eIF4G2-dependent translation restrains pancreatic cancer progression

16 Pancreatic ductal adenocarcinoma (PDA) is among the most lethal cancers, driven by cellular plasticity that 17 fuels therapeutic resistance and early dissemination. The contribution of translational control to this plasticity 18 remains poorly understood. Through an in vivo CRISPR/Cas9 screen, we identify the non-canonical initiation 19 factor eIF4G2 (DAP5/NAT1) as a translational checkpoint restraining PDA progression. Loss of eIF4G2 20 accelerated tumor growth, induced poorly differentiated, basal -like histology, and triggered widespread 21 metastasis. Ribosome profiling revealed that eIF4G2 loss does not alter bulk protein synthesis but instead 22 impairs translation of a selective regulon, including tumor suppressors such as PTEN and CREBBP . Functional 23 studies confirmed that PTEN loss was sufficient to drive dedifferentiation but insufficient to promote metastasis, 24 implicating the broader eIF4G2 -dependent program , including translational control of transcriptional 25 regulators like CREBBP , in limiting dissemination. Consistently, eIF4G2 -deficient tumors exhibited 26 transcriptomic enrichment of programs related to migration and wound healing. Computational inference from 27 human PDA datasets revealed reduced eIF4G2 activity in metastases, aligning with basal -like features and 28 predicting poorer survival. These results support a model in which eIF4G2 maintains epithelial identity and 29 restrains metastatic potential, highlighting selective translation as a determinant of PDA subtype and clinical 30 outcome. 31 32 Running title: eIF4G2 is a translational checkpoint in PDA. 33 34 35 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 2

36 Pancreatic ductal adenocarcinoma (PDA) remains one of the most lethal human malignancies, with a five-year 37 survival rate below 12% despite advances in surgery, systemic therapy, and supportive care (1). Poor outcomes 38 are driven by late diagnosis, intrinsic therapeutic resistance, and early metastatic spread (2,3). Genetically, PDA 39 is relatively homogeneous: activating mutations in KRAS occur in the vast majority of tumors, while TP53, 40 CDKN2A, and SMAD4 are inactivated in large subsets (4,5). These alterations provide a common oncogenic 41 backbone upon which extensive phenotypic heterogeneity emerges. Over the past decade, transcriptomic 42 studies have organized this heterogeneity into clinically meaningful subtypes. Tumors segregate along an 43 epithelial “Classical” axis, marked by well-differentiated glandular morphology, epithelial gene programs, and 44 relatively better outcomes , versus a “Basal -like” (or squamous) axis characterized by dedifferentiation, 45 mesenchymal features, therapeutic refractoriness, and poor prognosis (6). This framework has clarified 46 important differences in disease behavior and treatment response (7), yet the molecular mechanisms that permit 47 plasticity between these states remain incompletely understood. In particular, while transcriptional 48 reprogramming has been extensively studied, the role of translational checkpoints in governing PDA cell states 49 has remained unexplored in vivo. 50 51 Protein synthesis is among the most energy -consuming cellular processes, accounting for ~20% of energy 52 usage in normal cells and up to 40% in cancer (8). Accordingly, oncogenic signaling pathways, including 53 PI3K/AKT/mTOR and MEK/MNK, often converge on the translation machinery (9). Many of these pathways 54 impinge on the eIF4F complex, the central driver of cap -dependent translation (10). The eIF4F complex 55 consists of the cap-binding protein eIF4E, the scaffolding protein eIF4G1, and the RNA helicase eIF4A (11). 56 Elevated eIF4F activity promotes the translation of oncogenic transcripts and correlates with poor prognosis 57 and therapy resistance across various cancers (9,12,13). Indeed, we and others have shown that inhibition of 58 eIF4A, the helicase component of the eIF4F complex, strongly impairs PDA cell growth compared to normal 59 ductal cells (14-16). While canonical eIF4F -dependent translation has been intensively studied, far less is 60 known about non-canonical initiation factors. One such factor, eIF4G2 (also known as DAP5 or NAT1), is a 61 paralog of eIF4G1 but lacks the domain that eIF4G1 uses to interact with the cap -binding protein eIF4E (17). 62 Several models have been proposed for the function of eIF4G2 , including IRES -mediated (18), m(6)A -63 mediated (19), and eIF3d-mediated (20) translation initiation. However, these models are based largely on in 64 vitro or cultured cell studies, and the role of eIF4G2 in vivo has yet to be clarified. Here, we identify eIF4G2 65 as a translational checkpoint in PDA. Loss of eIF4G2 in PDA cells drives dedifferentiation, “Basal-like” 66 reprogramming, and widespread metastasis through impaired translation of tumor -suppressive regulators , 67 including Pten and Crebbp. Computational analysis further revealed that reduced eIF4G2 activity characterizes 68 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 3 metastatic human PDA and predicts poor patient survival, supporting eIF4G2 as a key barrier to the progression 69 and metastatic dissemination of PDA. 70 71

& DISCUSSION 72 Genome-wide CRISPR screen identifies eIF4G2 as a suppressor of PDA 73 To identify genes essential for pancreatic cancer growth in vivo , we performed a genome -wide CRISPR 74 knockout screen using a murine KPC (KrasG12D;Trp53R172H;Pdx1-Cre) PDA cell line (ccmT2) subcutaneously 75 injected into immunocompetent C57Bl/6J mice (Fig . 1A). The screen was performed by transducing ccmT2 76 with the Brie mouse whole-genome CRISPR knockout library. Transduced cells were selected using puromycin 77 for 3 days. Following puromycin selection, a baseline sample was collected (Day 0). Cells were then passaged 78 in vitro for 17 days to deplete gRNAs targeting common essential genes, at which point an engraftment-ready 79 cell sample was collected (Day 17). Endpoint tumors were harvested 10 days post -injection after reaching an 80 average volume of approximately 1.5 cm³. Finally, sgRNA abundance was quantified via deep sequencing from 81 the Day 0, Day 17, and endpoint tumor samples. We confirmed adequate library representation (Supplementary 82 Fig. 1A) and robust depletion of essential genes by Day 17 (Supplementary Fig. 1B), which correlated with 83 DepMap essentiality scores (Fig. 1B). We find that the initial in vitro passage successfully depleted common 84 essential genes from the library, resulting in their less pronounced depletion in the tumors (Fig . 1C, 85 Supplementary Fig. 1B) (21). Importantly, our in vivo screen successfully identified known tumor suppressors 86 such as Cdkn2b, Nf2, Pten, Rb1, and Smad4, as well as canonical oncogenes including Kras and Myc (Fig. 1D). 87 Our results also strongly correlate with an independent screen conducted on pancreatic cancer cells using a 88 metabolic library (Supplementary Fig. 1C) (22), further confirming the technical robustness of our screen. 89 Among the top enriched hits, we identified Eif4g2, a non -canonical translation initiation factor. Multiple 90 independent sgRNAs targeting Eif4g2 were consistently enriched across both the in vitro and in vivo arms (Fig. 91 1D and 1E; Supplementary Fig. 1E and 1F), indicating that Eif4g2 loss provides a selective advantage during 92 tumor growth. 93 94 eIF4G2 has been linked to selective translation of transcripts involved in stress adaptation and survival (23), 95 but its role in cancer remains unsettled. While some studies suggest oncogenic functions (24,25), others report 96 tumor-suppressive roles (26), and recurrent EIF4G2 mutations have been observed in tumor genomes (27). 