further stratified into a “high-molecular risk” group of patients predicted to have the worst outcomes and a “standard-molecular risk” prognosis group expected to have outcomes typical for patients with advanced PDAC. The risk groups were assigned using a 50th percentile threshold of the generated risk score (see Methods; Fig. 2). Together, approximately half of patients (n = 1105) received the opposite treatment from the model recommendation. 16.9% (n = 373/2202) of patients were treated with gem/nab-p first but were predicted to have superior response to first-line FOLFIRINOX. Conversely, 33.2% (n = 732/2202) of patients were treated with FOLFIRINOX first but were predicted to have similar or better outcomes on first-line gem/nab-p (Fig. 3).Fig. 2Schematic of AI model development.Hypothesis-driven and hypothesis-free feature selection were conducted. Hypothesis-driven features were subjected to nested 10-fold cross validation with a 90/10 train/test outer split and 90/10 train/validate inner split. Propensity score matched samples only were used for training and validation, but all samples were used for testing in the testing cohort. Model architectures evaluated were COXPH, Coxnet, ComponentWise Gradient Boosted Survival Analysis (GBSA), Gradient Boosted Survival Analysis, and IPCRidge. *Coxnet only was used for hypothesis-free models. The best models were selected based on performance and stability. Ensemble model voting was used to generate binary treatment recommendations using hypothesis-driven feature sets. A risk score was generated from the hypothesis-free models and applied to the binary treatment recommendation groups to generate standard- and high-molecular risk subgroups. Created in BioRender. Ribeiro, J. (2026) https://BioRender.com/dipv1av.Fig. 3Overlap of model treatment recommendations and actual treatments received.Middle number in Venn diagram illustrates patients who received treatment in alignment with the model recommendation. Schematic of AI model development. Hypothesis-driven and hypothesis-free feature selection were conducted. Hypothesis-driven features were subjected to nested 10-fold cross validation with a 90/10 train/test outer split and 90/10 train/validate inner split. Propensity score matched samples only were used for training and validation, but all samples were used for testing in the testing cohort. Model architectures evaluated were COXPH, Coxnet, ComponentWise Gradient Boosted Survival Analysis (GBSA), Gradient Boosted Survival Analysis, and IPCRidge. *Coxnet only was used for hypothesis-free models. The best models were selected based on performance and stability. Ensemble model voting was used to generate binary treatment recommendations using hypothesis-driven feature sets. A risk score was generated from the hypothesis-free models and applied to the binary treatment recommendation groups to generate standard- and high-molecular risk subgroups. Created in BioRender. Ribeiro, J. (2026) https://BioRender.com/dipv1av. Overlap of model treatment recommendations and actual treatments received. Middle number in Venn diagram illustrates patients who received treatment in alignment with the model recommendation. In the testing dataset, standard-risk patients predicted to have better response to first-line FOLFIRINOX compared to gem/nab-p had significantly prolonged median TTNT (10.3 months vs. 6.5 months; HR = 0.55 [95% CI: 0.40–0.76]; p < 0.001) and OS (16.0 months vs. 9.9 months; HR = 0.48 [95% CI: 0.34–0.67]; p < 0.0001) with FOLFIRINOX treatment (Fig. 4A). Patients identified as high-risk within the FOLFIRINOX recommendation group maintained modestly superior outcomes on FOLFIRINOX compared to gem/nab-p, but their overall prognosis on FOLFIRINOX was less favorable compared to the standard-risk group (median TTNT: 6.0 months vs. 5.2 months; HR = 0.68 [95% CI: 0.44–1.1]; p = 0.092 and median OS: 9.7 months vs. 7.9 months; HR = 0.55 [95% CI: 0.34–0.88]; p = 0.011) (Fig. 4B). Modeling results for the training cohort and entire cohort are shown in Supplementary Fig. S5.Fig. 4Testing cohort outcomes for patients recommended for FOLFIRINOX treatment by the model.Kaplan-Meier curves illustrate TTNT (A, B) and OS (C, D) for FOLFIRINOX-treated patients and gem/nab-p-treated patients among patients recommended by the model for first-line FOLFIRINOX in the testing cohort (n = 316). Patients were further stratified into standard- (A, C) and high-molecular risk (B, D) prognosis groups. Testing cohort outcomes for patients recommended for FOLFIRINOX treatment by the model. Kaplan-Meier curves illustrate TTNT (A, B) and OS (C, D) for FOLFIRINOX-treated patients and gem/nab-p-treated patients among patients recommended by the model for first-line FOLFIRINOX in the testing cohort (n = 316). Patients were further stratified into standard- (A, C) and high-molecular risk (B, D) prognosis groups. Patients within the “standard-risk recommend gem/nab-p” group had similar median TTNT and OS with FOLFIRINOX compared to gem/nab-p (TTNT: 8.0 months and 10.6 months, respectively; HR = 1.2 [95% CI: 0.75–1.8], p = 0.513 and OS: 12.2 months vs. 13.7 months, respectively; HR = 1.1 [95% CI: 0.68-1.8]; p = 0.725). The high-risk group also showed no significant difference in median TTNT or OS between FOLFIRINOX and gem/nab-p, but displayed a worse prognosis overall (TTNT: 5.6 months vs. 5.5 months, respectively; HR = 0.88 [95% CI: 0.66–1.2]; p = 0.367 and OS: 8.3 months vs. 7.6 months, respectively; HR = 0.87 [95% CI: 0.65–1.2]; p = 0.349) (Fig. 