Adipocyte-like signature in ovarian cancer minimal residual disease identifies metabolic vulnerabilities of tumor-initiating cells

Similar to tumor-initiating cells (TICs), minimal residual disease (MRD) is capable of reinitiating tumors and causing recurrence. However, the molecular characteristics of solid tumor MRD cells and drivers of their survival have remained elusive. Here we performed dense multiregion transcriptomics analysis of paired biopsies from 17 ovarian cancer patients before and after chemotherapy. We reveal that while MRD cells share important molecular signatures with TICs, they are also characterized by an adipocyte-like gene expression signature and a portion of them had undergone epithelial-mesenchymal transition (EMT). In a cell culture MRD model, MRD-mimic cells showed the same phenotype and were dependent on fatty acid oxidation (FAO) for survival and resistance to cytotoxic agents. These findings identify EMT and FAO as attractive targets to eradicate MRD in ovarian cancer and make a compelling case for the further testing of FAO inhibitors in treating MRD.


54
The term minimal residual disease (MRD) was originally coined in relation to hematological 55 malignancies to define the leukemic cells that remain after treatment (1). More generally, tumor MRD 56 describes cancer cells that remain following complete clinical and radiological response to therapeutic 57 interventions (2). Such cancer cells share phenotypic and genomic characteristics with the bulk tumor 58 that existed prior to the intervention. In hematological malignancies such as chronic myeloid leukemia 59 (CML) and acute lymphoblastic leukemia (ALL), personalized treatment of MRD demonstrated the 60 feasibility of achieving long term responses and cures presumably by eliminating residual cancer cells 61 (3-5). Importantly, these examples have shown that rationalized switching of treatment to circumvent 62 the development of resistance in MRD can still help achieve long term benefits in patients. 63 However, the concept of treating MRD in solid tumors remains largely unexplored because of limited 64 understanding of the drivers of MRD survival. Current knowledge is largely obtained from preclinical 65 models rather than directly from patients and suggests that MRD survival is related to key characteristics 66 that define tumor initiating cells such as overexpression of ABC transporters and over-activity of 67 aldehyde dehydrogenases (ALDH) (6). However, the clinical relevance of these observations and the 68 presence of any additional survival mechanisms for MRD in patients has remained unknown because 69 of the difficulty in isolating and characterizing MRD cells. 70 71 It is important to make a distinction between measuring the MRD load and the molecular 72 characterization of MRD cells. The former is largely a diagnostic process for predicting the probability 73 of recurrence while the latter aims to understand driving survival pathways and potential tumor 74 vulnerabilities for therapeutic intervention (2). Recent advances in isolating and quantifying circulating 75 tumor DNA and circulating tumor cells, among other technologies, have made it possible to predict the 76 load of MRD and the probability of recurrence with high precision (7,8). These methods could also be 77 extended to detect the presence or evolution of known mechanisms of resistance to therapies such as 78 the development of resistance mutations in the active site of a kinase that is being targeted 79 therapeutically (9). However, such approaches would only be helpful for predicting response to a 80 limited number of therapeutics (2). Conversely, an unbiased molecular characterization of MRD would 81 be ideal for the discovery of novel therapeutic strategies to treat MRD. In many hematological 82 malignancies, it is possible to sample MRD by obtaining bone marrow biopsies. In contrast, the 83 selection of the biopsy sites to harvest MRD from solid tumors is much more challenging since it is 84 difficult to predict their site of residence. 85 To overcome the aforementioned limitations, we designed a prospective observational study in patients 86 with advanced high grade serous ovarian cancer (HGSOC) allowing us to sample and characterize 87 MRD. Given HGSOC's relatively short latency before recurrence and its tendency not to spread outside 88 of the abdominal cavity, sampling MRD from the peritoneal cavity provides an opportunity to 89 characterize clinically relevant MRD lesions. To define MRD, we applied strict criteria of complete 90 responses that are based on clinical and radiological evidence, direct visualization of the peritoneal 91 cavity and histopathological evidence of significant response. We used Laser Capture Microdissection 92 (LCM) to enable further in-depth characterization of MRD sites. This work enabled us to identify a 93 highly selected and pure population of tumor cells that faithfully represent MRD directly in patients. 94 Our results provide potential opportunities for therapeutic intervention to treat MRD in patients with 95 HGSOCs. 96 97