97 Thus, to investigate the role of eIF4G2 in PDA progression, we targeted Eif4g2 expression by CRISPR-Cas9, 98 using two independent sgRNAs, in murine KPC pancreatic cancer cells (ccmT4, a different KPC line from the 99 ccmT2 screening line). To assess tumorigenic potential in vivo, we orthotopically implanted control (sg Rosa) 100 or sgEif4g2 KPC cells into the pancreas of syngeneic mice. At 5 weeks post -implantation, mice were 101 euthanized for endpoint analyses. Animals harboring Eif4g2-deficient tumors (Fig. 2A) exhibited a significant 102 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 4 increase in tumor burden, with an average 2 -fold increase in tumor volume relative to control (Fig s. 2B). 103 sgEif4g2 tumors also exhibited elevated mitotic activity, as indicated by increased phospho-Histone H3 (pH3) 104 staining (Fig. 2C). Blinded histopathological analysis of primary tumors revealed a profound shift in 105 differentiation status. While control tumors generated by sgRosa cells ranged from well - to moderately 106 differentiated histological grades (Figs. 2D and 2E), with approximately half displaying glandular structures 107 (Fig. 2F), sgEif4g2 tumors were poorly differentiated, exhibiting marked loss of glandular features (Figs. 2D-108 2F). Strikingly, this dedifferentiated state was accompanied by widespread metastasis (observed in 100% of 109 sgEif4g2 tumors; Fig. 2G, Supplementary Table 1 ). Together, these findings reveal eIF4G2 as a previously 110 unrecognized regulator of PDA, acting to limit tumor growth, maintain epithelial differentiation, and restrain 111 metastatic dissemination. These observations prompted us to investigate whether eIF4G2 enforces these effects 112 through a selective translational program. 113 114 eIF4G2 drives a selective tumor-suppressive translatome in PDA 115 In PDA, increased eIF4F activity, particularly through its helicase, eIF4A, promotes tumor growth by enabling 116 the translation of structured, oncogenic mRNAs (9,14-16). Because eIF4G2 lacks the N -terminal eIF4E -117 binding domain found in its paralog, eIF4G1, but retains the ability to interact with eIF4A, we initially 118 hypothesized that eIF4G2 loss might liberate eIF4A , thereby enhancing eIF4F-dependent, global translation, 119 potentially explaining the increased proliferation of sg Eif4g2 PDA cells (Supplementary Figs. 2A, 2B) . 120 However, we observed no increases in bulk translation, polysome loading, or mTORC1 activity 121 (Supplementary Figs. 2C–2H) upon eIF4G2 loss. These results indicate that eIF4G2 loss likely does not act 122 through canonical eIF4F-dependent programs, prompting us to investigate selective translational changes by 123 ribosome profiling. 124 125 To define the selective translational program regulated by eIF4G2, we performed ribosome profiling (Ribo -126 seq) in three independent murine PDA cell lines (ccmT1, ccmT2, and ccmT4) transduced with sg Rosa or 127 sgEif4g2. Libraries met the expected quality metrics (Supplementary Figs. 3A and 3B ), with no eIF4G2 -128 dependent changes in elongation dynamics (Supplementary Fig. 3C) and codon usage (Supplementary Fig. 129 3D). Integration of ribosome-protected fragments (RPF) and mRNA abundances identified 83 transcripts with 130 significantly altered translation efficiency (adjusted p < 0.05), of which 79 were decreased and 4 were increased 131 (Fig. 3A and Supplementary Table 2) in sgEif4g2 cells. Among the most strongly downregulated transcripts 132 was Pten, a well -established tumor suppressor. Notably, PTEN mutations are rare in human PDA (28), but 133 decreased PTEN protein expression and elevated PI3K-AKT signaling are frequently observed and associated 134 with aggressive disease (29-31), and Pten deletion cooperates with mutant Kras to accelerate PDA 135 development in experimental models of PDA (32). These findings suggest that translational regulation by 136 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 5 eIF4G2 may provide a mechanistic explanation for the loss of PTEN in PDA. Indeed, polysome fractionation 137 confirmed reduced association of Pten mRNA within translation-active heavy polysomes (Supplementary Fig. 138 3E) and decreased PTEN protein levels in both murine (Fig. 3B and Supplementary Fig. 3F) and patient-139 derived (Fig. 3C) PDA models lacking EIF4G2. To test whether PTEN loss functionally accounts for the 140 phenotype of Eif4g2 loss, we orthotopic ally transplanted sgPten KPC cells (Supplementary Fig. 3G). These 141 tumors were significantly larger (Supplementary Fig. 3H) and exhibited poorly differentiated histology (Figs. 142 3D and 3E), recapitulating the growth and dedifferentiation phenotype of sgEif4g2 tumors. However, unlike 143 sgEif4g2 tumors, in this same orthotopic model, sgPten tumors did not give rise to metastatic lesions . Thus, 144 while Pten translational suppression contributes to enhanced growth and dedifferentiation, additional eIF4G2-145 dependent targets are likely required to explain the metastatic phenotype. Collectively, these data indicate that 146 eIF4G2 supports a selective, tumor -suppressive translational program that includes PTEN and additional 147 regulators, acting to maintain differentiation and restrain PDA progression. 148 149 eIF4G2 restrains a stemness-associated translational program with clinical relevance in PDA 150 To define the broader scope of eIF4G2-dependent translation, we profiled the whole-cell proteome of sgRosa 151 and sgEif4g2 PDA cells. Of the 79 transcripts whose translation was impaired in sgEif4g2 PDA cells, 31 (~40%) 152 were detected by proteomic analysis and exhibited concordant decreases in protein abundance (Supplementary 153 Table 3). Notably, 12 of these 31 (~40%) belong to a stemness-associated signature recently linked to eIF4G2 154 activity in intestinal stem cells (33) (Supplementary Table 3 ). Although not canonical EMT drivers, their 155 combined loss would be expected to compromise epithelial integrity, deregulate polarity, and promote 156 transcriptional reprogramming toward stem -like states. Among these, CREBBP is notable because it is 157 frequently lost in human cancers (34-37). Indeed, we demonstrate that the heavy polysome association of the 158 Crebbp mRNA (Supplementary Fig. 3J) and its protein expression (Fig. 3F) are strongly dependent on eIF4G2 159 expression. Consistent with the transcriptional functions of CREBBP , loss of eIF4G2 induces broad 160 transcriptomic changes (Supplementary Fig. 3K, Supplementary Table 4) that are enriched for migration, 161 wound healing, and neuronal features (Fig. 3G). In line with this, we observed robust upregulation of KRT14, 162 a basal epithelial marker associated with poorly differentiated and stem-like PDA (7), in sgEif4g2 tumor lysates 163 (Fig. 3H). Importantly, KRT14 induction was absent in sgPten tumors (Supplementary Fig. 3L), supporting the 164 notion that the stemness shift upon Eif4g2 loss extends beyond PTEN suppression, reflecting a broader cell -165 state reprogramming program. 