High-Throughput Translational Profiling with riboPLATE-seq

Protein synthesis is dysregulated in many diseases, but we lack a systems-level picture of how signaling molecules and RNA binding proteins interact with the translational machinery, largely due to technological limitations. Here we present riboPLATE-seq, a scalable method for generating paired libraries of ribosome-associated and total mRNA. As an extension of the PLATE-seq protocol, riboPLATE-seq utilizes barcoded primers for pooled library preparation, but additionally leverages rRNA immunoprecipitation on whole polysomes to measure ribosome association (RA). We demonstrate the performance of riboPLATE-seq and its utility in detecting translational alterations induced by inhibition of protein kinases.

7 treatments in total consisted of two competitive mTOR inhibitors, PP242 and AZD-8055; an inhibitor of 142 PI3K upstream of mTOR, BKM120; a specific inhibitor of MNK1/2 activity, MNK-i1; and 4EGi-1, a 4E-BP 143 mimic that inhibits the association of eIF4E and eIF4G. We determined concentrations of these drugs 144 from an examination of the literature, ensuring values near the half-maximum inhibitory concentrations 145 (IC50) for the main substrates of the drugs in question: 625nM PP242 14 , 50nM AZD-8055 15 , 1μM 146 BKM120 16 , 100nM MNK-i1 17 , and 50μM 4EGi-1 18 . In order to analyze possible interactions between 147 kinases, we also treated samples with pairwise combinations of PP242, BKM120, and MNK-i1. 148 Previous studies of the effect of MAP kinase interacting kinases (MKNKs, a.k.a. MNKs 1/2) on 149 translational regulation utilized small-molecule inhibitors of these proteins, notably the compound 150 CGP57380 19,20 . However, this compound has been shown to be a nonspecific inhibitor of several 151 unrelated kinases, with effects on eIF4F formation independent of its effects on MNKs. CGP57380 has 152 low-micromolar IC50 values for MNK isoforms (0.87 μM/ 1.6μM for MNK1/MNK2, respectively) and 153 significantly inhibits other kinases at these concentrations, including MKK1, CK1, and BRSK2 21 . 154 Additionally, CGP57380 concentrations below that which affects eIF4E phosphorylation may still 155 decrease proliferation and survival, and an increase in eIF4E:4EBP binding occurs at concentrations 156 below those impacting MNK1, indicating broad off-target effects impinging translational regulation 22 . 157 Determination of the translational targets of MNK1 via ribosome or polysome profiling with this drug in 158 prior work is complicated by this lack of specificity, especially with regards to off-target effects directly 159 impacting translational machinery. In contrast, MNK-i1 has been recently identified as a highly specific 160 MNK inhibitor, with IC50 values of 0.023 μM and 0.016 μM for MNK1 and MNK2 respectively, and blocks 161 eIF4E phosphorylation without impacting other pathways converging on eIF4E 22 . We therefore sought to 162 clarify the effect of MNK1 on translation with this novel inhibitor. 163 8 For comparison, we performed ribosome profiling and RNA-seq on TS-543 neurospheres treated with 164 PP242 or MNK-i1 in identical regimens to the riboPLATE-seq study, in order to assess the similarity of 165 translational perturbations detected across experiment types. As riboPLATE-seq measures the fraction 166 of an expressed transcript associated with ribosomes (hereafter referred to as "ribosome association" or 167 RA) and ribosome profiling/RNA-seq measure the average number of ribosomes bound per transcript 168 (conventionally defined as "translation efficiency" or TE), we expected these two methods to give 169 quantitatively distinct results while identifying similar sets of targets for these translational regulators. 170