98
The Oxford Ovarian Cancer Predict Chemotherapy Response (OXO-PCR) study: a patient cohort 99 The study of MRD in ovarian cancer or other solid tumors in patients has been difficult because of the 100 inability to identify and sample microscopic deposits intra-operatively or by using traditional imaging 101 In order to compare the gene expression profiles across response groups, RNA-seq was performed on 129 cancer islets isolated using LCM from all samples collected at both timepoints. Due to the abundance 130 of tumor cells before treatment and in the "post-chemo" samples in sites where there was evidence of 131 poor response, scrolls of their samples were also collected to be used for bulk RNA-seq analyses ( Figure  132 1C). Both LCM and bulk RNA-seq pipelines included multiple quality control steps to avoid the 133 contamination from surrounding non-cancer tissue ( Figure S1A, S1B). After QC filtering (which 134 removed 37 out of the total 156 libraries), differential expression analyses were carried out across 135 timepoints and response groups as well as on each patient individually ( Figure 1C). 136 137 Pseudotime analysis reveals limited intra-patient heterogeneity 138 We first sought to evaluate how representative the sample set is to the known molecular profile of 139 HGSOC. To do this we used unsupervised pseudotime analysis (11) of the pre-chemotherapy sample 140 set and compared this to data obtained from the TCGA HGSOC dataset (12) which comprised, 141 predominantly, of pre-chemotherapy samples. This comparison revealed that our set clustered around 142 the centre of the pseudotime gradient of the TCGA cohort ( Figure S2A) indicating that the dataset is 143 highly representative of HGSOCs. Next, we examined the pseudotime data of the entire dataset (pre-144 chemo and post-chemo) and found that samples from the same patient clustered together on the 145 pseudotime gradient (Figure 2) in spite of the analysis being conducted in an unsupervised manner 146 without taking into account the patient identity or the timing of sampling. This result was also consistent 147 with the t-distributed stochastic neighbor embedding (t-SNE) of the entire dataset following batch 148 correction ( Figure S2B), strongly suggesting that, despite the existence of intra-patient heterogeneity, 149 the gene expression diversity observed in the OXO-PCR cohort is clearly dominated by inter-patient 150

heterogeneity. 151
We next identified pseudotime-dependent genes by fitting a linear model and highlighting genes that 152 are differentially expressed along the pseudotime gradient independent from chemotherapy effect. 153 Pathway analysis revealed an enrichment of genes that are involved in interferon signaling pathways, 154 L1CAM and MAPK pathways ( Figure S2C). Importantly, after accounting for the pseudotime effect, 155 and, as expected, analysis of the chemotherapy effect revealed a significant downregulation in FOXM1 156 expression, a master regulator of the expression of genes involved in mitosis (12). There was also 157 evidence of downregulation of the corresponding mitotic signature that is known to be highly expressed 158 in HGSOC (12), with downregulation of known mitosis genes such as AURKB, NCAPH, NCAPG, 159 Cyclin B2 and several kinesins (Table S2). Overall, these data show the robustness of the approach 160 chosen for the study and highlight that transcriptional heterogeneity is predominantly observed between 161 patients rather than within individual patients. 162