5A–D). Modeling results for the training cohort and entire cohort are shown in Supplementary Fig. S6.Fig. 5Testing cohort outcomes for patients recommended for gem/nab-p treatment by the model.Kaplan-Meier curves illustrate TTNT (A, B) and OS (C, D) for FOLFIRINOX-treated patients and gem/nab-p-treated patients among patients recommended by the model for first-line gem/nab-p in the testing cohort (n = 345). Patients were further stratified into standard- (A, C) and high-molecular risk (B, D) prognosis groups. (E, F) Kaplan-Meier analysis for time-on-treatment (TTNT, E) and overall survival (OS, F) was performed on patients expected to have the best responses to gem/nab-p, irrespective of prognostic risk score. These patients were identified using an upper tertile cutoff of the ensemble votes and comprised 17.1% of the testing cohort. Testing cohort outcomes for patients recommended for gem/nab-p treatment by the model. Kaplan-Meier curves illustrate TTNT (A, B) and OS (C, D) for FOLFIRINOX-treated patients and gem/nab-p-treated patients among patients recommended by the model for first-line gem/nab-p in the testing cohort (n = 345). Patients were further stratified into standard- (A, C) and high-molecular risk (B, D) prognosis groups. (E, F) Kaplan-Meier analysis for time-on-treatment (TTNT, E) and overall survival (OS, F) was performed on patients expected to have the best responses to gem/nab-p, irrespective of prognostic risk score. These patients were identified using an upper tertile cutoff of the ensemble votes and comprised 17.1% of the testing cohort. As shown in Supplementary Fig. S1, in the entire cohort, the median OS of patients treated with FOLFIRINOX was 11.8 month (95% CI: 10.9–12.7) and with gem/nab-p was 8.9 month (95% CI: 8.1–9.6). Relative to these outcomes, the prolonged OS observed for FOLFIRINOX-treated patients in the “recommend FOLFIRINOX group” and the equalization of outcomes in the “recommend gem/nab-p” indicates the robustness of the model in identifying patients predicted to do well on each therapy. However, because the gem/nab-p recommendation group showed no significant difference in outcomes between regimens, we concluded that these patients could reasonably receive either treatment and that survival would not be compromised using gem/nab-p for older patients with comorbidities in this cohort. However, we wanted to further explore the possibility that the model could identify patients with superior responses to first-line gem/nab-p. To do so, we analyzed patients in the testing cohort with ensemble model voting scores in the upper tertile (top third) for gem/nab-p, which corresponds to the model’s internal 30% quantile cutoff used to define fixed voting positions during ensemble training and was not selected by optimizing against outcomes in the testing cohort (additional post-hoc sensitivity analyses were conducted and are shown in Supplementary Table S1). In this group of patients (n = 115), median OS was significantly shorter for FOLFIRINOX-treated patients compared to gem/nab-p-treated patients (8.3 months vs. 13.3 months, HR = 1.7 [95% CI: 1.1–2.8], p = 0.016; Fig. 5E, F). To confirm that these findings for the gem/nab-p recommendation group were not merely prognostic, we also examined TTNT and OS across score tertiles, which illustrated a monotonic increase in TTNT and OS for gem/nab-p treated patients but not for FOLFIRINOX-treated patients (Supplementary Table S2). These exploratory findings further emphasize the robustness of the binary treatment recommendations and imply that the model can extract patients with superior outcomes on gem/nab-p compared to FOLFIRINOX, favoring gem/nab-p selection for this group. Multivariate analysis According to a multivariate Cox proportional-hazards model analysis using all data, clinical covariates (sex, age, frailty score, PurIST subtype, and primary tumor site) and technical covariates (NGS assay type, percent tumor nuclei, average coverage depth, and total mapped read pairs) were not independently associated with survival (Table 2). There was a modest relationship between primary tumor site and survival. Interestingly, despite improved prognosis previously reported for the PurIST classical subtype compared to basal19,20, there was not a significant independent effect of subtype on survival in our analysis (a modest, non-significant trend was noted for strongly classical tumors). There was a significant interaction between the model treatment recommendation and actual treatment regimen the patient received (HR = 0.35 [95% CI: 0.28–0.42], p = 2.1e-26), indicating that the association between the regimen received and survival differed according to the model recommended regimen.Table 2Cox multivariate analysis for overall survivalTermHazard ratiop valueModel Recommendation1.87 (1.64–2.13)1.8e-20Regimen Cohort Gemcitabine/Nab-Paclitaxel2.35 (2.01–2.74)8.9e-27Sex: Male1.00 (0.91–1.10)0.94Age1.00 (0.99–1.00)0.43Frailty score1.00 (0.99–1.00)0.88PurIST result: Lean Classical0.86 (0.58–1.30)0.48PurIST result: Likely Basal-Like0.92 (0.67–1.28)0.63PurIST result: Likely Classical0.92 (0.66–1.27)0.60PurIST result: Strong Basal-Like1.05 (0.77–1.43)0.77PurIST result: Strong Classical0.74 (0.54–1.01)0.054Primary Tumor Site: Body of pancreas0.85 (0.12–6.17)0.87Primary Tumor Site: Head of pancreas0.79 (0.11–5.71)0.81Primary Tumor Site: Neck of pancreas1.00 (0.13–7.67)1.0Primary Tumor Site: Other specified parts of pancreas0.73 (0.08–6.37)0.78Primary Tumor Site: Overlapping lesion of