Performance of riboPLATE-seq 171
First, we assessed the quality of the pooled ribosome-associated riboPLATE-seq and normal PLATE-seq 172 libraries in terms of library complexity and saturation. Figures 2A and 2B show saturation curves for 173 riboPLATE-seq and PLATE-seq, respectively, demonstrating the dependence of these libraries' 174 sensitivities on read depth. The two curves are comparable, with ~10-11K unique genes detected at 175 saturating depth, though riboPLATE-seq requires about twice the number of aligned reads as PLATE-seq 176 to achieve this saturation. As expected, PLATE-seq libraries contain more unique genes on average than 177 riboPLATE-seq libraries despite a shallower sequencing depth. Asymmetric division of initial lysate 178 volumes favoring ribosome IP over unmodified PLATE-seq (90%/10%) helped to combat this inherent 179 inequality, generating sufficiently complex libraries in both cases. 180 Figure 2C highlights the differences between riboPLATE-seq and PLATE-seq in terms of library 181 complexity and sequencing depth, and compares these with libraries generated by ligation-free 182 ribosome profiling and conventional RNA-seq. On average, riboPLATE-seq detects approximately 9,800 183 unique genes in 1.4 million uniquely mapped reads, while PLATE-seq detects an average of 10,400 genes 184 in 0.6 million reads per sample in this study. These measurements are comparable to the initial report 185 characterizing PLATE-seq, in which Bush et al. detected an average of approximately 10,200 genes from 9 0.67 million uniquely mapped reads per sample 10 . In contrast, ligation-free ribosome profiling and total 187 RNA sequencing libraries downsampled to the respective median read depths of riboPLATE-seq and 188 PLATE-seq libraries still detect ~14,000 genes each, reflecting the inherent complexity of these libraries 189 even at reduced sequencing depths (Additional File #1: Supplementary Figure S1). At full depth, 190 ribosome profiling libraries detect an average of 14,000 genes per 2 million reads, while RNA-seq 191 detects 15,000 genes in 1.6 million reads per sample. In summary, both riboPLATE-seq and PLATE-seq 192 generate libraries of a lower overall complexity than ribosome profiling and RNA sequencing, and 193 require substantially fewer reads to achieve saturation. 194 To determine the specificity of pan-ribosomal IP for ribosome-bound RNA, we measured the depletion 195 of RNA species for which we expected little or no ribosome association (RA) in the riboPLATE-seq vs 196 PLATE-seq libraries. We measured depletion extrinsically, with respect to a set of polyadenylated spike-197 in RNAs added after lysis, and intrinsically with respect to the set of highly-expressed, polyadenylated 198 non-coding RNA transcripts (ncRNA) contained in the UCSC "known genes" RefSeq annotation. Figures 199 2D-E summarize these two analyses. 200 To assess the depletion of spike-in RNA in riboPLATE-seq, we added polyadenylated RNA standards 201 (ERCC spike-ins) to half of the wells in a riboPLATE-seq experiment, after lysis but prior to ribosome IP. 202 Figure 2D shows the distribution of the log2-ratio of spike-in abundance in riboPLATE-seq to PLATE-seq, 203 demonstrating depletion of spike-ins associated with ribosomal IP across most wells. The wells exhibit a 204 median 4.5-fold depletion ratio, and 31/48 wells exhibit 4-fold depletion or greater. However, two wells 205 exhibit a modest (<2-fold) enrichment of spike-ins after ribosome IP, and the wide distribution of 206 depletion log2-ratios ranges from 11 to 0.7 (equivalent to ~1.6-fold relative enrichment). Aside from 207 potential non-specific pulldown of spike-ins, random re-initiation of free ribosomes on polyadenylated 208 transcripts in lysate could result in their capture by ribosome immunoprecipitation, resulting in their 209 enrichment in riboPLATE-seq. This might be minimized in future riboPLATE-seq studies by inclusion of 210 GDPNP in the lysis buffer. As GTP hydrolysis is required in both start site selection and subunit joining 211 steps of 80S initiation complex formation, inclusion of a non-hydrolyzable GTP analogue such as GDPNP 212 would prevent re-initiation of free ribosomes in lysate 23 . In summary, inclusion of ERCC RNA spike-ins 213 provides a valuable, internal quality control measurement to check IP fidelity in riboPLATE-seq. 214 As ribosome profiling has revealed low but significant levels of ribosome occupancy among ncRNA 24 , we 215 sought to contrast ribosome association between ncRNA and mRNA transcripts with riboPLATE-seq. We 216 expected noncoding transcripts to be generally depleted in our riboPLATE-seq libraries compared with 217 PLATE-seq libraries from the same samples. Indeed, we observed lower RA for the set of highly-218 expressed ncRNA transcripts than mRNA within the same sample ( Figure 2E). Examining the relationship 219 between RA and transcript abundance across ncRNA and mRNA gene sets also uncovered lower RAs for 220 ncRNA than mRNA at all expression levels ( Figure 2F), with similar patterns observed between 221 translation efficiency (TE) and expression level in our ribosome profiling and RNA-seq data ( Figure 2G). 222 Combined with the observed spike-in depletion, our results are consistent with the depletion of RNA 223 that is not bound to ribosomes by the pan-ribosomal immunoprecipitation implemented in riboPLATE-224 seq. 225