163
Transcriptomic signatures related to tumor initiating cells and lipid metabolism characterize HGSOC 164

MRD cells 165
We next sought to evaluate MRD in exceptional responders (patients 11152, 1016 and 1036), who all 166 had no visible residual disease at the post-chemotherapy laparoscopy. We compared the gene expression 167 profiles of LCM samples obtained before and after treatment from these tumors ( Figure 3A). This 168 analysis identified 356 differentially expressed genes ( Figure 3A, Table S3). 169 The post-chemotherapy tumor cells showed significantly higher expression of ATP-binding cassette 170 transporters (ABCA12, ABCB5, ABCA9, ABCA6, ABCA10, ABCA8; log fold change > 3.3; p value < 171 1.68E-04) as well as other known markers of tumor initiating cells (TICs) (ALDH1L1, ALDH1A1, 172 ALDH2, MS4A1/CD20; log fold change > 1.8; p value < 9.4E-04) ( Figure 3B) (13,14). 173 These results suggest that MRD has characteristics that are consistent with previously identified features 174 of tumor initiating cells from preclinical models. Surprisingly, we also found a significant increase in 175 the expression of genes that are involved in lipid metabolism such as PLIN1, PLIN4, CD36, ACACB, 176 G0S2, LIPE, LPL, GPAM, SCD ( Figure 3B) (log fold change > 3.7, p value < 2.1E-04). Notably, the 177 MRD cells were also characterized by the upregulation of 60 different small nucleolar RNAs (Table  178 S3), non-coding RNAs whose traditional role is to guide the post-transcriptional modification of 179 ribosomal and small nuclear RNAs. More recently, however, snoRNAs have been shown to also play 180 important roles in tumorigenesis and in the regulation of lipotoxic and oxidative stress responses (15). 181 Importantly, the expression of HGSOC marker genes such as PAX8, MUC16 and WT1 or the epithelial 182 marker EPCAM was maintained after chemotherapy, indicating that the MRD LCM cells kept their 183 HGSOC identity ( Figure S3A). 184 Moreover, to rule out the possibility of cross contamination of MRD with adipocytes, we attempted to 185 perform LCM and RNA-seq on large areas of adipose tissue adjacent to the MRD lesions and compare 186 the RNA expression results. However, this did not yield sufficient RNA for downstream analysis. We 187 conclude that the possibility of contamination of MRD with small number of adipocytes is highly 188 unlikely to have biased the differential expression analysis. 189 We next examined the differentially expressed genes per individual patient from the exceptional 190 responders. We noted that for patient 1016, who had the most notable microscopic response, the main To determine whether the features described above are specific to MRD or shared among all 204 chemotherapy resistant cells, we compared the LCM data from the exceptional and the poor responders 205 after treatment. Even though both cell populations survived NACT, significant transcriptional 206 differences were observed with 867 genes found to be differentially expressed between the two groups. 207 Notably, the MRD samples showed upregulation of genes related to lipid metabolism and those 208 previously known to be associated with TICs ( Figure 3D, S3F) (log fold change > 2.7, p value < 0.001) 209 as well as snoRNAs (Table S4). These results further support the notion that fatty acid metabolism is 210 specifically upregulated in MRD highlighting a previously unrecognized feature of these cancer cells 211 in HGSOC patients. 212

The transcriptome of MRD cells resembles differentiated adipocytes 213
The observation that the identified MRD-upregulated genes are involved in both anabolic and catabolic 214 lipid metabolic processes implied that the purpose of such upregulation was not to simply increase ATP 215 uptake following chemotherapy. Instead, these observations pointed to a more complex transcriptional 216 program of MRD cancer cells that may contribute to the acquisition of chemotherapy resistance. These 217 transcriptional changes were reminiscent of those observed in adipocytes where both synthesis and 218 turnover of storage lipids such as triacylglycerols are highly active. To test the hypothesis that MRD 219 cancer cells selected by chemotherapy are adipocyte-like, we compared the transcriptional changes to 220 those that occur during differentiation of adipocytes from fibroblast-like precursors. To this end, we 221 differentiated 3T3-L1 cells into adipocytes as previously described (17). To monitor transcriptional 222 changes, we performed RNA-seq at day -2 (pre-adipocyte/fibroblast stage), day 0 (start of the 223 differentiation protocol) and day 6 (adipocyte stage) ( Figure 4A). Strikingly, we found that the 224 expression of key adipocyte markers such as CD36, PLIN1, CIDEC, LIPE, LPL and ACACB, all 225 strongly upregulated upon adipocyte differentiation ( Figure S4), correlated significantly with the 226 expression changes observed in exceptional responders (Pearson's correlation coefficient of 0.8, p value 227 of 0.03) ( Figure 4B). In contrast, over-expression of genes that were known to be upregulated in TICs 228 such as ABC transporters was only observed in MRD ( Figure 4C) and not during adipocyte 229 differentiation. These results strongly suggest that MRD cancer cells, while retaining features of TICs, 230 reflect a transcriptional state that resembles adipocytes. 231 Altogether, these findings represent the first in vivo characterization of MRD cells isolated from 232 HGSOC patients and identify specific markers that are unique to this population of chemotherapy 233 resistant cells. 234