pancreas0.75 (0.10–5.60)0.78Primary Tumor Site: Pancreas, NOS0.86 (0.12–6.21)0.88Primary Tumor Site: Pancreatic duct0.64 (0.08–4.84)0.67Primary Tumor Site: Tail of pancreas0.91 (0.13–6.62)0.93Specimen Site Type: Primary0.79 (0.71–0.88)2.2e-05Percent Tumor Nuclei1.00 (1.00–1.00)0.58Total Mapped Read Pairs1.00 (1.00–1.00)0.63Average Coverage Depth1.00 (1.00–1.00)0.2Model Recommendation:Regimen Cohort Gemcitabine/Nab-Paclitaxel0.35 (0.28–0.42)2.1e-26 Table 2Cox multivariate analysis for overall survivalTermHazard ratiop valueModel Recommendation1.87 (1.64–2.13)1.8e-20Regimen Cohort Gemcitabine/Nab-Paclitaxel2.35 (2.01–2.74)8.9e-27Sex: Male1.00 (0.91–1.10)0.94Age1.00 (0.99–1.00)0.43Frailty score1.00 (0.99–1.00)0.88PurIST result: Lean Classical0.86 (0.58–1.30)0.48PurIST result: Likely Basal-Like0.92 (0.67–1.28)0.63PurIST result: Likely Classical0.92 (0.66–1.27)0.60PurIST result: Strong Basal-Like1.05 (0.77–1.43)0.77PurIST result: Strong Classical0.74 (0.54–1.01)0.054Primary Tumor Site: Body of pancreas0.85 (0.12–6.17)0.87Primary Tumor Site: Head of pancreas0.79 (0.11–5.71)0.81Primary Tumor Site: Neck of pancreas1.00 (0.13–7.67)1.0Primary Tumor Site: Other specified parts of pancreas0.73 (0.08–6.37)0.78Primary Tumor Site: Overlapping lesion of pancreas0.75 (0.10–5.60)0.78Primary Tumor Site: Pancreas, NOS0.86 (0.12–6.21)0.88Primary Tumor Site: Pancreatic duct0.64 (0.08–4.84)0.67Primary Tumor Site: Tail of pancreas0.91 (0.13–6.62)0.93Specimen Site Type: Primary0.79 (0.71–0.88)2.2e-05Percent Tumor Nuclei1.00 (1.00–1.00)0.58Total Mapped Read Pairs1.00 (1.00–1.00)0.63Average Coverage Depth1.00 (1.00–1.00)0.2Model Recommendation:Regimen Cohort Gemcitabine/Nab-Paclitaxel0.35 (0.28–0.42)2.1e-26 Cox multivariate analysis for overall survival Association of model recommendations with molecular features There was no significant relationship between mutations in ATM, BRAF, BRCA1, CDKN2A, PALB2, or SMAD4; NTRK3 fusions; microsatellite instability-high status; or tumor mutational burden-high status with model treatment recommendations. However, compared to the “recommend first-line gem/nab-p” group, significantly more tumors with expected superior response to FOLFIRINOX had BRCA2 mutations (1.3% vs 4.6%) and HRD (3.15% vs. 7.36%; Fisher’s exact test, p < 0.0001; Fig. 6), aligning with the known response of tumors with HRD to platinum-based therapies. This was also expected given that the HR pathway was included as a feature in the model.Fig. 6Association of treatment recommendation with molecular features.A Percentage of tumors with identified molecular features across each modeling group. The denominator represents the entire cohort for all alterations other than homologous recombination repair deficiency (HRD); a clinical HRD call was unable to be made for 77 of 2202 tumors. B Positive percentages for each alteration are shown. Biomarkers indicated in red had significantly different rates of mutation between the “recommend FOLFIRINOX” and “recommend gem/nab-p” groups, independent of risk group (Fisher’s exact test, p < 0.05). Association of treatment recommendation with molecular features. A Percentage of tumors with identified molecular features across each modeling group. The denominator represents the entire cohort for all alterations other than homologous recombination repair deficiency (HRD); a clinical HRD call was unable to be made for 77 of 2202 tumors. B Positive percentages for each alteration are shown. Biomarkers indicated in red had significantly different rates of mutation between the “recommend FOLFIRINOX” and “recommend gem/nab-p” groups, independent of risk group (Fisher’s exact test, p < 0.05). The rate of KRAS mutations was significantly lower in tumors with predicted FOLFIRINOX benefit over gem/nab-p (89% vs. 94%, Fisher’s exact test, p < 0.0001; Fig. 5). Overall, there was a tendency for patients with KRAS-wildtype (WT) tumors to be recommended for FOLFIRINOX and for patients with KRAS-mutated (mut) tumors to be recommended for gem/nab-p, while these preferences were somewhat discrepant from the actual treatments they received (Supplementary Fig. S7A, B). Patients with both KRAS-mut and KRAS-WT tumors derived significant TTNT and OS benefit from first-line FOLFIRINOX over gem/nab-p, as previously reported26, although the difference was more striking in the KRAS-WT group (Supplementary Fig. S8). When we applied our predictive model, the FOLFIRINOX benefit for patients in the “recommend first-line FOLFIRINOX” group was observed in both KRAS-mut and KRAS-WT subgroups but again was more pronounced in the KRAS-WT subgroup (Supplementary Figs. S7C, D and S9–10). Finally, we investigated PurIST scores across model recommendation groups. While more patients were treated with FOLFIRINOX than gem/nab-p across all PurIST categories, the model recommendations favored gem/nab-p for strong-basal tumors and modestly favored FOLFIRINOX for strong-classical tumors (Fig. 7A-B). This result bolsters previous findings that strong-classical tumors have superior FOLFIRINOX response19, but offers a more nuanced view of this assertion given that a substantial number of strong-classical tumors were also recommended by the model for gem/nab-p. Furthermore, while patients actually treated with gem/nab-p had higher frailty scores across all PurIST categories, the model treatment recommendations were decoupled from frailty scores (Fig. 7C-D).Fig. 7Association of treatment recommendation with PurIST subtypes.A Percentage of patients treated with FOLFIRINOX or gem/nab-p across