Massively parallel interrogation of protein fragment secretability using SECRiFY reveals features influencing secretory system transit

While transcriptome- and proteome-wide technologies to assess processes in protein biogenesis are now widely available, we still lack global approaches to assay post-ribosomal biogenesis events, in particular those occurring in the eukaryotic secretory system. We here developed a method, SECRiFY, to simultaneously assess the secretability of >105 protein fragments by two yeast species, S. cerevisiae and P. pastoris, using custom fragment libraries, surface display and a sequencing-based readout. Screening human proteome fragments with a median size of 50 - 100 amino acids, we generated datasets that enable datamining into protein features underlying secretability, revealing a striking role for intrinsic disorder and chain flexibility. SECRiFY is the first methodology that generates sufficient amounts of annotated data for advanced machine learning methods to deduce secretability predictors. The finding that secretability is indeed a learnable feature of protein sequences is of significant impact in the broad area of recombinant protein expression and de novo protein design.

protein folds after entry in the ER. Mass spectrometry (MS)-based proteomics is the 48 predominant approach for the interrogation of post-translational events, but despite many 49 technological advances, its breadth and depth is limited and decreases steeply with sample 50 complexity; in routine MS setups, generally less than 70% of all transcribed mammalian 51 protein-coding genes are detected 12,13 , and full protein coverage is rarely achieved. 52 In practice, the absence of such an integrated picture of the secretory system is most apparent 53 in the unpredictability of heterologous protein secretion. Indeed, four decades after the 54 recombinant DNA revolution, obtaining detectable levels of functional protein in a particular 55 heterologous host system, secretory or not, is still principally a process of trial and error. 56 Statistics from worldwide Structural Genomics Consortia initiatives suggest that the likelihood 57 of progression from cloned target to soluble, purifiable protein lies between 20-50%, even with 58 target selection 14,15 . These low and unpredictable expression success rates slow down 59 progress in the many fields of basic and applied life science where recombinant protein 60 production comes into play. Although expression screening of parallel constructs or variant 61 libraries of a protein of interest to increase heterologous protein expression success rates has 62 gained momentum [16][17][18][19] , it has to be repeated for each new target. Additionally, alternative 63 comprehensive strategies to assess heterologous expression across entire proteomes have 64 generally been limited to intracellular expression in E. coli, small proteomes, and cumbersome 65 clone-by-clone strategies 20-23 . As the availability of large-scale protein expressability data will 66 help to demystify which and why proteins fail to express, new methods for measuring 67 expression in high throughput are needed. 68 We here developed an approach to evaluate the secretory potential ('secretability') of proteins 69 on a proteome-wide scale. SECRiFY (SECretability screening of Recombinant Fragments in 70 Yeast) combines yeast surface display screening of protein libraries and a deep sequencing 71 readout, enabling the systematic identification of heterologous polypeptides that can pass (or 72 evade) the secretory quality control checkpoints of the yeast ER, Golgi, secretory vesicles and 73 plasma membrane, and be secreted. As a first fundamental question to be addressed using 74 this new methodology, we asked whether, given a particular sequence of a protein fragment, 75 we could 1) learn what features contribute to its secretability and 2) generate machine-learned 76 classifiers that predicts its secretability. Hence, we fragmented the human proteome and 77 screened these fragments for secretability in two different yeast species, Saccharomyces 78 cerevisiae and Pichia pastoris (Komagataella phaffii). We generated a large repository of more 79 than 20,000 experimentally determined yeast-producible human protein fragments, available 80 to the research community at https://iomics.ugent.be/secrify. These datasets of experimentally 81 verified secretable fragments will assist experimental design of recombinant protein secretion. 82 We used them to train (deep) machine learning models for secretability prediction, which 83 unveiled sequence and structural determinants of productive secretory system transit, 84 highlighting the utility of SECRiFY to provide further insight into the basic mechanisms of 85 secretory processing. More application-focused implementation of SECRiFY (focusing on 86 fixed-boundary protein domains or multi-domain fragments) should enable to generate 87 databases of experimentally validated secretable native protein domain fragments, which 88 could tremendously speed up experimental protein expression in many fields of study. 89 90