MRD cells show mesenchymal characteristics 236
We observed that the resistant cancer cells that were laser-captured from exceptional responders prior 237 to RNA sequencing showed an elongated and spindle-like shape in contrast to the more rounded 238 appearance of cancer cells isolated from poor responders. The appearance of MRD cancer cells was 239 consistent with that of mesenchymal cells suggesting that they may represent epithelial-to-240 mesenchymal transition (EMT). This was a reasonable assumption given that chemotherapy resistance 241 (18,19) and the acquisition of stem cell properties (20, 21) have been clearly associated with cancer-242 related EMT. 243 To test this assumption, we next quantified the EMT cell state in the OXO-PCR samples using our 244 recently described deconvolution-based classifier (22,23). This analysis revealed that MRD cells are 245 particularly enriched in genes belonging to the EMT signature. Specifically, among all the post-246 chemotherapy samples, those isolated from exceptional responders showed the highest EMT score 247 ( Figure 5A). These MRD samples were characterized by a very high proportion of the EMT-high cell 248 state (EMT fraction > 0.85) compared to other cell states, regardless of whether a high EMT level was 249 already widely observed before treatment (as in patient 1016) or not (as in patients 11152 and 1036) 250 ( Figure 5B). In contrast, the samples from poor responders were more heterogeneous after 251 chemotherapy, showing the co-existence of multiple cell states (e.g. EMT-high, differentiated, Krt17, 252 Given the EMT enrichment in MRD, we reasoned that the adipocyte-like gene signature observed in 254 MRD may be a defining feature of the EMT-high cancer cell state, selected in the chemotherapy-255 resistant EMT-high MRD cells ( Figure 5C). The alternative explanation is that either the adipocyte-like 256 state, the EMT-like state or both are induced by chemotherapy ( Figure 5C). To test these alternatives, 257 we compared the adipocyte-like gene signature between EMT-high and EMT-low pre-chemotherapy 258 tumors using publicly available datasets of pre-chemotherapy HGSOC, The Cancer Genome Atlas 259 (TCGA)(12) and the Australian Ovarian Cancer Study (AOCS) (24). We divided the samples according 260 to our previously described EMT score (22, 23) into EMT-high and EMT-low. 261 Our analysis indicated that many genes from the adipocyte-like gene signature showed significantly 262 higher expression in the EMT-high group ( Figure 5D). This suggests that EMT-high cancer cells are 263 enriched in genes related to lipid metabolism and that this cell state becomes selected for after 264 chemotherapy treatment. However, our results cannot completely rule out the alternative explanations 265 that the adipocyte-like gene signature or the EMT-high signature are, at least in part, induced by 266