Pharmacological Screening of Mitogenic Signaling with riboPLATE-seq 226
After establishing the performance of the riboPLATE-seq, we sought to characterize its ability to detect 227 differential expression and RA. In principal component analyses (PCA) of the PLATE-seq and riboPLATE-228 seq profiles, samples segregate according to the drug with which they were treated and related drugs 229 co-cluster ( Figure 3A, B). Principal component 1 (PC1) separates DMSO-treated controls from drug-230 treated samples, and PC2 separates samples treated with one kinase inhibitor from those treated with 231 4EGi1 or a combination of kinase inhibitors. Combination-treated samples co-cluster more readily with 232 each other than with any of their singularly-treated counterparts due to separation on PC2. 233 We further calculated RA for all genes in all samples using riboPLATE-and PLATE-seq counts and 234 differences in RA across the genome relative to DMSO-treated controls. PCA plots of RA and log-fold 235 change in RA are shown in Figures 3C and 3D, respectively. In both, PC2 separates combination-and 236 4EGi1-treated samples from singularly-treated samples (PP242, BKM120, MNK-i1, or AZD8055 alone), as 237 it does in the previous plots in Figure 3A  This division between strong and weak inhibitors also appears in the RA-dependent plots in Figure 3C  line, clusters which correspond to combination and 4EGi-1 treatments are similarly arranged along PC1 250 (4EGi-1, MNK-i1/BKM120, PP242/MNK-i1, PP242/BKM120; Spearman rank correlation p=0.0) ( Table 2). 251

riboPLATE-seq Differential Translation Efficiency Analysis 252
In order to more rigorously analyze differential ribosome association as a function of drug treatment, we 253 utilized DESeq2 to compare the replicates for each drug treatment to vehicle controls. We first identified 12 the total set of significantly differentially ribosome-associated genes across all singular drug treatments 255 and generated a hierarchically-clustered heatmap of differential RA across conditions ( Figure 4A). 256 Signatures of differential RA due to treatment with PP242 (mTOR inhibitor), AZD8055 (mTOR inhibitor), 257 and BKM120 (PI3K inhibitor), which target the PI3K/mTOR pathway, form a cluster in the column 258 dendrogram. As expected, MNK-i1 and 4EGi-1 targets cluster separately. MNK-i1 targets MNK1 in a 259 separate pathway converging on the ribosome at eIF4E, and 4EGi-1 is a broad inhibitor of eIF4E and cap-260 dependent translation in general. 261 The closely related differential-RA signatures for PP242, BKM120, and AZD8055 also include strong 262 downregulation of the 5'TOP motif-containing genes, canonical translational targets of mTOR signaling, 263 as indicated by the black tick marks in Figure 4A. In the rightmost two columns of Figure 4A, signatures 264 of differential translation efficiency obtained via ribosome profiling and RNA-seq recapitulate the major 265 patterns seen in differential RA. Both up-and down-regulated targets of PP242 are in good agreement 266 between the two methods. The 5'TOP motif-containing genes exhibit low TE by ribosome profiling after 267 PP242 treatment, whereas MNKi-1 treatment is far less effective on these genes based on ribosome 268 profiling and leads to fewer differentially translated genes in general, consistent with riboPLATE-seq. 269 We used gene set enrichment analysis (GSEA) to identify gene ontologies associated with differential 270 ribosome association under each drug treatment ( suggestive of mTOR-independent inhibition of translation. Surprisingly, all kinase-dependent treatments 277 13 exhibit stronger, more consistent downregulation(s) of these genes than 4EGi-1; as this drug targets 278 eIF4E directly, we expected TOP genes to be included in the set of its strongly-inhibited targets. 279

Attenuation of Perturbations to Ribosome Association in Drug Combinations 280
We expected that drug combinations would elicit greater changes in RA than the individual drug 281 treatments alone. Specifically, we expected at least additivity if not outright synergy from simultaneous 282 inhibition of the PI3K/Akt/mTOR and MAPK/ERK pathways, and a similar but perhaps less pronounced 283 additivity of effects from inhibiting kinases in the same pathway (i.e. mTOR and PI3K). Surprisingly, we 284 instead found a pattern of attenuation of the strongest effects of individual drugs when combined. 285  Table 1). 297