166 167 To contextualize these findings for human PDA, we projected the sgEif4g2 expression profile onto established 168 human PDA transcriptional subtypes (6,38-40) and found strong enrichment for the “Basal” subtype (Fig. 4A), 169 as well as EMT-related programs (Fig. 4B). In accordance with this observation, laser-capture microdissection 170 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 6 of human PDA sections followed by RNA sequencing (19) revealed that EIF4G2 mRNA is strongly depleted 171 in poorly differentiated lesions (Supplementary Fig. 4A), providing a direct human correlate to our 172 experimental observations. To further ascertain clinical relevance, we used VIPER (Virtual Inference of Protein 173 activity by Enriched Regulon analysis) (41) with a curated EIF4G2 regulon derived from our Riboseq and 174 proteomic datasets to estimate eIF4G2 activity in human PDA transcriptomes. Inferred eIF4G2 activity was 175 significantly reduced in metastases compared with primary tumors (Fig. 4C), and low activity correlated with 176 poorer overall survival (Fig. 4D). Together, these results support a model in which eIF4G2 acts as a 177 translational checkpoint linking selective mRNA control to PDA progression. By buffering against stemness-178 associated and EMT programs through regulators such as PTEN and CREBBP , eIF4G2 constrains metastatic 179 plasticity, and loss of its activity marks a clinically aggressive subset of PDA tumors. 180 181 Collectively, our findings reveal that not all translational outputs drive oncogenesis ; in PDA, eIF4G2 -182 dependent translation actively restrains tumor progression . By sustaining translation of regulators such as 183 PTEN and CREBBP , eIF4G2 maintains epithelial identity, and its loss drives dedifferentiation and metastatic 184 spread. Reduced eIF4G2 activity in human metastases, together with its correlation with poor survival, 185 highlights its potential as a prognostic biomarker. Its alignment with “Basal -like” features, including KRT14 186 induction, further supports integrating eIF4G2 activity scores into PDA subtype classifiers. Notably, because 187 Basal-like PDA cells are more sensitive to KRAS inhibition (42), eIF4G2 activity scores may augment existing 188 subtype classifiers and help stratify sensitivity to emerging KRAS-targeted therapies. 189 190 191 192 193 194 195 196 197 198 199 200 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 7

201 202 Orthotopic allograft model of pancreatic cancer 203 Orthotopic engraftment of mouse pancreatic cancer cells was conducted as described (43). In brief, mice were 204 anesthetized using isoflurane and subcutaneously administered 5 mg/kg Carprofen. 100,000 cells were 205 transplanted to the parenchyma of the pancreas. The abdominal wall was sutured with absorbable Vicryl sutures 206 (Ethicon Cat# J392H), and the skin was closed with wound clips (CellPoint Scientific Inc. Cat# 203 -1000). 207 C57Bl/6J mice (RRID:IMSR_JAX:000664) were purchased from the Jackson Laboratory for syngeneic 208 orthotopic transplant experiments. All animal experiments were conducted in accordance with procedures 209 approved by the IACUC at Columbia University (AC-AABK554). 210 211 Cell culture conditions for monolayer cultures 212 Monolayer KPC primary pancreatic cancer cells were generated from tumors derived from Trp53+/LSL-R172H, 213 Kras+/LSL-G12D, Pdx1-Cre (KPC) mice (44). HEK293T cell line was from ATCC (Cat# CRL -3216, 214 RRID:CVCL_0063). All cells were cultured in DMEM (Gibco Cat#11995073) supplemented with 10% fetal 215 bovine serum (FBS) (Corning Cat# 35-010-cv) and 1% PS unless stated otherwise. All cells were cultured at 216 37 °C with 5% CO2. 217 218 CRISPR/Cas9-mediated gene deletion lentiviral production and transduction 219 KPC cells were transduced with lentivirus expressing Cas9 and sgRNAs (lentiCRISPRV2, Addgene # 52961). 220 Single sgRNAs were cloned by annealing two DNA oligos and T4 DNA ligation into a BsmB I-digested 221 lentiCRISPRV2. LentiCRISPRV2 lentiviruses were produced in HEK293T cells co-expressing the packaging 222 vectors (pPAX2, Addgene # 12260 and VSV -G, Addgene # 12259), concentrated with LentiX concentrator 223 (Clontech Cat# 631232), and resuspended with DMEM supplemented with 10% FBS and 1% PS at 10 X 224 concentration. Ecotropic pBabe-puro (Addgene Plasmid #1764) retroviruses were produced in Phoenix -ECO 225 cells (ATCC Cat# CRL -3214), concentrated with RetroX Concentrator (Clontech Cat# 631456), and 226 resuspended in DMEM supplemented with 10% FBS and 1% PS at 10X concentration. 100,000 cells were 227 plated and infected with 5X concentrated viruses and spinoculated at 600 × g for 45 min at room temperature. 228 One day after infection, cells were treated with 3 µg/mL puromycin (Sigma-Aldrich Cat # P9620) for selection. 229 230 sgRNA sequences used: 231 sgEif4g2-2 sense, 5′-CACCGGTTCAGATAGTCAGTCACAA-3′ 232 sgEif4g2-2 antisense, 5′-AAACTTGTGACTGACTATCTGAACC-3′ 233 sgEif4g2-4 sense, 5′-CACCGATTAGACCATGAACGAGCCA-3′ 234 sgEif4g2-4 antisense, 5′-AAACTGGCTCGTTCATGGTCTAATC-3′ 235 sgPten-1 sense, 5′-CACCGACTATTCCAATGTTCAGTGG-3′ 236 sgPten-1 antisense, 5′-AAACCCACTGAACATTGGAATAGTC-3′ 237 sgPten-2 sense, 5′-CACCGGGTTTGATAAGTTCTAGCTG-3′ 238 sgPten-2 antisense, 5′-AAACCAGCTAGAACTTATCAAACCC-3′ 239 sgRosa sense 5′- CACCGAAGATGGGCGGGAGTCTTC-3′ 240 sgRosa antisense, 5′-AAACGAAGACTCCCGCCCATCTTC-3′ 241 242 CRISPR screen and analysis 243 CRISPR Library and Lentivirus Production : A genome-wide knockout screen was performed using the Brie 244 mouse whole-genome CRISPR knockout library, which contains 78,637 gRNAs targeting 19,674 genes. For 245 lentivirus production, 1.8 × 107 HEK293T cells were seeded in a 175 cm 2 tissue culture flask and the 246 transfection was performed using a DNA mixture of VSV-G (5 µg), psPAX2 (50 µg), and 40 µg of the transfer 247 vector. Flasks were transferred to a 37 °C incubator for 6 –8 hours; after this, the media was aspirated and 248 replaced with BSA-supplemented media. Virus was harvested 36 hours after this media change, pooled, and 249 filtered. 250 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 8 Genome-wide CRISPR-Cas9 Knockout Screen: 100 million ccmT2 cells, a monoclonal murine KPC PDA cell 251 line, were transduced with the Brie lentiviral library at a multiplicity of infection (MOI) < 0.6. Three days post-252 transduction, the cells underwent selection with puromycin at 2 µg/mL for 3 days. After selection, a baseline 253 cell sample was collected (Day 0). The remaining puromycin-selected cells were passaged in vitro for 17 days 254 to deplete cells with knockouts of common essential genes. A sample of these "engraftment -ready" cells was 255 collected (Day 17). For the in vivo screen, 1 x 10⁶ engraftment -ready cells were injected subcutaneously into 256 the bilateral hind limbs of 10-week-old male B6J mice. The experiment was conducted with five cages of mice, 257 each containing five male mice. Tumors were allowed to grow for 10 days, at which point they were harvested 258 when they reached an average volume of approximately 1.5 cm³. Genomic DNA was isolated from the Day 0, 259 Day 17, and endpoint tumor samples using the QIAamp DNA Mini Kit (Qiagen) according to the 260 manufacturer's instructions. The gDNA from 5 mice in the same cage were combined as one sample. The gRNA 261 cassettes were amplified from the genomic DNA via a two-step PCR protocol. For the first step PCR, 500 µg 262 of gDNA were used as input for one PCR reaction and in total 400 mg of gDNA were used for one cage of 263 mice. The resulting amplicons were purified and subjected to deep sequencing on a n Illumina NextSeq 500 264 platform to quantify gRNA abundance. 265 266 Data Processing and Analysis : Sequencing reads were processed and analyzed using the MAGeCK (v0.5.9) 267 computational pipeline. Quality control analyses confirmed sufficient gRNA representation, high read -268 mapping rates (average > 85%), and uniform read distribution across all samples. Gene essentiality scores 269 (beta scores) for the in vitro portion of the screen (Day 17 vs. Day 0) were calculated using MAGeCK -MLE 270 and correlated with pan-cancer essentiality scores from the DepMap database (2025Q2). Gene-level depletion 271 or enrichment in the in vivo screen (endpoint tumors vs. Day 17) was analyzed using MAGeCK -RRA. The 272