chemotherapy. 267
Collectively, these data highlight the mesenchymal characteristics of MRD that encompass an elevated 268 adipocyte-like signature. We speculate that active lipid metabolism might confer a survival advantage 269 for chemotherapy-resistant MRD. Similar results were also obtained at the end of treatment timepoint compared to untreated cells ( Figure  295 S6D-S6E), indicating that OXPHOS plays an important role in chemotherapy resistant cells. 296 To elucidate which substrates are key for this process and evaluate a potential role of FA in the survival 297 of MRD-mimic cancer cells, we first blocked FAO using etomoxir, an inhibitor of the carnitine 298 palmitoyl transferase (CPT1) that imports FA into mitochondria for β-oxidation (25). Colony forming 299 assays were performed using OVCAR5 and OVCAR8 cells treated with etomoxir concentrations 300 previously shown not to elicit off-target effects (26). 301 This approach revealed that the MRD-mimic cells of both cell lines displayed a significantly higher 302 sensitivity than their carboplatin-untreated counterparts ( Figure 6E-S6F). The results were also 303 confirmed using perhexiline, another CPT1 inhibitor currently used in the clinic as a prophylactic 304 antianginal agent (27) ( Figure 6E-S6F), showing that FAO is indeed crucial for chemotherapy resistant 305 cells, as already suggested by the transcriptome analyses described above. 306 Given that the lipid signature observed both in vivo and in vitro included not only genes related to FAO 307 but also FA synthesis, we performed mass spectrometry-based lipidomics (28) to determine if the 308 transcriptional increase of genes involved in de novo lipogenesis is reflected in an increase of storage 309 lipids such as triacylglycerols (TAGs). Interestingly, for both OVCAR5 and OVCAR8 cell lines, the 310 total concentration of glycerolipids, which comprises both TAGs and their immediate precursor 311 diacylglcyerols, did not differ significantly before and after treatment with carboplatin ( Figure S6G). 312 This suggests that, unlike FAO, the transcriptional changes observed in the FA synthesis pathway do 313 not lead to an increase in lipid storage or that the newly synthesized FA are immediately oxidized and 314 therefore do not accumulate, as shown in the lipidome analysis. 315 Taken together these data confirm the robustness of our in vitro model and highlight that FAO is Fatty acid oxidation is a general mechanism of resistance in MRD that is independent from the 322 cytotoxic agent 323 Both our in vivo and in vitro data show that the ovarian cancer cells that survive DNA-targeting 324 cytotoxic (carboplatin) or microtubule-stabilizing (paclitaxel) chemotherapy treatment are 325 characterized by a transcriptional upregulation of their lipid metabolic pathways that appears to be a 326 survival mechanism in MRD cancer cells. However, whether these observations represent a general 327 survival mechanism under cytotoxic treatment or are more specifically related to the chemotherapeutics 328 used remained unclear. 329 To investigate if the upregulation of lipid metabolism was a general survival mechanism in ovarian 330 cancer cells, we tested the effect of poly-ADP ribose polymerase (PARP) inhibitor treatment on ovarian 331 cancer cell lines. We selected this cytotoxic agent because its use is rapidly becoming standard of 332 practice in patients with HGSOC, due to the defects in the homologous recombination repair pathway 333 often found in these tumors (29). Moreover, it is known that PARP activation decreases the 334 concentration of nicotinamide adenine dinucleotide (NAD+) and this has been linked to lipid 335 accumulation (30). 336 First, we measured the expression of key lipid metabolism genes using qPCR following olaparib 337 treatment. KURAMOCHI and OVCAR5 cells showed an upregulation of PPARA and genes belonging 338 to the CPT family at both concentrations of olaparib that were tested ( Figure 7A). In hepatocyte, 339 adipocyte, and myoblast cells, it was reported that PARP inhibitors activate the expression of FAO-340 related genes through SIRT1 activation (31, 32); however, in our system, SIRT1 knockdown did not 341 change the upregulation of FAO genes upon treatment with olaparib ( Figure S7A). 342 Next, we examined if the inhibition of FAO had any effect on the sensitivity of ovarian cancer cells to 343 olaparib. The colony formation ability for both KURAMOCHI and OVCAR5 lines was significantly 344 reduced when the cells were treated with a combination of etomoxir and olaparib compared with single 345 treatment ( Figure 7B-7C). Similar results were also obtained for OVCAR8 and SKOv3 cell lines 346 ( Figure S7B). 347 These findings indicate that the upregulation of lipid metabolism may be a more general mechanism 348 through which ovarian cancer cells survive cytotoxic stress as it is not restricted to any particular type 349 of chemotherapeutic treatment. Furthermore, inhibiting fatty acid oxidation may represent a therapeutic 350 strategy to enhance the efficacy of cytotoxic treatment in HGSOC. 351 Moreover, these findings suggest that preventing cells from entering into an adipocyte-like cell state 352 could represent a new therapeutic strategy to sensitize ovarian cancer cells to cytotoxic treatment. analyzing preclinical models. In this study, we have designed a clinical trial to specifically address these 361 issues and successfully obtained a pure collection of MRD samples that enabled us to gain informative 362 insights about MRD biology in ovarian cancer patients. Our analysis revealed a previously 363 unrecognized adipocyte-like signature in MRD in HGSOC. We have complemented our in vivo 364 approach with validation in an in vitro MRD-mimic model that we developed. Using this model we 365 demonstrate that MRD upregulates fatty acid oxidation and that the specific inhibition of this process 366 significantly synergizes with chemotherapeutics, increasing their cancer cell killing potential. We show 367 that upregulation of fatty acid oxidation seems to be a general survival mechanism in MRD following 368 chemotherapy or PARP inhibition and thus, despite the small number of patients analyzed and the 369 limited mechanistic studies, our results have important therapeutic implications. 370