PurIST subtypes. B Percentage of patients in each modeling group across PurIST subtypes. C Frailty scores across PurIST subtypes according to treatment regimen. Midline shows median. D Frailty scores in each modeling group across PurIST subtypes. Midline shows median. x- and y-axis labels carry across graphs. Association of treatment recommendation with PurIST subtypes. A Percentage of patients treated with FOLFIRINOX or gem/nab-p across PurIST subtypes. B Percentage of patients in each modeling group across PurIST subtypes. C Frailty scores across PurIST subtypes according to treatment regimen. Midline shows median. D Frailty scores in each modeling group across PurIST subtypes. Midline shows median. x- and y-axis labels carry across graphs. Discussion Personalized approaches towards treatment selection have the potential to extend survival for patients with PDAC27. However, there are few comprehensive approaches to define optimal first-line treatment regimens in the advanced setting. To address this need, we developed an AI model to provide clinically relevant first-line treatment recommendations for patients with advanced PDAC. Our model successfully identified patients with superior response to first-line FOLFIRINOX compared to gem/nab-p and also identified patients who were likely to have similar outcomes regardless of which treatment was given first—although gem/nab-p was favored for a subset of these patients. Clinically, this information could support selection of FOLFIRINOX over gem/nab-p for a patient with borderline performance status, or allow greater flexibility in first-line treatment choice for certain patients. Importantly, model results are generated immediately following the completion of sequencing, with no impact on average case turnaround time of nine days for molecular profiling. Limited direct head-to-head comparisons of FOLFIRINOX and gem/nab-p in the advanced setting makes treatment decisions between these two regimens more complex28. While some studies show improved clinical endpoints with FOLFIRINOX3,7,9,11—which we also observed in our cohort—several conflicting studies exist including recent prospective trials12–16. As we were unable to control all prognostic clinical variables in our dataset (e.g., performance status, number of metastatic sites, CA19-9 levels, albumin), the debate over the overall superiority of one regimen over the other remains unsettled. Furthermore, while it’s generally considered that gem/nab-p is more tolerable for older patients, each regimen in fact has been associated with a unique toxicity profile3,7,9,12,14,15. Harnessing AI can enable physicians to select treatments in an informed manner with greater confidence, while providing treatment flexibility for patients who are predicted to respond similarly to both regimens. Cost considerations may also be relevant in these decisions; economic evaluations on the use of different regimens in patients with advanced PDAC in the United States suggest that gem/nab-p is the more cost-effective treatment29,30, although costs associated with different treatment-related expenses can vary for each regimen31. In addition to limited randomized, controlled clinical trial data comparing FOLFIRINOX with gem/nab-p, biomarkers to predict patient response to each regimen are also lacking. The transcriptomic-based “Pancreas-View”21–24 and the more recently developed whole-slide image-based platform CHAI25 employed similar approaches to identifying patients predicted to have superior outcomes to fluoropyrimidine-based or gemcitabine-based regimens. There are a few key features of our dual prediction model that differ from one or both of these tools. Most notably, our model was developed and validated in the advanced setting in a large cohort of patients (1,541 for training and 661 for validation) and provides a binary treatment recommendation specifically between FOLFIRINOX and gem/nab-p regimens. In validation, CHAI identified a 2.7 month survival benefit for patients recommended for fluoropyrimidine-based regimens when treated accordingly, while our model identified a 6.1 month survival benefit for standard-risk patients recommended for FOLFIRINOX when treated accordingly. Pancreas-View was developed for the adjuvant setting, so outcome analyses are not directly comparable. Despite differences, the success of these models in predicting treatment outcomes further emphasizes the validity of advanced AI approaches in personalizing therapy in PDAC. The PurIST classifier, which predicts molecular subtypes (basal, classical), has been used to inform prognosis and treatment response in PDAC, with the classical subtype demonstrating improved prognosis and FOLFIRINOX outcomes20. A more recent evaluation of PurIST subtypes in PDAC confirmed a survival benefit of first-line FOLFIRINOX for patients with classical subtype tumors, while demonstrating no such benefit in patients with basal subtype tumors19. In contrast to these findings, our multivariate analysis revealed only a non-significant trend relationship between strongly classical tumors and improved survival (HR = 0.75 [95% CI: 0.55–1.02], p = 0.066) and a modest association between the strongly-classical subtype and FOLFIRINOX recommendation. Similar to prior studies, strongly-basal tumors in our cohort were more likely to be predicted by the model to have equivalent outcomes with gem/nab-p. The treatment recommendation × regimen cohort interaction was significant even after correcting for PurIST subtype and other clinical