Motif-Based Target Classification in a Translation Control Network 298
Finally, we displayed the results of our study in network form and mapped occurrences of a known 299 translational cis-regulatory element, the 5' TOP motif, across this network. Following the observation of 300 14 concordant regulation of canonical TOP genes in drug treatments impacting the mTOR signaling 301 pathway, we first sought to expand the potential set of TOP genes within the strongly-perturbed genes 302 in this study. We first obtained sets of canonical TOP genes and candidates containing previously 303 uncatalogued 5'TOP tracts, a subset of which have known TOP-containing homologues in the mouse 304 genome, from the comprehensive analysis of human transcription start sites performed by Yamashita et 305 al 25 . This yielded a total set of 1,626 TOP gene candidates: within this set, 237 candidates have mouse 306 homologues that are known TOP genes, and this subset overlaps substantially with the 83-member set 307 of canonical TOP genes identified (54/83 mouse homologues/canonical TOP). We found these TOP 308 candidates to behave similarly to canonical TOP genes in terms of perturbed RA. In the strip plots in 309 Figure 6A, TOP genes and candidates within the significant targets (FDR<0.05) of each drug on the plate 310 are color-coded, allowing comparison of their differential RA between conditions. 311 We then constructed a simple translational regulatory network from our riboPLATE-seq data and 312 overlaid it with these canonical and novel TOP genes ( Figure 6B). We considered the genes 313 demonstrating significant reductions in RA due to treatment with PP242, MNK-i1, and BKM120 as the 314 positive translational targets of mTOR, PI3K, and MNK1, respectively, as these drugs are specific 315 inhibitors of these kinases. Here, we used a typical threshold for significance (FDR<0.05), but 316 additionally required target genes to have at least 20 average normalized read counts across all samples. 317 As expected, the targets of mTOR are enriched heavily for TOP genes and candidates. Nearly all known 318 TOP genes and candidates with mouse homology, as well as a significant majority of remaining TOP gene 319 candidates, are targets of mTOR. The largest fraction of each set is found in either the joint targets of 320 mTOR and PI3K or the exclusive targets of mTOR. The largest set of targets belongs to mTOR (228 321 exclusive/386 total), followed by PI3K (57 exclusive/213 total) and then MNK1 (21 exclusive/43 total), 322 and TOP genes and candidates comprise a larger fraction of mTOR targets than PI3K or MNK1 targets. 323

15
The set of targets common to all three kinases is also highly enriched for curated and novel TOP genes 324 (Fisher's exact test p=0.00006), suggesting a subset of TOP genes impacted by MNK1, though this three-325 way intersection is vastly smaller and less significantly-enriched for these genes than the two-way 326 intersection of mTOR and PI3K targets (Fisher's exact test p=6x10-24). Furthermore, the targets 327 exclusive to either MNK1 or PI3K are not significantly enriched for TOP genes and candidates (Fisher's 328 exact test p=0.760 and p=0.219, respectively) in contrast to the exclusive targets of mTOR (p=5x10-8). 329 Supplementary Table 2 details a more comprehensive statistical analysis of the network in Figure 6B, 330 which considers TOP genes and candidates separately, while the calculations above consider enrichment 331 across the combined set of genes and candidates (Additional File #1: Supplementary Ribosome association is frequently used to infer translational activity. This can be measured by sucrose 337 gradient fractionation of intact RNA in polysome profiling, or of digested monosomes and their 338 ribosome-protected footprints in ribosome profiling. Translation efficiency, defined as the rate of 339 protein production per transcript, is approximated differently in these two methods. In polysome 340 profiling, it is calculated as the ratio of transcripts that sediment in "heavy" vs "light" fractions, similar to 341 the ratio of ribosome association in riboPLATE-seq. Ribosome profiling refines this measurement with its 342 focus on ribosome footprints, calculating instead a per-transcript ribosome occupancy with additional 343 information about position, regional density, and ribosome arrest 26 . riboPLATE-seq sacrifices the specific 344 positional information provided by ribosome profiling for a general measurement of ribosome 345 association, obtained by IP rather than sucrose gradient fractionation. With pooled library construction, 346 greater throughput is possible with riboPLATE-seq than with either ribosome profiling or polysome 347 profiling. 348 349 Though it is more scalable than ribosome profiling, riboPLATE-seq is not without limitations. The lack of 350 resolution of individual ribosome positions means the method cannot resolve location-specific effects, 351 such as the effect of ribosome association in 5' leader sequences on translation in downstream coding 352 sequences 7 . More generally, riboPLATE-seq is insensitive to translational regulation at the levels of 353 elongation or termination due to its inability to distinguish active from stalled ribosomes, a limitation 354 common to many ribosome-centric measurements 27 including ribosome profiling. 355