371
To the best of our knowledge, this work represents the first comprehensive characterization of MRD 372 from HGSOC clinical samples. Through an LCM-guided RNA-seq approach, we demonstrate that these 373 microscopic tumor foci, isolated from exceptional responders after NACT, not only have features of 374 TICs but also show altered lipid metabolism that has clinical relevance. Alongside a marked EMT 375 phenotype and the upregulation of several genes belonging to the ABC and ALDH families, these cells 376 have increased expression of the desaturase SCD, which is consistent with previous observations that 377 ovarian TICs have high levels of unsaturated lipids (33). The identification of several transcriptomic 378 features that are typical of TICs strongly supports the hypothesis that MRD in the peritoneal cavity is 379 indeed responsible for relapse. Targeting these residual chemotherapy-resistant cells would therefore 380 be highly promising. Our data suggest that targeting lipid metabolism could represent an attractive 381 therapeutic option, similar to what has been observed in vitro or in preclinical models of other cancer 382 types (34-37). 383 384 Using our deconvolution-based classifier (22, 23), we have shown that MRD from the exceptional 385 responders is enriched in EMT-high cancer cells. The EMT process is known to facilitate tumor 386 progression. For example, several mechanisms through which EMT induces stemness have now been 387 elucidated (38). Moreover, metabolic reprogramming has been associated with EMT plasticity (39) and 388

TGF-β1-induced mesenchymal cells display a shift from glycolysis to OXPHOS (40). A similar 389
metabolic shift has been previously described in breast (41) and pancreatic (42) cancers using mouse 390 models of oncogenic pathway inhibition to mimic MRD. However, the transcriptomics changes 391 described in those studies were not as extreme as the adipocyte-like signature observed here, possibly 392 due to differences across species or the specific organ tropism of ovarian cancer. 393