and technical variables (HR = 0.34 [95% CI: 0.27–0.42], p = 3.5e-21), which Hendifar et al. also reported for their CHAI platform25. Notably, the Pass-01 trial found mostly non-significant differences in outcomes between regimens for each PurIST subtype, although OS favored gem/nab-p for the classical subtype. The reason for this difference in results is not clear but could be related to differences in therapy (modified FOLFIRINOX for Pass-01 vs. FOLFIRINOX for our study), cohort size (n = 160 for Pass-01 vs. n = 2202 for our study), and study design (prospective for Pass-01 and real-world, retrospective for our study). Examination of underlying molecular data beyond PurIST subtypes did not reveal striking differences between treatment recommendation groups in our analysis, but HRD likely contributed to the FOLFIRINOX treatment recommendation in a subset of tumors. Together, these findings suggest that a comprehensive AI approach is better at modeling outcomes compared to individual tumor-intrinsic factors or patient characteristics. KRAS mutations are ubiquitously prevalent in PDAC and confer worse survival26,32. Prior studies show that patients with KRAS-mut and KRAS-WT tumors both have superior outcomes with FOLFIRINOX compared to gem/nab-p26, which we also observed in our cohort. However, our model predicted that a substantial subset of patients with KRAS mutations (57.4%, n = 1153/2008) would likely not derive any additional benefit from FOLFIRINOX-first compared to gem/nab-p. In the realm of clinical trials, this model has the potential to inform backbone chemotherapy regimens for highly anticipated KRAS inhibitors rather than arbitrarily pairing with gem/nab-p. Conversely, patients with KRAS-WT tumors were more frequently predicted to exhibit superior response to FOLFIRINOX-first and appeared to derive an even greater absolute TTNT and OS benefit from this regimen compared to gem/nab-p. These results imply the importance of proper first-line chemotherapy selection, especially for patients with KRAS-WT tumors. Recent clinical trials have explored alternative chemotherapy regimens and sequencing strategies for PDAC. The phase I/II SEQUENCE trial (NCT02504333) evaluated a sequential chemotherapy strategy in metastatic PDAC in which gem/nab-p was sequenced with modified FOLFOX, producing favorable outcomes compared to gem/nab-p alone (13.2 month vs. 9.7 month OS)33. More recently, the phase III CASSANDRA trial (NCT04793932) compared neoadjuvant PAXG (cisplatin, nab-paclitaxel, capecitabine, and gemcitabine) with modified FOLFIRINOX in patients with resectable or borderline-resectable PDAC. PAXG demonstrated improved event-free survival and higher disease-control metrics, suggesting potential advantages for PAXG over modified FOLFIRINOX in earlier-stage disease34. While these studies highlight the potential of novel regimens and treatment sequencing strategies, additional research is required to fully realize the benefit of these approaches in advanced stage disease, as the current standard first-line treatment for advanced PDAC remains FOLFIRINOX or gem/nab-p. In contrast to empiric “umbrella” chemotherapy approaches that may expose patients to additional agents without clear biological rationale, the molecular-based AI treatment selection strategy described here aims to provide a more personalized and biologically informed framework for therapy selection. Beyond treatment recommendations, our model also provides relevant prognostic information for patients within each treatment recommendation group, which could complement current prognostic frameworks involving ctDNA detection and PurIST subtyping20,35. Furthermore, this prognostic information could perhaps be applied retrospectively to clinical trial participant data to ensure tumors with “high-molecular risk” and “standard-molecular risk” biologies respond equally to novel agents. In the context of model treatment recommendations, the identification of high-risk patients for clinical trials could be important, as they likely have the most to gain from novel interventions. There are a few limitations to our model. First, due to the more recent approval of NALIRIFOX in 2023, we were unable to include this regimen due to fewer treated patients and shorter follow-up periods. We postulate that the treatment predictions for FOLFIRINOX may be extrapolated to NALIRIFOX given the similarities between these regimens and clinical trial meta-analysis showing that they produce similar outcomes3. Notably, the median OS in the current study for FOLFIRINOX (11.8 months) and gem/nab-p (8.9 months) cohorts were quite similar to those reported in the NAPOLI-3 trial (11.1 months with NALIRIFOX and 9.2 months with gem/nab-p4, distinct from a recent real world evidence study reporting a median OS for FOLFIRINOX of 9.1 months36. Ultimately, further research will be required to establish the validity of extending our AI model to NALIRIFOX selection. Secondly, our model also does not predict risk of toxicities to either regimen; as such, physicians must still consider the different toxicity profiles of each regimen and factor individual patient characteristics into treatment decisions. This model is not intended to address the separate issue of toxicity, but rather fit within a broader treatment selection scheme by providing personalized information on predicted treatment efficacy. Lastly, our analysis is somewhat limited by the use of real-world data for training and validation. We did not have access to more