356
In this study, we interrogated translational regulation in mitogenic signaling pathways in cancer cells. 357 Focal amplification of PDGFRA in TS-543 glioma neurospheres leads to constitutive activation of several 358 members of these pathways, including ERK, Akt, and PI3K 12 . We observed the expected results of mTOR 359 inhibition on translation with riboPLATE-seq, including decreased ribosome association with TOP genes, 360 and further clarified targets for the kinases PI3K and MNK1. We found PI3K to target a subset of the TOP 361 genes impacted by mTOR, without a strong impact on known TOP genes or candidate transcripts 362 separate from that shared with mTOR. This suggests the effect of PI3K on the TOP genes may be wholly 363 mediated by mTOR, consistent with the known organization of their signaling pathway. In contrast, 364 treatment with either the highly-specific MNK inhibitor MNK-i1 or the eIF4E inhibitor 4EGi-1 did not 365 significantly impact the TOP genes. Despite both drugs impacting the effector through which mTOR is 366 thought to mediate TOP gene translation, eIF4E, off-target effects of commonly-used MNK inhibitors in 367 past studies 22 may have overemphasized previous observations to this effect, and potential off-target 368 effects of 4EGi-1 28 complicate interpretation of its specific, eIF4E-dependent targets. This study serves as a proof-of-concept for larger-scale perturbation screens of potential translational 381 regulators. Here, riboPLATE-seq revealed signatures of specific translational targets for kinases in related 382 signaling pathways. Our results are consistent with the known structure of these pathways, including the 383 previously established mechanism by which mTOR controls translation of the TOP motif-containing 384 genes. However, the majority of the ~500 known protein kinases remain unstudied at the level of 385 translational regulation. The technology described here could enable a more comprehensive screen of 386 protein kinases and/or RNA binding proteins, allowing inference of translational regulatory networks 387 and de novo identification of regulatory motifs important to these networks that might be validated by 388 high-resolution techniques like ribosome profiling and CLIP-seq. We anticipate that the ability to dissect 389 these networks at scale will advance our understanding of translational regulation and the design of 390 specific therapies for diseases involving aberrant translation. 391

Cell Lysis 407
Following treatment, we centrifuged the plate of TS-543 for 7 minutes at 1800RPM on a Sorvall Legend 408 XTR at room temperature and removed supernatants by aspiration. Placing the plate on ice, we 409 resuspended the pelleted cells in each well in 30uL of polysome lysis buffer (20mM Tris-HCl, pH=7.4, 410 250mM NaCl, 15 mM MgCl2),0.1mg/mL cycloheximide, 0.5% Triton X-100, 1mM DTT, 0.5U/mL 411 SUPERase-In (ThermoFisher, AM2696), 0.024U/mL TURBO DNase (Life Technologies, AM2222), 1x 412 Protease Inhibitor (Sigma, P8340)), mixed 5 times by pipetting, and rested the plate on ice for 5 minutes. 413 We then centrifuged the plate for 5 minutes at 1400RPM at 4°C to remove bubbles before performing a 414 quick freeze-thaw, placing the plate first in a -80°C freezer and then resting at room temperature for 5 415 minutes each. Following an additional 10 minutes rest on ice, we viewed the plate under a microscope 416 to check the extent of cell lysis. At this point, we added standard RNA spike-ins (ERCC Spike-In Mix #1, 1 417 19 uL/well of 1:5000 diluted stock) (Life Technologies, #4456740) to half of the wells for spike-in depletion 418 experiments. We then prepared a new 96-well plate containing 3.5uL 2x TCL buffer (Qiagen, #1070498) 419 per well, to which we transferred 3.5uL of lysate (approximately 10% total volume). 420