394
The perturbation of lipid metabolism observed in MRD can be explained by at least two models that 395 are not necessarily mutually exclusive. The first one is that a sub-population of cells in the primary 396 tumor already expresses the adipocyte-like gene signature and that these cells become selected upon 397 treatment because such altered metabolism confers a survival advantage for MRD. Given the EMT 398 enrichment in MRD, one might argue that the adipocyte-like gene signature is an inherent feature of the 399 EMT-high cancer cell state. Through the analysis of publicly available datasets of pre-chemotherapy 400 HGSOC, we have shown that this might indeed be the case, since many genes from the adipocyte-like 401 signature showed significantly higher expression in EMT-high tumors. Additional evidence supporting 402 this idea of a selection process is provided by the observation that primary pre-chemotherapy HGSOC 403 displays OXPHOS metabolic heterogeneity: the high-OXPHOS group exhibits better short-term 404 survival because its chronic oxidative stress makes it more sensitive to chemotherapy (43). This is 405 consistent with the initial good clinical response observed in the exceptional responders; however, we 406 would argue that the high-OXPHOS cells that survive treatment may eventually lead to recurrence 407 because they have found mechanisms to reduce oxidative stress such as the activation of a temporary 408 dormant state. 409 An alternative hypothesis is that cells with altered regulation of lipid metabolism are absent before 410 chemotherapy and that this metabolic rewiring occurs in response to chemotherapy. Some degree of 411 chemotherapy induction cannot be excluded from our data and detailed time-lapse metabolic analysis 412 will be needed to further investigate this possibility. 413 Irrespective of how this adipocyte-like signature becomes so preponderant in MRD (whether it is by 414 selection or induction), the implications for treatment remain clear and suggest that the inclusion of 415 therapeutics targeting fatty acid oxidation may be beneficial for HGSOC patients. The successful 416 inhibition of CPT1, for example, could represent a new therapeutic approach to sensitize ovarian cancer 417 cells to different cytotoxic treatments, such as carboplatin and olaparib. In addition, our work clearly 418 shows that the MRD cells have marked mesenchymal characteristics. It is now widely recognized that 419 EMT causes resistance to several anti-cancer agents, spanning from chemotherapy to immunotherapy. 420 All research efforts to tackle EMT-induced multi drug resistance have so far focused on strategies to 421 prevent or reverse EMT (44) and only recently the significant metabolic rewiring associated with EMT 422 has started to gain attention as a potential therapeutic target (45). Our findings provide new insights in 423 this ongoing debate and proposes an alternative way forward to overcome EMT-related resistance, at 424 least in MRD. 425

426
In conclusion, we suggest that targeting fatty acid oxidation may be an attractive strategy to eradicate 427 MRD in HGSOC and improve the long-term survival of exceptional responders. 3T3-L1 pre-adipocytes were treated as previously described (17)  Cells were harvested in HKM buffer (50 mM HEPES, 50 mM KOH, 150 mM KCl, 5 mM MgCl2, pH 514 7.5) and lipidomics analysis was performed as previously described (17,28,46). 515 Preprocessing of RNA-seq data 516 Sequencing reads from fastQ files were trimmed for adapter sequences and quality with Trim Galore!, 517 mapped to the UCSC hg19 human genome assembly using STAR (v2.4.2a) and read counts were 518 obtained using subread . 519 520

Pseudotime analysis 521
We used the R package PhenoPath (11) to perform the pseudotime analysis that projected the high-522 dimensional transcriptomic data to one dimension, in which we compared the OXO-PCR cohort and 523 the TCGA dataset. i.e. scores, were calculated by applying the linear support vector regression, which was incorporated in 540 the CIBERSORT function, on the raw expression profiles of each tumor sample. The deconvolution 541 analysis was performed in the relative mode and, thus, for each tumor the scores of five molecular 542 signatures added up to one. 543 Analysis of TCGA and AOCS data 544 TCGA ''IlluminaHiSeq UNC'' RNA-seq dataset (version: 2017-10-13) was downloaded from the 545 UCSC Xena Data Hub (https://tcga.xenahubs.net) (48) (12). The AOCS dataset was downloaded from 546 GSE9899 (24). TCGA and AOCS data were transferred to a non-log-linear space and then deconvolved 547 in the same way as the OXO-PCR RNA-Seq data. Samples were partitioned into three groups, EMT-548 low, -middle and -high groups for each dataset. We compared the expression of five genes related to 549 lipid metabolism between EMT-high and EMT-low samples by using differential expression analysis 550    concentrations that achieved more than 90% cell killing (end of treatment timepoint) after which the 804 surviving cells were allowed to recover in regular medium for an additional 2 weeks (MRD-mimic 805 timepoint). 806