were compared with a previously published in vivo screen (22) to confirm technical robustness. All 273 correlation analyses were performed using the Pearson correlation coefficient. 274 275 Cell proliferation assays 276 CellTitre-Glo: Cell proliferation assay was performed by seeding 3,000 pancreatic cancer cells per well in 277 opaque 96-well plates (Corning Cat# 3917). Cells were seeded in Gibco Fluorobrite DMEM (ThermoFisher 278 Cat# A1896701) supplemented with 10% FBS, 1% P/S and 1% GlutaMAX (ThermoFisher Cat# 35050061). 279 Cell viability was measured using a luminescent ATP-based assay (CellTiter-Glo, Promega Cat# G7573) with 280 a plate reader (SpectraMax i3x, Molecular Devices). Data were analysed with GraphPad Prism. 281 EdU incorporation: EdU incorporation was assessed using the Click -iT™ EdU Alexa Fluor ™ 488 Flow 282 Cytometry Assay Kit (Thermo Fisher Scientific). Cells were treated with 10 µM EdU for 2 hours at 37°C, then 283 trypsinized and fixed in 200 µL of 1% PFA for 10 minutes at room temperature with gentle rocking. After two 284 washes with flow buffer (PBS + 1% BSA), cells were permeabilized in 200 µL of 1× BD Perm Buffer for 15 285 minutes at room temperature. Following two additional washes, cells were incubated in 100 µL of freshly 286 prepared Click-iT™ reaction cocktail (430 µL reaction buffer, 20 µL CuSO₄, 1.2 µL Alexa Fluor 488 azide, 287 and 50 µL buffer additive) for 30 minutes at room temperature protected from light. Cells were then washed 288 twice and analyzed by flow cytometry (Attune NxT, Thermo Fisher Scientific). 289 290 Western blot analysis 291 Standard techniques were employed for immunoblotting of organoids. Protein lysates were prepared using 292 0.1% SDS lysis buffer in 50 mM Tris pH 8, 0.5% Deoxycholate, 150 mM NaCl, 2 mM EDTA, 1% NP40, with 293 1 tablet of PhosSTOP (Roche Cat# 4906837001) and 1 tablet of cOmplete ™, Mini, EDTA -free Protease 294 Inhibitor Cocktail (Roche Cat# 11836170001) per 10 ml buffer, and separated on 4 -12% Bis-Tris NuPAGE 295 gels (Invitrogen Cat# NP0335BOX) or 12% Bis-Tris SurePAGE Gel (Genscript Cat#M00669; M00667) or 3-296 8% Tris-Acetate NuPAGE gels (Invitrogen Cat# EA0378BOX), transferred onto a PVDF membrane (Millipore 297 Cat# IPVH00010) and incubated with the indicated antibodies for immunoblotting. The following primary 298 antibodies were used at 1:1000 dilution: eIF4G2 (D88B6) [Cell Signaling #5169, AB_10622189], eIF4A 299 (C32B4)[Cell Signaling# 2013, AB_2097363], 4EBP1[Cell Signaling# 9452, AB_331692], pS6(D68F8)[Cell 300 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 9 Signaling # 5364, AB_10694233], S6 (54D2) [Cell Signaling # 2317, AB_2238583], eIF4E(C46H6) [Cell 301 Signaling # 2067, AB_2097675], PTEN (138G6) [Cell Signaling # 9559, AB_390810], CREBBP(D6C5)[ Cell 302 Signaling # 7389, AB_2616020]. 303 304 Proteomic Analysis 305 Global lysate proteome sample preparation: 306 For each sample, 50 µg of protein were reduced with 5 mM dithiothreitol for 30 minutes at 25°C and 1 ,000 307 rpm, then alkylated in the dark with 10 mM iodoacetamide for 45 minutes at 25 °C and 1,000 rpm. Samples 308 were then diluted with 50 mM Tris for a final urea concentration of < 2M. EDTA was added for a final 309 concentration of 10 mM; lastly, SDS was added for a final concentration of 1%. Magnetic SP3 beads were 310 made by combining equal volumes of carboxylate -modified hydrophilic (Cytiva: 45152105050250) and 311 hydrophobic beads (Cytiva: 65152105050250). 500 µg of SP3 beads were added to each sample (input, 312 flowthrough and IP). 100% ethanol was added at a 1:1 volumetric ratio with the sample to precipitate the 313 protein material onto the beads. The samples were then incubated for 15 minutes at room temperature. 314 Following incubation, the beads were washed thrice with 1 mL of 80% ethanol and reconstituted in 100 µL of 315 freshly prepared ammonium bicarbonate with 0.5 µg of trypsin. The samples were incubated overnight at 37 316 °C degrees and 700 rpm to digest the proteins off of the SP3 beads. Tryptic peptides were then dried in a 317 vacuum concentrator and resuspended in 3% acetonitrile/0.2% formic acid for a final peptide concentration of 318 0.25 µg/µL. 319 320 LC-MS/MS analysis on a Q-Exactive HF: 321 Approximately 1 μg of total peptides were analyzed on a Waters M -Class UPLC using a 15 cm × 75 µm 322 IonOpticks C18 1.7 µm column coupled to a benchtop Thermo Fisher Scientific Orbitrap Q Exactive HF mass 323 spectrometer. Peptides were separated at a flow rate of 400 nL/min with a 150 -minute gradient, including 324 sample loading and column equilibration times. Data were acquired in data-independent mode using Xcalibur 325 software. MS1 spectra were measured with a resolution of 120,000, an AGC target of 3e6 and a scan range 326 from 350 to 1600 m/z. 35 isolation windows of 36.0 m/z were measured with a resolution of 30,000, an AGC 327 target of 3e6, normalized collision energies of 22.5, 25, and 27.5, and a fixed first mass of 200 m/z. 328 329 Raw data were searched against the mus musculus proteome (UP000000589) with Spectronaut (version 19.4) 330 and the DirectDIA workflow. The Biognosys Global Standard (BGS) factory settings were used, meaning 331 cross-run normalization was implemented under the “Automatic” setting (RT-dependent local regression) and 332 no imputation strategy was used. To measure relative protein abundance, iBAQ values were used as reported 333 by Spectronaut. No further normalization was applied. 334 335 Global translatome analysis 336 For each starting sample, cells were lysed and 1 µg of total RNA was isolated for RNA-Seq while 9 µg of total 337 RNA was digested with RNase I to isolate ribosome protected RNA footprints (RPFs). Both sample types were 338 taken through rRNA depletion and RNA-seq samples were heat fragmented. Both sample types then underwent 339 adapter ligation. Ligated RNA was reverse transcribed, libraries were generated via PCR, and then sequenced 340 on a NextSeq 2000. After sequencing, samples were processed with Eclipsebio's proprietary analysis pipeline 341 (v1). Unique molecular identifiers (UMIs) were pruned from read sequences using UMI -tools (v1.1.1). Next 342 3' adapters were trimmed from reads using Cutadapt (v3.2). Reads were then mapped to a custom, curated 343 database of repetitive elements and rRNA sequences. All non-repeat mapped reads were mapped to the genome 344 hg38 using STAR (v2.7.7a), during the alignment multimapping reads were not retained. The exclusion of 345 multimapping reads reduces false positive differential and pausing calls. The removal of multimapping reads 346 may lead to underestimates of counts for genes with high rates of pseudogenization or duplication events. PCR 347 duplicates were removed using UMI -tools (v1.1.1). Gene coverage was calculated using Eclipsebio's 348 proprietary feature counting algorithm and ribosome occupancy was calculated by dividing coverage in the 349 ribosome protected footprint (RPF) library by the RNA -Seq library. Differentially expressed and occupied 350 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 10 genes were identified using DESeq2. For transcriptome analyses, genomic aligned reads were aligned to the 351 transcriptome with mudskipper (v0.1.0). Transcriptome aligned reads were processed with riboWaltz and 352 custom analysis programs developed by Eclipsebio to determine codon usage, stall sites, 5' UTR loading, and 353 3' read through. ORF detection was performed on genome alignments using ribotricer (v1.3.3). 