Normalized fragment libraries for screening at domain-level resolution 92
Multi-domain proteins often fail to express or secrete in their entirety due to local issues with 93 misfolding of particular protein areas, translation inhibitory sequences, protease susceptibility, 94 the absence of stabilizing interaction partners or modifications, or toxicity. The structural, 95 functional and evolutionary modularity of proteins in domains, however, implies that individual 96 expression of certain protein parts, especially domains, can often nonetheless be achieved. 97 Chopping up difficult proteins into experimentally tractable fragments has been exploited by 98 structural biologists for years, both in rational target design as well as in random library screens 99 for soluble expression 24-28 . Moreover, screening of protein domains or fragments can provide 100 valuable information that is not immediately attainable or obvious from screens with full-length 101 proteins 29 . Some domain-focused interactome studies, for example, have allowed immediate 102 delineation of the minimal interacting regions and the detection of more interactors without 103 increasing the number of false positives 30 . We thus rationalized that screening libraries of 104 domains or domain-sized polypeptides, rather than full-length proteins, would allow for a higher 105 resolution measurement of secretability across proteomes, and facilitate the identification of 106 sequence or structural determinants of secretion. 107 Domain boundary prediction, however, is notoriously inaccurate, and even with a reliable 108 estimate, small variations in the exact N-and C-terminus of the fragment can lead to dramatic 109 differences in expressability 31 . Random approaches, on the other hand, can generate libraries 110 of fragments that encompass most domains of a proteome by sheer oversampling. We 111 therefore designed and built directional, randomly fragmented cDNA libraries covering the 112 human transcriptome with fragments coding for approx. 50-100 amino acids, which is the 113 median domain size of human proteins (Fig. 1a, c). Due to the large dynamic range of mRNA 114 transcripts in human cells (in abundance of 4 orders of magnitude), however, capturing the full 115 diversity of fragments would require unfeasibly large libraries, even at 100 bp resolution (Fig.  116   1b). We therefore reduced fragment abundance differences by relying on the second-order 117 kinetics of nucleic acid rehybridization after denaturation, and subsequent digestion with the 118 Kamchatka crab duplex-specific nuclease (DSN) 32,33 (Fig. 1c-d). More abundant DNA species 119 rehybridize faster and are therefore digested first; as such, even a single round of normalization 120 substantially reduces abundance differences between DNA fragments (Fig. 1g). Crucially, this 121 allowed us to downsize the libraries to a scale that is feasibly compatible with downstream 122 cDNA library cloning and yeast transformation efficiencies (+/-5*10 6 -5*10 7 ). 123 To ensure directionality, random primers were tagged with a rare-cutter restriction site (PacI), 124 which is distinct from the SfiI site incorporated in the library adapters (Fig. 1c). We initially 125 observed that the random primer tag sequence is susceptible to degradation due to endo-and 126 exonuclease activity of the E. coli DNA polymerase I during second strand synthesis 34-37 . As a 127 result, less than 20% of fragment sequences contained a full-length PacI site, negatively 128 affecting ligation into the surface display vector (Fig. 1e). Both nuclease-resistant 129 phosphorothioate bonds and buffer sequences could partially protect the tag from degradation, 130 and for the final library design, we settled on the primer where protection efficiency was 131 maximal (Fig. 1e, grey bar). 132 Tag composition also affected abundance normalization efficiencies. In an earlier design with 133 a GC-rich tag, normalization was less effective (Fig. 1f) than the design with PacI tag (Fig.  134   1g), where an approx. 1000-fold normalization could routinely be obtained. The tag sequence 135 is present on all sequence fragments, and most likely, when using a GC-rich tag, 136 rehybridization kinetics (and therefore degradation) is dominated by the tag rather than by the 137 sequence of the fragments themselves. 138 In all, this library construction protocol allows for efficient capture of protein-coding fragments 139 tiled along eukaryotic transcriptomes. It is, to our knowledge, the first time that an effective 140 method for normalization of tagged random-primed cDNA fragment libraries is reported, and it 141 should find many applications in areas where the protein-coding potential of a cell needs to be 142 effectively covered in expression libraries. 143 144

SECRiFY as a platform for secretability screening in yeast 145