Automated Pan-Ribosome Immunoprecipitation 421
To the remaining lysate, we added 1 uL of SUPERase-in (ThermoFisher, AM2696) and 1 uL of biotinylated 422 y10b antibody (ThermoFisher, MA516060) to each well, then sealed the plate and allowed it to incubate 423 while gently shaking for 4 hours at 4°C. During this incubation, we washed 500uL of Dynabeads MyOne 424 Streptavidin C1 streptavidin-coated magnetic beads (ThermoFisher, #65001) 3 times with polysome 425 wash buffer (20mM Tris-HCl (pH 7.4), 250mM NaCl, 15mM MgCl2, 1mM DTT, 0.1mg/mL cycloheximide, 426 0.05% v/v Triton X-100), using 1mL per wash and resuspending in 500uL. We added 5uL of washed 427 beads to each well, then incubated while gently shaking at 4°C for an additional hour. After this short 428 incubation, we placed the plate on a magnet, removed and reserved supernatants, and washed the 429 wells 3 times with 200uL per well of polysome wash buffer supplemented with 1uL/mL SUPERase-in on 430 the Biomek 4000 automated liquid handling system. 431 Following the final wash, we resuspended the beads in 15uL of ribosome release buffer (20mM Tris-HCl 432 (pH 7.4), 250mM NaCl, 0.5% Triton X-100, 50mM EDTA) per well. During a 15-minute incubation at 4C 433 on a Peltier module, with continuous pipet mixing on the Biomek 4000 in order to maximize elution, we 434 distributed 15uL of 2x TCL buffer to each well of a new 96-well plate. Finally, we replaced the sample 435 plate on the magnet and transferred eluants to the TCL-containing plate. 436

PLATE-seq Library Preparation and Sequencing 437
The plates of ribosome-associated and previously reserved total lysate in TCL buffer were submitted to 438 the Columbia Genome Center for processing by the previously-described PLATE-seq method of RNA-seq 439 library preparation 10 , which involves poly-A selection of transcripts, incorporation of sequence barcodes 440 20 in poly(T)-primed reverse transcription, and pooling for subsequent library preparation steps, generating 441 a single 3'-end RNA-seq library from each 96-well plate. We pooled total and ribosome-associated 442 PLATE-seq libraries and sequenced the combined libraries on the Illumina NextSeq 550 with a 75-cycle 443 high-output kit. With paired-end sequencing, the first read corresponds to the 3' end of a transcript, and 444 the second read contains the barcode identifying the library in which the read was obtained. 445

Ribosome Profiling and RNA Sequencing 446
We seeded TS-543 neurospheres in a 6-well plate at a starting density of 50,000 cells/mL in 2 milliliters 447 of NS-A complete medium per well, and allowed the plate to rest for 36 hours. After preparing PP242 448 and MNK-i1 solutions in DMSO as above, we treated two wells each with 625nM PP242, 1.0μM MNK-i1, 449 or DMSO vehicle for 6 hours in the tissue culture incubator. Following treatment, we transferred 450 samples to 15mL conical vials for centrifugation at 640 RCF for 7 minutes, then removed supernatants 451 and added 400 uL polysome lysis buffer (recipe above). After mixing by rapid pipetting, we transferred 452 samples to 1.8mL microcentrifuge tubes, rested them on ice for 5 minutes, and triturated by 5 passages 453 through a 23-gauge needle. Following a clarifying spin of 11K RCF for 10 minutes at 4C on a benchtop 454 centrifuge, we transferred supernatants to a new set of microcentrifuge tubes and discarded pellets. 455 Finally, we prepared ligation-free ribosome profiling and total RNA-seq libraries from these clarified 456 polysome lysates following the instructions provided with their respective kits (smarter-seq smRNA-seq 457 kit, Takara-Clontech; NEBnext Ultra-Directional II), augmented with our previously-published ligation-458 free ribosome profiling protocol 8 . We

Definition of Gene Sets of Interest 473
As PLATE-and riboPLATE-seq depend on isolation of RNA by poly(T) pulldown, they can only be used to 474 measure polyadenylated transcripts. We defined a set of questionably-polyadenylated transcripts by the 475 union of the set of non-polyadenylated and variably-polyadenylated genes identified in a screen of 476 polyadenylation status across the transcriptome 33 . We removed these genes from consideration in our 477 study to leave only consistently polyadenylated transcripts. We additionally obtained a set of known 5' 478 terminal oligopyrimidine motif-containing genes (TOP genes), as well as novel TOP candidates with and 479 without known TOP-containing analogues in mice, from a comprehensive search of transcription start 480 sites 25 . 481