detailed clinical information such as performance status, disease burden, and treatment intensity (although we used frailty score as a proxy for performance status). To minimize the potential bias of real-world data, we used a stringent filtering process for inclusion of patients so as to reduce as many confounding treatment variables as possible, and used a propensity score matching approach to minimize the impact of frailty, age, and sex on the model. Additional validation of this model in independent cohorts will be important to fully establish its broad applicability across PDAC populations. In conclusion, we present a new AI model to aid treatment selection in patients with advanced PDAC. Use of this model will provide greater confidence in selecting patients predicted to derive survival benefit from first-line FOLFIRINOX or gem/nab-p, while allowing greater consideration of toxicity and cost for patients who are predicted to have similar outcomes regardless of treatment regimen. Methods Patients The cohort used in this study comprised 2202 molecularly-profiled samples from primary or metastatic tumors from patients with advanced or metastatic PDAC. Only patients with both successful molecular profiling results ( ≥ 3 million mapped reads, ≥300 average DNA coverage depth) and clinical outcome data were included. Tissue collection occurred between May 3, 2017 and March 7, 2025 and sequencing occurred between October 1, 2020 and April 18, 2025. All specimens were derived from patients within the United States. Inclusion criteria were patients who received at least three cycles of first-line FOLFIRINOX or gem/nab-p with the first cycle occurring within 30 days prior to or within 100 days after tissue collection. This criterion was implemented in order to ensure quality survival data and real treatment-specific outcomes were included (the model does not assume that a patient must have completed three cycles of therapy to generate a treatment recommendation). Patients who received FOLFIRINOX or gem/nab-p outside this window and patients who underwent radiotherapy or other procedures were excluded. While patients who received incomplete or modified regimens were excluded, patients who were switched from FOLFIRINOX to FOLFIRI due to neuropathy from oxaliplatin were considered as if they had continued on FOLFIRINOX. The final cohort included 1332 patients who were treated with FOLFIRINOX and 870 patients who were treated with gem/nab-p (Fig. 1). Real-world clinical outcomes Real-world patient outcome data was obtained from insurance claims through Caris’ clinico-genomic database. Overall survival (OS) was calculated from start of first-line treatment to last contact or death. Time-to-next-treatment (TTNT) was used as a surrogate for progression-free survival and was calculated from start of treatment with FOLFIRINOX or gem/nab-p to start of next treatment, or last contact or death when no treatment change occurred. In the absence of a treatment change or death, patients were censored in model training and analyses using TTNT. Similarly, patients without a date of death were censored in the OS analysis. There was no data imputation performed. Censoring rules were applied according to standard survival analysis, and there was no administrative censoring. Propensity score matching Propensity score matching was performed to match samples in the FOLFIRINOX and gem/nab-p treated groups based on clinical covariates (frailty, age, and sex) and specimen age (denoted by activation year; Supplementary Fig. S2). Activation year is an integer value from the year 2020, referring to when Caris activated the case. Frailty score was calculated as previously described by Gilbert et al., using a validated tool based on ICD-10 codes characterizing diagnoses associated with frailty37. Propensity score matching was conducted prior to splitting the cohort into training and testing cohorts in order to reduce bias in treatment selection, which depended upon age, sex, and frailty of the patient, and the year the profiling was performed. Other variables were checked for confounding (primary tumor site, average coverage depth, total unique mapped read pairs, specimen site type, percent tumor nuclei, and purist classification) before settling on these four confounders. Nine-hundred thirty-two FOLFIRINOX-treated patients and 609 gem/nab-p-treated patients were used for training and 400 FOLFIRINOX-treated patients and 261 gem/nab-p-treated patients were used for testing (Fig. 1). Only propensity score matched samples were used for feature selection and model training, but all samples were used for out-of-fold predictions and for model evaluation in the testing cohort. Clinical and demographic characteristics of the training, testing, and total cohort are shown in Table 1 and Supplementary Table S3. Ethical compliance This study was conducted in accordance with guidelines of the Declaration of Helsinki, Belmont report, and U.S. Common rule. In keeping with 45 CFR 46.101(b)4, this study was performed utilizing retrospective, deidentified clinical data, and is considered IRB exempt with waivers of patient informed consent. Waiver of patient consent and exempt status were determined by WCG IRB. Molecular profiling Molecular profiling was performed at Caris Life Sciences (Phoenix, AZ, USA), a College of American Pathologists (CAP)/Clinical Laboratory Improvement Amendments (CLIA)-certified laboratory. Whole exome and whole transcriptome sequencing Whole exome sequencing (WES) and whole