354 355 O-propargyl-puromycin labeling 356 Click-iT assays were performed using 1 × 106 cells per assay according to the manufacturer’s protocol. In brief, 357 O-propargyl-puromycin (OP -Puro; 20 μM) (Life Technologies Cat# C10456,) was added to the cells and 358 incubated for 30 minutes. Cells were washed in ice -cold PBS and then fixed and permeabilized. The amount 359 of Alexa-Fluor-488 conjugated OP -Puro was quantified using flow cytometry using a BD LSR Fortessa 360 Cytometer (BD Biosciences). Data analysis was performed using FlowJo v9.7.6 (FlowJo, Ashland, 361 RRID:SCR_008520). 362 363 Polysome fractionation and qRT-PCR 364 100 μg/ml of cycloheximide was supplemented into the media. The cells were then harvested on ice in PBS 365 containing 100 μg/ml cycloheximide. Cells were pelleted and lysed in 10 mM Tris-HCl (pH 8), 140 mM NaCl, 366 1.5 mM MgCl 2, 0.25% NP -40, 0.1% Triton X -100, 50 µM DTT, 150 μg/mL cycloheximide, and 640 U/mL 367 RNaseIN (Sigma Cat# 3335399001) for 15 minutes. Lysates were cleared and then loaded onto a 10 -50% 368 sucrose gradient made using a Biocomp Gradient Master 108 and centrifuged for 2 h ours and 15 minutes at 369 35,000 rpm in a SW41 rotor using a Sorvall Discovery 90SE. The gradients were fractionated on a Teledyne 370 ISCO Foxy R1 apparatus while monitoring the OD254. The following SybrGreen primers were used: 371 Pten: Forward 5'-GAAAGGGACGGACTGGTGTA, Reverse 5'-TAGGGCCTCTTGTGCCTTTA 372 Crebbp: Forward 5'-CACCATCTGTGGCTACTCCTCA, Reverse 5'- GGTTTCAGCACTGGTCACAGAG 373 374 L-Azidohomoalanine (AHA)-based detection of nascent protein synthesis by immunoblot 375 Cells were incubated in methionine -free medium at 37 °C for 30 minutes to deplete intracellular methionine. 376 Nascent protein synthesis was then pulse -labeled by adding AHA or methionine (negative control) to the 377 methionine-free medium and incubating for at least 10 minutes at 37 °C. Cycloheximide was used in parallel 378 as a translational inhibition control. Following treatment, cells were washed with ice -cold PBS and lysed in 379 RIPA buffer lacking EDTA (containing Tris pH 8.0, NaCl, NP -40, sodium deoxycholate, and SDS). Protein 380 concentration was measured using a BCA assay. Equal amounts of total protein were prepared in Click -iT™ 381 reaction buffer and incubated with a freshly prepared cocktail containing CuSO₄, biotin-alkyne, and Click-iT™ 382 buffer additive. Samples were incubated at room temperature to allow copper -catalyzed azide –alkyne 383 cycloaddition (CuAAC). Proteins were separated by SDS-PAGE and transferred to nitrocellulose membranes. 384 Biotin-labeled proteins were detected by probing with streptavidin–HRP (Cell Signaling Cat #3999) at a 1:1000 385 dilution and visualized by chemiluminescence. 386 387 Histological annotation and quantification of tumors 388 Whole slide imaging of H&E (hematoxylin and eosin)-stained tumor sections was performed using the Aperio 389 AT2 slide scanner (Leica), and the images were analyzed using QuPath software. For each sample, regions of 390 well-, moderately-, and poorly differentiated morphology were annotated based on standard histopathological 391 criteria by an experienced pathologist. The percentage of each category was quantified relative to the total 392 annotated tumor area per slide, with values summing to 100% per sample. All annotations were performed 393 blinded to sample identity. To enable subtype -relevant comparisons, a secondary classification was applied 394 using a glandular architecture -based system previously shown to correlate with transcriptional subtypes of 395 pancreatic ductal adenocarcinoma (PDAC ) (45). In this schema, gland -forming regions are enriched in 396 Classical subtype tumors, whereas non-gland-forming regions are characteristic of Basal-like PDAC. All well-397 differentiated areas were gland -forming, and all poorly differentiated areas were non -gland-forming. 398 Moderately differentiated areas were further subclassified into gland -forming and non -gland-forming 399 categories based on the presence or absence of defined glandular architecture. Quantitative measurements of 400 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 11 glandular and non –glandular composition were generated accordingly, offering a subtype -informed 401 stratification of tumor differentiation status. 402 403 eIF4G2 activity analysis on metastasis and survival outcomes 404 Expression data and clinicopathological metadata, including sample type (“Primary” vs. “Metastasis”) and 405 epithelial subtype (basal, classical, precursor), were obtained from the UNC pancreatic ductal adenocarcinoma 406 cohort (40). Protein activity was inferred using VIPER (41) applied to expression signatures. EIF4G2 activity 407 values were extracted from the resulting activity matrix, and the activity difference between primary and 408 metastatic tumors was evaluated with an unpaired, two-sided Wilcoxon rank-sum test. Since this was a single, 409 pre-specified comparison, no multiple-testing correction was applied. Analyses were performed in R (version 410 ≥4.3), and boxplots were generated using ggplot2. For survival analysis, samples were divided into two groups 411 (EIF4G2 High: activity > 0; EIF4G2 Low: activity ≤ 0), a threshold that corresponds to the sign change of the 412 standardized VIPER score. Overall survival was defined as the time in months from diagnosis (or the cohort -413 defined index) to death. Kaplan –Meier survival curves were estimated nonparametrically, and differences 414 between groups were compared using a two -sided log-rank (Mantel–Cox) test with α = 0.05. Analyses were 415 performed in R (version ≥4.3) using the survival package to create Survival objects, fit Kaplan –Meier 416 estimators, and perform log -rank tests, and with survival for visualization with number -at-risk tables. 417 Greenwood standard errors were used by default, and unadjusted log-rank P values are reported. 418 419 eIF4G2 z-score analysis on basal/classical/Epithelial/EMT signatures 420 Bulk RNA-seq counts from control (sgROSA) and EIF4G2 knockout (sgEIF4G2) cells were analyzed with 421 DESeq2 (46). For signature scoring, variance -stabilized expression values were calculated and converted to 422 per-gene Z -scores by centering and scaling across all samples. Gene sets included (i) Moffitt epithelial 423 programs defining Classical and Basal -like PDAC (40), and (ii) curated epithelial and mesenchymal marker 424 panels for EMT (47). Signature scores were obtained for each sample as the mean Z -score of genes from the 425 respective set present in the expression matrix. An EMT index was calculated as (Mesenchymal score – 426 Epithelial score). Marks used for each analysis are listed below: 427 428 Classical markers Basal-like markers TFF1 KRT5 TFF2 KRT6A TFF3 KRT14 GATA6 KRT17 KRT19 TP63 KRT7 S100A2 KRT8 ZEB1 MUC1 ZEB2 MUC13 VIM CEACAM5 FN1 CEACAM6 CDH2 AGR2 SNAI1 EPCAM SNAI2 SLC1A2 TWIST1 LGALS4 MMP2 SPINK1 MMP9 Classical markers Basal-like markers REG1A ITGA3 REG3A ITGB1 ALB LAMB3 APOA1 LAMC2 APOA2 CXCL1 CFTR CXCL8 DPP4 FOSL1 FABP1 JUN HNF1A IL1A HNF4A IL1B PIGR MYC SLC2A2 SOX9 VIL1 429 Epithelial markers Mesenchymal markers CDH1 VIM (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 12 Epithelial markers Mesenchymal markers EPCAM FN1 KRT8 CDH2 KRT18 SNAI1 KRT19 SNAI2 KRT7 TWIST1 CLDN1 ZEB1 CLDN3 ZEB2 CLDN4 MMP2 CLDN7 MMP9 CLDN8 ITGA5 OCLN ITGB1 TJP1 ITGB3 MUC1 LAMB1 MUC13 LAMC2 AGR2 COL1A1 AGR3 COL1A2 Epithelial markers Mesenchymal markers EHF COL3A1 ELF3 COL5A1 GATA6 COL6A1 LGALS4 THY1 PIGR PDGFRA ERBB2 PDGFRB ERBB3 TAGLN KLF5 ACTA2 SLC9A3R1 DES CEACAM5 SPARC CEACAM6 SERPINE1 CXCL12 FAP LOX POSTN 430 431 Quantifications and Statistical Analysis 432 Biochemical experiments in vitro were repeated at least three times, and the repeat number was increased 433 according to effect size or sample variation. We estimated the sample size considering the variation and mean 434 of the samples. No statistical method was used to predetermine the sample size. No animals or samples were 435 excluded from any analysis. Animals were randomly assigned groups for in vivo studies; no formal 436 randomization method was applied when assigning animals for treatment. All western blotting experiments 437 with quantification were performed a minimum of three times with independent biological samples and 438 analyzed by ImageJ 1.52q. Investigators were blinded to group allocation during data analysis. Statistical 439 analyses were performed using GraphPad Prism 8. All tests and p values are provided in the corresponding 440 Figs. or Fig. legends. 