Regularized Logarithm Normalization and Outlier Removal 482
After subsetting the count matrices for all libraries to remove counts for reads aligned to non-483 polyadenylated and spike-in transcripts, we constructed an overall count matrix of all 192 libraries for all 484 96 samples. We loaded this matrix into DESeq2 with corresponding column data describing the sample 485 ID, library type (ribo or RNA PLATE-seq), and drug treatment for each library. We then used the 486 After setting PLATE-seq as the reference level for library type and DMSO as the reference level for drug 509 treatment condition, we executed the DESeq2 function using fitType=local, then retrieved results for 510 each drug treatment and comparison of interest. For differential ribosome association, we used function 511 calls of the following format: 512 res <-results(dds, name=paste0("typeribo.condition",cond)) 513 Where dds is the DESeqDataSet object and cond is any of the drug treatments. The interaction between 514 library type and drug treatment is equivalent to the calculation of ribosome association (RA). For 515 differential expression, we considered only changes in gene abundance between PLATE-seq libraries 516 independent of riboPLATE-seq. This required function calls of the following format: 517 res <-results(dds, name=paste0("condition",cond)) 518 Where dds is the DESeqDataSet object and cond is any of the drug treatments. This isolates the main 519 effect of drug treatment, defined across total RNA PLATE-seq libraries only. 520 For ribosome profiling and RNA sequencing, we followed a similar workflow, utilizing a design formula 521 without sample pairing: 522 design = ~condition + type + condition:type 523 where condition corresponds to drug treatment (MNKi1, PP242, or DMSO) and type corresponds to 524 RNA-seq vs ribosome profiling. We then set the base level of type to RNA and the base condition to 525 DMSO, and executed DESeq2 with fitType=local. Finally, we retrieved results specific to the 526 interaction effect, here equal to the calculated change in translation efficiency (TE) as a result of drug 527 treatment, with the following two function calls: 528 res_242 <-results(dds, name = "typeribo.conditionPP242") 529 res_MNK <-results(dds, name = "typeribo.conditionMNKi1") 530

Identification of Perturbed Gene Ontologies by Gene Set Enrichment Analysis 531
We constructed ranked lists for gene set enrichment analysis (GSEA) using the results of these 532 differential expression and ribosome association analyses. We first removed any gene with invalid 533 results reported by DESeq2 in any drug-vs-control comparison (i.e. genes with assigned p-value 'NA') 534 from consideration. Next, we created a ranked list of genes for each drug treatment, using each gene's 535 log-fold change in RA as a rank statistic. We then performed a preranked GSEA on the C5 collection of 536 gene ontology terms against these ranked lists, using the Preranked option in the GSEA-P desktop 537 application 36 , with default parameters and scoring method set to 'classic'. 538

Network Visualization 539
To create a basic network, we interpreted the abundant genes exhibiting highly significant reductions in 540 RA under treatment with kinase inhibitors (FDR<0.01, baseMean>=20) as positive targets of the kinases 541 in question. We loaded these gene sets into CytoScape 37 (v2.7.1) as individual networks for each kinase, 542 merged the three networks, and used the yFiles 38 Organic automatic layout to organize the resulting 543 In each, these specific targets have largely shifted down and rightward, indicating less significant and 637 smaller perturbation in RA of these targets in combination. G-I) Scatterplots comparing the effect under 638 single-drug treatment and the difference between combination and single-drug effects for the same 639 target sets from (A-C). These plots show the relationship between initial effect size under single drug 640 treatment and the degree of attenuation or amplification in this effect under combined treatment, 641 excluding targets that change sign (Additional File #1: Supplementary Table 1 for details). The majority 642 of targets fall in the first and third quadrants of each plot, indicating attenuation of most targets, though 643 a greater fraction of targets in quadrants 2 and 3 for the combination of PP242+BKM120 suggests 644 additivity in some of their effects. The plots additionally demonstrate a consistent pattern of increased 645 attenuation (e.g. increased differences between combination and singular effect size) with increasing 646 single-treatment effect size (ρ = -0.48, p = 1.9*10 -32 for PP242+BKM120; ρ = -0.63, p = 9.7*10 -15 for 647 PP242+MNK-i1; ρ = -0.55, p = 5.2*10 -45 for BKM120+MNK-i1). 648 TOP genes and candidates significantly perturbed in RA by drug treatments. Strip plots along the X axis, 651 labeled for each drug treatment in our riboPLATE-seq study, contain log-fold changes in RA (Y axis) for 652 the genes exhibiting significant RA perturbations (FDR<0.05) under each treatment relative to DMSO 653