transcriptome sequencing (WTS) were performed on a NovaSeq 6000 System (Illumina, San Diego, CA, USA) (RRID:SCR_016387) either separately or using a hybrid WES/WTS assay, as previously described38. Raw WTS data in FASTQ files was demultiplexed, trimmed, counted, and sequences aligned to human reference genome hg19 (for separate WTS) or hg38 (for hybrid WES/WTS) by Spliced Transcripts Alignment to a Reference (STAR) (RRID:SCR_004463)39. Clinically relevant fusions and/or splice variants were detected using STAR-Fusion. Transcripts per million (TPM) molecules were generated using the Salmon expression pipeline (RRID:SCR_017036)40. Homologous recombination repair deficiency was determined as previously described41. Variant detection Pathogenic and likely-pathogenic mutations were identified from NGS data. Variants detected were mapped to reference genome hg38 using the Burrows-Wheeler Aligner (BWA 0.7.17) (RRID:SCR_010910) embedded in the analysis tools licensed from Sentieon®. Bioinformatic tools including SAMtools (RRID:SCR_002105), Pindel (RRID:SCR_000560), and SnpEff (RRID:SCR_005191) were incorporated to perform variant calling functions, and annotations were standardized to the Human Genome Variation Society format. Filtering was performed to remove benign variants, low quality and low depth variants, or variants determined to be unreliable in several analyses like strand bias and repeat analysis. A threshold of 5% variant allele frequency and five alignments supporting a variant were required to call positive variants. Model development A schematic of model development is shown in Fig. 2. Claims data preparation Each specimen underwent a three-step process of (1) algorithmic regimen-block construction, (2) algorithmic TTNT calculations and (3) manual review of marked TTNT lines and sample quality to produce the final data used in model training and evaluation. Algorithmic regimen block construction began with initial block seeding, where each drug administration event generated a single-day block containing all drugs administered on that date. Backbone identification distinguished protocol-defining core drugs from supportive add-on agents and was followed by gap-based window merging that collapsed consecutive blocks sharing the same backbone. Additional logic allowed for individual mono-block absorption as both prelude and tail to backbone regimen blocks. Catalog-based regimen labeling then mapped the final drug set of each block into a canonical regimen name (e.g., FOLFIRINOX, gem/nab-p, etc.) when the combination matched an entry from a previously constructed treatment catalog. Statistics on administration dates, doses, and duration were stored for parsing. Algorithmic TTNT calculations operated on the regimen blocks constructed above. First, detection of first-line start day was based on the first qualifying treatment block that passed regimen, dose, duration, and collection day-window criteria with similar treatment-switching criteria assigned for second-line detection. Next, the algorithm performed an alignment of TTNT window boundaries to actual core drug administration dates within the blocks, tightening start/end to first and last observed doses of protocol-defining drugs. TTNT calculations were then the difference between second-line start day and first-line start day when a valid second line was detected or the difference between last contact or death day and first-line start day when no valid second-line existed, with censoring flag set based on vital status. Finally, manual review of all samples, for both study inclusion and TTNT calculations was done using an interactive treatment viewer. Each sample’s treatment trajectory and associated regimen blocks and TTNT lines were reviewed, visually, confirming or adjusting TTNT endpoints for any non-standard cases incorrectly calculated by the TTNT determination algorithm. Molecular data preparation Data from WES and WTS sequencing for the 2202 PDAC cases was collected across two profiling assays. To harmonize data across assays, assay specific percentile lookup tables for copy number and expression data were built using 16,778 PDAC cases that did not have matched claims records meeting the claims curation criteria. All copy number and expression values of the 2202 cases were percentile transformed according to the assay used to obtain WES and WTS results38. Data leakage was prevented by removing the cases used in the training and testing cohorts from the list of cases used to create the harmonization lookup tables. A combined data set consisting of the percentile transformed data, pathogenic or likely pathogenic mutations, PurIST categories, and probability was generated for the training and testing cohorts. There were no exclusions based on targetable mutations or fusions. Feature selection Feature sets were derived using a hypothesis-driven or hypothesis-free approach (Fig. 2). For the hypothesis-driven approach, feature sets were formed through pathway aggregates or drug-specific genes across multiple pathways from literature review. Pathways and drugs represented in the final hypothesis-driven feature sets are shown in Supplementary Table S4. For the hypothesis-free approach, for each training fold all possible transcriptomic feature sets were ranked by concordance index (C-index) for each treatment regimen, where C-index = (number of concordant pairs)/(number of comparable pairs) (Supplementary Data S1). Training/testing Nested 10-fold cross-validation was performed for model development