441 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 13 AUTHOR’S DISCLOSURE 442 A. Chavez has a series of CRISPR-related patents managed by Harvard and Columbia University. 443 444 AUTHORS’ CONTRIBUTIONS 445 J. Powers : Conceptualization, data curation, methodology , writing—review and editing. W. L ai: 446 Conceptualization, formal analysis, methodology, writing —original draft, writing —review and editing. P. 447 Ahmadi: Data curation, writing—review and editing. H. Kobayashi: formal analysis, methodology, writing—448 original draft, writing —review and editing. A. Curiel -Garcia: formal analysis, methodology, writing —449 original draft, writing —review and editing. E. Valenzuela: data curation, formal analysis, methodology, 450 writing—original draft, writing—review and editing. M. Jovanovic: formal analysis, methodology, writing—451 original draft, writing —review and editing. A. Chavez : Conceptualization, formal analysis, methodology, 452 writing—original draft, writing—review and editing. I.I.C. Chio: Conceptualization, resources, data curation, 453 formal analysis, supervision, funding acquisition, investigation, project administration, writing—original draft, 454 writing–review and editing. 455 456

457 We thank other members of the Chio lab, as well as Drs. Richard Baer, Laura Pasqualucci, and Luke Berchowitz 458 for discussion. This work was performed with the support of the Herbert Irving Comprehensive Cancer Center 459 (Columbia University Irving Medical Center) Proteomics, Flow Cytometry, Genomics and High Throughput 460 Screening, Oncology Precision Therapeutics and Imaging Core, and Molecular Pathology (MPSR) Shared 461 Resources, as well as the Mass Spectrometry Core Facility (Chemistry Department at Columbia University). 462 Funding: This work was supported by National Institute of Health (NIH) grants (R01 -CA240654, 463 1R01CA267870, 1R01CA273023 to I.I.C.C.), Pershing Square Sohn Research Alliance (to I.I.C.C.), Irma 464 Hirchl Trust (to I.I.C.C). A.C. is supported by an award from the NIH Director’s fund (DP2NS131566). P30-465 CA13696 supports the Herbert Irving Comprehensive Cancer Center at Columbia University, and the Columbia 466 University flow cytometry core is supported by P30CA36727. 467 468 469 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 14 FIGURES 470 471 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 15 Figure 1. A genome-wide in vivo CRISPR screen nominates eIF4G2 as a suppressor of PDA 472 (A) Schematic of the in vivo CRISPR-Cas9 screening workflow, from library transduction of KPC PDAC cells 473 to tumor harvest in 5 cages of 10-week-old mice. 474 (B) Correlation of gene essentiality scores from the in vitro portion of the screen (beta scores from MAGeCK-475 MLE, Day 17 vs. Day 0) with pan-cancer essentiality scores from the DepMap database. 476 (C) Reduced correlation of in vivo gene essentiality scores (beta scores, endpoint tumors vs. Day 17) with 477 DepMap scores. 478 (D) Changes in gRNA enrichment between the in vitro outgrowth (Day 17) and in vivo tumors, known tumor 479 suppressors (e.g., Cdkn2b, Pten) and oncogenes (e.g., Kras, Myc) are highlighted. 480 (E) Normalized read counts for four independent gRNAs targeting Eif4g2, showing their strong enrichment in 481 endpoint tumors relative to the Day 17 engraftment-ready cell population. Data are represented as the median 482 of all primary tumors. 483 484 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 16 485 Figure 2. eIF4G2 restrains tumor growth, maintains differentiation, and limits metastasis in vivo 486 (A) Immunoblot analysis of eIF4G2 expression in bulk tumor lysates. 487 (B) Mass of pancreatic tumors at the 5 -week endpoint after orthotopic implantation of control (sg Rosa) and 488 eIF4G2-deficient (sgEif4g2) KPC cells. sgEif4g2-2 and sgEif4g2-4 are two sgRNAs targeting Eif4g2. 489 (C) Immunohistochemical staining and quantification of phospho-histone H3 in orthotopic PDA tumors. 490 (D-F) Representative H&E images (D) and pathological assessments of sgRosa and sgEif4G2 tumor sections 491 based on differentiation status (E) and glandular morphology (F). 492 (G) Number of metastatic lesions across different organ sites. 493 Error bars in this Fig. are means ± SDs. Student’s t-test was performed. ns (not significant) for P ≥ 0.05, * (one 494 asterisk) for P < 0.05, ** (two asterisks) for P < 0.01, *** (three asterisks) for P < 0.001, and **** (four 495 asterisks) for P < 0.0001. 496 497 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 17 498 499 Figure 3. Loss of eIF4G2 impairs a tumor-suppressive translational program. 500 (A) V olcano plot showing changes in translation efficiency in sgEif4g2 vs sgRosa KPC cells. 501 (B) Immunoblot analysis of PTEN expression in control (sgRosa) and eIF4G2-deficient (sgEif4g2) KPC cells. 502 (C) Immunoblot analysis of PTEN expression in two independent control (sg Rosa) and eIF4G2 -deficient 503 (sgEif4g2) patient-derived PDA cell lines. 504 (D) Representative H&E images from sgRosa and sgPten tumor sections. 505 (E) Pathological assessment of sg Rosa and sg Pten tumor sections based on glandular features and 506 differentiation status. 507 (F) Representative immunoblots of CREBBP in control (sgRosa) and eIF4G2-deficient (sgEif4g2) KPC cells. 508 (G) Gene set enrichment analysis of transcripts upregulated upon eIF4G2 loss. 509 (H) Representative immunoblot analysis of KRT14 expression in control (sg Rosa) and eIF4G2 -deficient 510 (sgEif4g2) bulk tumor lysates. 511 Error bars in this Fig. are means ± SDs. Student’s t-test was performed. ns (not significant) for P ≥ 0.05, * (one 512 asterisk) for P < 0.05, ** (two asterisks) for P < 0.01, *** (three asterisks) for P < 0.001, and **** (four 513 asterisks) for P < 0.0001. 514 515 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 18 516 517 518 Figure 4. Loss of eIF4G2 promotes basal-like and metastasis-associated programs in human PDA. 519 (A) Boxplots showing Z-score–based expression signature scores for Basal and Classical PDAC programs in 520 control (sgRosa) and EIF4G2 knockout (sgEif4g2) cells. 521 (B) Boxplots of Epithelial, Mesenchymal, and EMT index scores comparing control and eIF4G2 knockout 522 cells. 523 (C) EIF4G2 protein activity analysis in primary versus metastatic tumors from the UNC PDAC cohort (40). 524 Differences between groups were evaluated using an unpaired, two -sided Wilcoxon rank -sum test; p-value 525 indicated. 526 (D) Kaplan-Meier analysis of overall survival of patients stratified by EIF4G2 activity (High > 0, Low ≤ 0) in 527 the UNC PDAC cohort (40). Survival probability was estimated nonparametrically and compared with a two-528 sided log-rank test; the number at risk is shown at the bottom. 529 530 531 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 19