Relying on the sophisticated quality control (QC) machinery of the eukaryotic secretory system, 146 which ensures efficient degradation of unstable or misfolded proteins before reaching the 147 plasma membrane 19,20 , we further reasoned that surface display could be used as a proxy for 148 productive secretion. As such, once cloned into the surface display vector and transferred to 149 yeast, library polypeptides are directed to the secretory system by an N-terminal secretory 150 leader sequence derived from the yeast a mating factor (MFa1 prepro), and furthermore on 151 the yeast cell wall via C-terminal fusion to the GPI-anchoring region of the S. cerevisiae cell 152 wall protein Sag1 ( Fig. 2a-b). Fragments for which the fragment-Sag1 fusion successfully 153 passes (or escapes) secretory system QC without proteolytic degradation are recognized 154 through their N-and C-terminal epitope tags (FLAG and V5, resp.), and are segregated from 155 the rest using iterations of high-efficiency magnetic-and fluorescence-activated cell sorting 156 (MACS/FACS) (Fig. 2b-c). Finally, fragment identification and classification are achieved by 157 deep sequencing of fragment amplicons from the unsorted and sorted cell populations. In 158 short, SECRiFY assesses secretability, i.e. the potential of a polypeptide to transit through the 159 secretory system of ER, Golgi, vesicles, and PM without degradation, in a manner that is 160 independent of the original endogenous localization of the protein of interest. For the present 161 study, we focus on basic principles of secretability. While in practice, any protein-coding mRNA 162 pool is compatible with SECRiFY, considering its biomedical importance and structural 163 complexity, we here focused on the human proteome for our screens, as encoded by the 164 transcriptome of various human cell lines. 165 We first benchmarked the method by building a 1.96*10 6 clone fragment library of the 166 HEK293T transcriptome and performed triplicate screens in S. cerevisiae. On average, 1.76% 167 ± 0.12% of library cells displayed a fragment with an intact N-terminus (FLAG-tag) and intact 168 C-terminus (V5-tag) (Supplementary Fig. 1). Accounting for a 1/9 chance of up-and 169 downstream in-frame cloning, this means that approximately 15.8% of in-frame fragments are 170 detectably displayed and hence, potentially secretable. After a 32-fold enrichment of these 171 double positive cells through a single round of MACS and two subsequent rounds of FACS 172 ( Supplementary Fig. 1), both pre-and post-sort population were sequenced at a per-base 173 average coverage of minimally 150 reads. On average, 1.12*10 6 unique fragments/replicate 174 were detected, covering on average 26.45% ± 0.86% of the human canonical transcriptome 175 with at least three reads (Supp . Table 1-3). To assess the secretion-predictive value of the 176 method, we picked random clones from the sorted population of a single experiment 177 ( Supplementary Fig. 2, Supplementary Table 4) and tested the secretion of their encoded 178 fragments when not fused to the anchor protein Sag1. The N-or C-terminal tags of 18/20 (90%) 179 fragments could be reliably detected on western blot from the growth medium, and for 16/20 180 (80%) fragments, both tags were recognized (Fig. 1d, Supp. Table 5). As such, fragments 181 displayed by sorted cells are indeed 'secretable' with a high probability. We further classified 182 fragments into those that were enriched (also referred to as secretable) and those that were  . Table 6, 188 Supplementary Fig. 3). Thus, using this metric, these screens were reproducible with an 189 87.03% concordance between replicates. These final stratified groups of fragments, which 190 were concordantly enriched or depleted, will further be referred to as secretable and depleted, 191

respectively. 192
Since we only performed positive selection for secretable fragments during screening, the 193 depleted fraction contains only passively depleted fragments, and the negative predictive value 194 is relatively low (40-73%, Supplementary Fig. 4). In light of this, interpretation of features 195 affecting secretion must focus on those that affect secretability, and not non-secretability. 196 However, as there are ± 15 times as many fragments (data points) in this depleted set, this 197 relatively low negative predictive value still provides for sufficient signal to allow machine 198 learning methods to learn (see below) . 