from the hypothesis-driven feature sets. The outer fold was split into 90% training and 10% testing, and the inner fold split the outer fold training set into 90% training and 10% validation. Using all LLM-aggregated sets of putative gene features, model performance was evaluated across all available architectures (CoxPH, Coxnet, Component Wise Gradient Boosted Survival Analysis, Gradient Boosted Survival Analysis, and IPCRidge) and log-risk calibrations (Fig. 2). The validation data was used as a benchmark from which hazard ratios (HR) were calculated between patients who received FOLFIRINOX and patients who received gem/nab-p (and associated p values) at log-risk extrema, e.g., top 30% and bottom 30% by log-risk for each outer fold. Models underwent a five-step filtering process: to find models with favorable FOLFIRINOX outcomes and, separately, more favorable gem/nab-p outcomes, models with high performance were filtered to those separating out FOLFIRINOX from gem/nab-p. Furthermore, only those models that exhibited this behavior across at least 8 of the 10 outer folds with significant p values (≤0.05) were retained. From the first filtered pool, putative feature sets were further reduced by stipulating that across all possible architectures they must have met the minimum thresholding of 20 requirements in across all architectures and log-risk calibrations. Each feature set was then reduced to its single best-performing architecture. Finally, each remaining feature set and architecture was filtered to a maximum of four unique log-risk calibrations; one standard, non-time based transformation, and up to three survival probability-based transformations. The above resulted in a short-list set of architecture, feature set, and log-risks, with some representative of detecting favorable FOLFIRINOX performance and others favorable gem/nab-p performance. Ten FOLFIRINOX models and five gem/nab-p models were ultimately selected for ensemble model voting based on performance and stability (Supplementary Data S2). Models were then stacked in multiple configuration sizes and feature set groups, first as a loose grid search, then iteratively with upper limits on the number of models used for each voting arm. Final performance was evaluated with hazard ratio (HR), p value, and size of the resultant recommendation buckets on validation followed by comparison of the top performing validation models to the performance on the cross-validation test sets associated with each outer fold. Log-risk calibrations Since results from models were combined across the treatment regimens, calibrations to log-risk were applied to linearly map between the log-risk vector spaces from the independent models, as well as supply time dependent predictions with a consistent baseline risk. Not only did this approach provide accurate log-risks when comparing and combining model results, but also added a necessary stability check used in evaluating model performance (Supplementary Data S3). Ensemble model voting Using the ten selected FOLFIRINOX models and the five selected gemcitabine models, ensemble model voting was conducted for each patient (Fig. 2). Per-fold voting was conducted where each model casts a vote when a patient exceeds its quantile cutoff. Majority normalized vote counts determined which treatment recommendation group the patient was assigned. For tie-breakers, the higher mean of per-fold proportions won. If means were equal, gem/nab-p was the default recommendation. This method assigned all patients into two recommendation groups: FOLFIRINOX or gem/nab-p. Risk-scoring Using the top hypothesis-free features, models were evaluated for performance stepwise by five features. Highly correlated features (Spearman rank) were removed in favor of the feature with a higher C-index. The process resulted in the selection of two sets of Coxnet survival models with 30 and 50 features per fold for FOLFIRINOX and gem/nab-p treated cases, respectively (Supplementary Data S4). These models were used to generate an overall risk-score for a median split applied to the FOLFIRINOX and gem/nab-p recommendation buckets. Patients with a risk score in the top 50th percentile were expected to have the worst prognosis and were termed “high-molecular risk”, as the risk score was generated from the model that was trained on molecular data. Patients in the lower 50th percentile were termed “standard-molecular risk” rather than “low-molecular risk” due to the overall poor prognosis expected for patients with PDAC (Fig. 2). Statistical analysis All measurements were taken from unique samples. To analyze survival data, HRs and 95% confidence intervals (CI) were calculated using Cox proportional hazards model and p values by log-rank test. Significance was determined as p < 0.05. A multivariate analysis was conducted for OS using model recommendation, treatment received (regimen cohort), sex, age, frailty score, PurIST classifier subtype, primary tumor site, specimen site type, percent tumor nuclei, total mapped read pairs, and average coverage depth as covariates. The possibility of multicollinearity was evaluated by examining models with and without the interaction term (model recommendation × regimen cohort) before finalizing the Cox model with the interaction term. Fisher’s exact test was used to determine significant differences in molecular features between treatment recommendation groups. Supplementary information Supplementary Materials Supplementary Data Supplementary Materials Supplementary Data