532 1. Siegel RL, Kratzer TB, Giaquinto AN, Sung H, Jemal A. Cancer statistics, 2025. CA Cancer J Clin 533 2025;75:10-45 534 2. Ghaneh P, Costello E, Neoptolemos JP. Biology and management of pancreatic cancer. Gut 535 2007;56:1134-52 536 3. Maitra A, Hruban RH. Pancreatic cancer. Annu Rev Pathol 2008;3:157-88 537 4. Dreyer SB, Chang DK, Bailey P, Biankin AV . Pancreatic Cancer Genomes: Implications for Clinical 538 Management and Therapeutic Development. Clin Cancer Res 2017;23:1638-46 539 5. Yachida S, White CM, Naito Y , Zhong Y , Brosnan JA, Macgregor-Das AM, et al. Clinical significance 540 of the genetic landscape of pancreatic cancer and implications for identification of potential long-term 541 survivors. Clin Cancer Res 2012;18:6339-47 542 6. Collisson EA, Bailey P, Chang DK, Biankin AV . Molecular subtypes of pancreatic cancer. Nat Rev 543 Gastroenterol Hepatol 2019;16:207-20 544 7. Bailey P, Zhou X, An J, Peccerella T, Hu K, Springfeld C, et al. Refining the Treatment of Pancreatic 545 Cancer From Big Data to Improved Individual Survival. Function (Oxf) 2023;4:zqad011 546 8. Buttgereit F, Brand MD. A hierarchy of ATP -consuming processes in mammalian cells. The 547 Biochemical journal 1995;312 ( Pt 1):163-7 548 9. Pelletier J, Graff J, Ruggero D, Sonenberg N. Targeting the eIF4F translation initiation complex: a 549 critical nexus for cancer development. Cancer Res 2015;75:250-63 550 10. Ruggero D, Sonenberg N. The Akt of translational control. Oncogene 2005;24:7426-34 551 11. Sonenberg N, Hinnebusch AG. Regulation of translation initiation in eukaryotes: mechanisms and 552 biological targets. Cell 2009;136:731-45 553 12. Ruggero D, Montanaro L, Ma L, Xu W, Londei P, Cardon-Cardo C, et al. The translation factor eIF-4E 554 promotes tumor formation and cooperates with c -Myc in lymphomagenesis. Nature medicine 555 2004;10:484-6 556 13. Avdulov S, Li S, Michalek V , Burrichter D, Peterson M, Perlman DM, et al. Activation of translation 557 complex eIF4F is essential for the genesis and maintenance of the malignant phenotype in human 558 mammary epithelial cells. Cancer Cell 2004;5:553-63 559 14. Wolfe AL, Singh K, Zhong Y , Drewe P, Rajasekhar VK, Sanghvi VR, et al. RNA G-quadruplexes cause 560 eIF4A-dependent oncogene translation in cancer. Nature 2014;513:65-+ 561 15. Chan K, Robert F, Oertlin C, Kapeller-Libermann D, Avizonis D, Gutierrez J, et al. eIF4A supports an 562 oncogenic translation program in pancreatic ductal adenocarcinoma. Nature communications 563 2019;10:5151 564 16. Singh K, Lin J, Lecomte N, Mohan P, Gokce A, Sanghvi VR , et al. Targeting eIF4A-Dependent 565 Translation of KRAS Signaling Molecules. Cancer Res 2021;81:2002-14 566 17. Imataka H, Olsen HS, Sonenberg N. A new translational regulator with homology to eukaryotic 567 translation initiation factor 4G. The EMBO journal 1997;16:817-25 568 18. Liberman N, Gandin V , Svitkin YV , David M, Virgili G, Jaramillo M , et al. DAP5 associates with 569 eIF2beta and eIF4AI to promote Internal Ribosome Entry Site driven translation. Nucleic acids research 570 2015;43:3764-75 571 19. Di Timoteo G, Dattilo D, Centron-Broco A, Colantoni A, Guarnacci M, Rossi F , et al. Modulation of 572 circRNA Metabolism by m(6)A Modification. Cell reports 2020;31:107641 573 20. Lee ASY , Kranzusch PJ, Doudna JA, Cate JHD. eIF3d is an mRNA cap-binding protein that is required 574 for specialized translation initiation. Nature 2016;536:96-+ 575 21. Eirew P, O'Flanagan C, Ting J, Salehi S, Brimhall J, Wang B, et al. Accurate determination of CRISPR-576 mediated gene fitness in transplantable tumours. Nature communications 2022;13:4534 577 22. Zhu XG, Chudnovskiy A, Baudrier L, Prizer B, Liu Y , Ostendorf BN , et al. Functional Genomics In 578 Vivo Reveal Metabolic Dependencies of Pancreatic Cancer Cells. Cell Metab 2021;33:211-21 e6 579 23. Mahe M, Rios -Fuller T, Katsara O, Schneider RJ. Non -canonical mRNA translation initiation in cell 580 stress and cancer. NAR Cancer 2024;6:zcae026 581 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 20 24. Fu L, Wang Z, Jiang FX, Wei GH, Sun LE, Guo CY , et al. High Expression of EIF4G2 Mediated by 582 the TUG1/Hsa-miR-26a-5p Axis Is Associated with Poor Prognosis and Immune Infiltration of Gastric 583 Cancer. J Oncol 2022;2022 584 25. Alard A, Katsara O, Rios -Fuller T, Parra C, Ozerdem U, Ernlund A , et al. Breast cancer cell 585 mesenchymal transition and metastasis directed by DAP5/eIF3d-mediated selective mRNA translation. 586 Cell reports 2023;42:112646 587 26. Meril S, Muhlbauer Avni M, Lior C, Bahlsen M, Olender T, Savidor A, et al. Loss of EIF4G2 mediates 588 aggressiveness in distinct human endometrial cancer subpopulations with poor survival outcome in 589 patients. Oncogene 2024;43:1098-112 590 27. Meril S, Bahlsen M, Eisenstein M, Savidor A, Levin Y , Bialik S, et al. Loss-of-function cancer-linked 591 mutations in the EIF4G2 non-canonical translation initiation factor. Life Sci Alliance 2024;7 592 28. Asano T, Yao YX, Zhu JJ, Li DH, Abbruzzese JL, Reddy SAG. The PI 3-kinase/Akt signaling pathway 593 is activated due to aberrant Pten expression and targets transcription factors NF -κB and c -Myc in 594 pancreatic cancer cells. Oncogene 2004;23:8571-80 595 29. Son M, Kim H, Han D, Kim Y , Huh I, Han Y , et al. A Clinically Applicable 24 -Protein Model for 596 Classifying Risk Subgroups in Pancreatic Ductal Adenocarcinomas using Multiple Reaction 597 Monitoring-Mass Spectrometry. Clin Cancer Res 2021;27:3370-82 598 30. Law HC, Clement EJ, Grandgenett PM, Hollingsworth MA, Woods NT. The Proteomic Landscape of 599 Pancreatic Ductal Adenocarcinoma Liver Metastases Identifies Molecular Subtypes and Associations 600 with Clinical Response-Response. Clin Cancer Res 2021;27:4127 601 31. Cao L, Huang C, Cui Zhou D, Hu Y , Lih TM, Savage SR , et al. Proteogenomic characterization of 602 pancreatic ductal adenocarcinoma. Cell 2021;184:5031-52 e26 603 32. Ying HQ, Elpek KG, Vinjamoori A, Zimmerman SM, Chu GC, Yan HY, et al. PTEN Is a Major Tumor 604 Suppressor in Pancreatic Ductal Adenocarcinoma and Regulates an NF-κB-Cytokine Network. Cancer 605 discovery 2011;1:158-69 606 33. Kunitomi H, Khaine AM, Jamee R, Lancero M, Raychaudhuri E, Perli S , et al. eIF4G2-Mediated 607 Translation Initiation of Histone Modifiers Is Essential for Intestinal Stem Cell Maintenance and 608 Differentiation. bioRxiv 2025:2025.05.19.652834 609 34. Horton SJ, Giotopoulos G, Yun H, V ohra S, Sheppard O, Bashford-Rogers R, et al. Early loss of Crebbp 610 confers malignant stem cell properties on lymphoid progenitors. Nat Cell Biol 2017;19:1093-104 611 35. Huntly BJP, Horton SJ, Giotopoulos G, Yun HY , V ohra S, Sheppard O, et al. Early Loss of CREBBP 612 Confers Malignant Stem Cell Properties on Lymphoid Progenitors. Blood 2016;128 613 36. Jia D, Augert A, Kim DW, Eastwood E, Wu N, Ibrahim AH, et al. Crebbp Loss Drives Small Cell Lung 614 Cancer and Increases Sensitivity to HDAC Inhibition. Cancer discovery 2018;8:1422-37 615 37. Zhang JY , Vlasevska S, Wells V A, Nataraj S, Holmes AB, Duval R , et al. The CREBBP 616 Acetyltransferase Is a Haploinsufficient Tumor Suppressor in B -cell Lymphoma. Cancer discovery 617 2017;7:322-37 618 38. Bailey P, Chang DK, Nones K, Johns AL, Patch AM, Gingras MC , et al. Genomic analyses identify 619 molecular subtypes of pancreatic cancer. Nature 2016;531:47-52 620 39. Maurer C, Holmstrom SR, He J, Laise P, Su T, Ahmed A, et al. Experimental microdissection enables 621 functional harmonisation of pancreatic cancer subtypes. Gut 2019;68:1034-43 622 40. Moffitt RA, Marayati R, Flate EL, V olmar KE, Loeza SG, Hoadley KA, et al. Virtual microdissection 623 identifies distinct tumor - and stroma-specific subtypes of pancreatic ductal adenocarcinoma. Nature 624 genetics 2015;47:1168-78 625 41. Alvarez MJ, Shen Y , Giorgi FM, Lachmann A, Ding BB, Ye BH, et al. Functional characterization of 626 somatic mutations in cancer using network -based inference of protein activity. Nature genetics 627 2016;48:838-47 628 42. Singhal A, Styers HC, Rub J, Li Z, Torborg SR, Kim JY, et al. A Classical Epithelial State Drives Acute 629 Resistance to KRAS Inhibition in Pancreatic Cancer. Cancer discovery 2024;14:2122-34 630 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint 21 43. He D, Feng H, Sundberg B, Yang J, Powers J, Christian AH , et al. Methionine oxidation activates 631 pyruvate kinase M2 to promote pancreatic cancer metastasis. Mol Cell 2022;82:3045-60 e11 632 44. Hingorani SR, Wang L, Multani AS, Combs C, Deramaudt TB, Hruban RH , et al. Trp53R172H and 633 KrasG12D cooperate to promote chromosomal instability and widely metastatic pancreatic ductal 634 adenocarcinoma in mice. Cancer Cell 2005;7:469-83 635 45. Kalimuthu SN, Wilson GW, Grant RC, Seto M, O'Kane G, Vajpeyi R , et al. Morphological 636 classification of pancreatic ductal adenocarcinoma that predicts molecular subtypes and correlates with 637 clinical outcome. Gut 2020;69:317-28 638 46. Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA -seq data 639 with DESeq2. Genome Biol 2014;15:550 640 47. Nieto MA, Huang RY , Jackson RA, Thiery JP. Emt: 2016. Cell 2016;166:21-45 641 642 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint SUPPLEMENTARY FIGURES Supplementary Figure 1. Quality assessment and validation of the genome-wide CRISPR screen. (A) Distribution of log-transformed sgRNA read counts. Other metrics for screen quality are also shown such as read mapping rates (average > 85%), sgRNA dropout, and uniform read distribution (average Gini index ≈ 0.1). (B) Distribution of sgRNA log2-fold change for the in vitro (Day 17 vs. Day 0) and in vivo (endpoint tumors vs. Day 17) portions of the screen. (C) Correlation matrix of sgRNA read counts across different screening conditions. (D) The correlation of the sgRNA fold change from our in vivo screen with the gene score generated from the PDA in vivo screen performed in the Birsoy lab (22). (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint (E) The fold change of sgRNAs and volcano plot of the in vitro screen results (Day 17 vs. Day 0), the enrichment of Eif4g2 is highlighted. (F) Normalized read counts for four independent sgRNAs targeting Eif4g2, between the Day 0 baseline and the Day 17 cell population. Supplementary Figure 2. Loss of eIF4G2 does not alter global protein synthesis (A, B) Proliferation of control (sgRosa) and eIF4G2-deficient (sgEif4g2) KPC cells measured by CellTiter-Glo (A) and EdU incorporation (B). (C) Viability of control and eIF4G2 -deficient KPC cells upon increasing concentrations of eIF4A inhibitor (CR-1-31-B) for 72 hours. (D) L -Azidohomoalanine (AHA) incorporation in sg Rosa and sg Eif4g2 KPC cells was analyzed by immunoblot, with cells treated with cycloheximide serving as negative controls. (E) OP-Puro incorporation was similarly analyzed by immunoblot, with cycloheximide-treated cells as controls. (F) Polysome profiles of control and eIF4G2-deficient KPC cells. Absorbance at 254 nm is plotted as a function of sedimentation. (G and H) Representative immunoblot analysis of the abundance of eIF4F components ( I) and downstream substrates of mTOR activation (J). Error bars in this Fig. are means ± SDs. Student’s t-test was performed. ns (not significant) for P ≥ 0.05, * (one asterisk) for P < 0.05, ** (two asterisks) for P < 0.01, *** (three asterisks) for P < 0.001, and **** (four asterisks) for P < 0.0001. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint Supplementary Figure 3. Ribosome profiling to characterize eIF4G2-dependent translatome (A) Pairs of ribosome-protected footprint (RPF) and RNA-Seq libraries were prepared from each sample. The plot shows the distribution of RPM (reads per million) normalized gene counts in each of the libraries. (B) This plot shows the location of scaled read ends that have been corrected for their estimated P -site (the second binding site for tRNA in the ribosome) from the ribosome -protected footprint (RPF) libraries. This correction is to show where ribosomes are directly binding, as the sequenced read ends have an overhang past the ribosome complex. RPF libraries are expected to show a trinucleotide periodicity. (C) RPF reads will accumulate where a ribosome pauses or stalls during translation. This plot shows putative stall locations for each transcript in the transcriptome, where stalls were detected through the use of a Kolmogorov-Smirnov (KS) test on normalize d cumulative P -site densities. This analysis is performed per transcript, so the listed distances refer to the distance in transcriptomic space, not genomic space. (D) This scatter plot shows the mean codon usage in each condition, where codon usage is defined as the scaled normalized frequency that P -sites were found to be enriched in each codon. Each dot is a different codon, colored by whether it is a start, stop, or an amino acid -coding codon. Off -diagonal codons can indicate differential codon usage between the two conditions. Codon usage was determined using riboWaltz. If a gene has multiple transcripts, only the longest transcript is included in this analysis. (E) qRT-PCR analysis of Pten mRNA distribution across polysome fractions. (F) immunoblot analysis of PTEN expression in control (sgRosa) and eIF4G2-deficient (sgEif4g2) bulk tumor lysates. (G) Immunoblot analysis of PTEN expression in sgRosa and sgPten KPC cells. (H) Mass of pancreatic tumors from in sgRosa and sgPten KPC cells 5 weeks after orthotopic implantation. (I) Immunoblot analysis of eIF4G2 expression in control (sg Rosa) and Pten-deficient (sgPten) bulk tumor lysates. (J) qRT-PCR analysis of Crebbp mRNA distribution across polysome fractions. (K) V olcano plot of eIF4G2-dependent transcriptional changes. (L) KRT14 and PTEN expression in sg Rosa and sgPten bulk tumor lysates. PTEN and HSP90 immunoblots are also used in panel (I). Error bars in this Fig. are means ± SDs. Student’s t-test was performed. ns (not significant) for P ≥ 0.05, * (one asterisk) for P < 0.05, ** (two asterisks) for P < 0.01, *** (three asterisks) for P < 0.001, and **** (four asterisks) for P < 0.0001. Supplementary Figure 4. eIF4G2 expression in human PDA. (A) EIF4G2 mRNA abundance based on a recently published morpho-biotype framework for PDA, relative to well-differentiated lesions. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint Supplementary Table 1. Summary of metastatic lesions identified in each mouse transplanted with sgRosa or sgEif4g2 PDA cells by organ site. Supplementary Table 2. List of mRNAs that exhibit significantly different ribosomal occupancy, i.e., translation efficiency, in sgEif4g2 PDA cells compared to sgRosa control PDA cells. Supplementary Table 3. Translationally downregulated transcripts with matching protein decreases in sgEif4g2 vs sgRosa PDA cells. Supplementary Table 4. List of differentially expressed transcripts in sgEif4g2 PDA cells compared to sgRosa control PDA cells. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.14.711799doi: bioRxiv preprint