199 Although we initially tested our method in the model yeast S. cerevisiae, in practice, the 200 methylotroph Pichia pastoris (Komagataella phaffii) is an increasingly popular choice of host 201 for recombinant protein production. Mostly, this has been attributed to this yeast's formidable 202 capacity for high-density growth, the secretion of relatively few endogenous secreted proteins, 203 and the availability of very tightly repressed and extraordinarily strong inducible promoters 204 derived from the yeast methanol metabolism genes 38,39 . Key to the adaptation of SECRiFY for 205 use in P. pastoris was the development of a modified protocol for high-efficiency large-scale 206 P. pastoris transformation, which resulted in an improvement in transformation efficiency of 2 207 -3 orders of magnitude (Online Methods and Supplementary Fig. 5). While we previously 208 observed a slight bias towards detecting small fragments in both enriched and depleted 209 classes in our S. cerevisiae pilot screen, reducing the number of PCR amplification cycles 210 during library generation for sequencing largely eliminated this trend, although small skews 211 occurring during both cloning and sequencing were still observed (Supplementary Fig. 6). For 212 the P. pastoris screens presented here, we first generated a new fragment library with slightly 213 larger fragment insert sizes from the pooled transcriptome of four different human cell lines 214 (SK-N-SH_RA, GM12878, HepG2 and MCF-7) originating from diverse human tissues (brain, 215 blood, liver and breast) and selected to maximize the number of expressed human genes 216 based on ENCODE transcriptome data 40 . Our high-efficiency transformation to P. pastoris 217 generated a library with an estimated diversity of 9.8*10 6 clones. Averaged over three replicate 218 screens, 4.06% ± 0.68% of cells from this library were FLAG + V5 + (Supplementary Fig. 7), 219 which, accounting for the frequent presence of multi-copy inserts (Supplementary Fig. 8), 220 suggests that 12.18% of in-frame fragments are displayed and hence, potentially secretable. 221 Sequencing the fragments of unsorted cells and cells sorted after 1 round of MACS and 1 222 round of FACS, we detected ± 1.5 million unique fragments per replicate, either in the enriched 223 protein-displaying library, in the non-enriched starting library, or in both, covering 38.38% ± 224 2.25% of the human canonical transcriptome with at least three reads (Suppl.  Fig. 9). 228 Overall, these data show that SECRiFY is a reproducible and a reliable method to estimate 229 the secretability of protein fragments. This dataset now represents by far the largest resource 230 on eukaryotic secretability of protein fragments. To enable researchers to easily access and 231 analyze the data from the screens in this study, we built an interactive database 232 (http://iomics.ugent.be/secrify/) allowing visualization of the protein fragments detected in 233 these screens and their mapping on available PDB structures (Supplementary Fig. 10, 234 We first examined whether secretable fragments differed from depleted ones in their probability 244 to form secondary structures. To maximize the accuracy of feature prediction, fragments were 245 filtered for size and exact match to Uniprot proteins, and condensed to an unambiguous subset 246 of consolidated sequences in order to reduce sequence redundancy (see Online methods). 247 Secondary structure prediction of this consolidated subset shows that secretable fragments 248 most prominently have a lower propensity to form a-helical structures (p= 2.95*10 -124 , Mann-249 Whitney-Wilcoxon test) (Fig. 3a, Supplementary Fig. 11a). Indeed, when clustering 250 overlapping sequences to representative fragments and mapping these to solved structures in 251 PDB (roughly 50% of representative fragments, Supplementary Fig. 12), secretability 252 similarly inversely correlates with a-helical content (Fig. 3b, Supplementary Fig. 13a) (p= 253 2.35*10 -8 , Mann-Whitney-Wilcoxon test). In contrast, differences in β-sheet content are only 254 minimal ( Fig. 3a-b). 255 Since secretable fragments also tend to more readily form random coils than depleted 256 fragments, based on secondary structure predictions (p= 1.99*10 -197 , Mann-Whitney-Wilcoxon 257 test) as well as PDB mapping (p= 1.24*10 -4 , Mann-Whitney-Wilcoxon test) (Fig. 3a-b), we 258 further examined how backbone dynamics and intrinsic disorder relate to secretability. As 259 predicted using Dynamine 45,46 , secretable fragments are distinctly more flexible (p= 5.08*10 -260 118 , Mann-Whitney-Wilcoxon test) (Fig. 3c, Supplementary Fig. 14a). Disorder calculations on 261 the full secretable vs depleted sets with RAPID 47 also confirmed a higher average intrinsic 262 disorder content in secretable fragments (p < 2.2*10-16 , Mann-Whitney-Wilcoxon test) (Fig. 3d, 263 Supplementary Fig. 15). In line with this, on average, fragments from both subsets appear 264 compositionally biased. A larger fraction of secretable fragments has a higher proportion of 265 negatively charged residues and prolines, and a tendency towards lower hydrophobicity 266 (Supplementary Fig. 16). Possibly, this increase disorder in secretable fragments reflects how 267 unstructured fragment sequences that lack typically exposed hydrophobic amino acids are 268 missed by ER chaperones and can subsequently travel downstream. This is particularly 269 striking since endogenous secretory system proteins in both human and yeast are, on average, 270 less disordered than the whole proteome, both when considering overall disorder content (p < 271 2.2*10 -16 and p= 2.46*10 -5 resp., Fisher exact test) as well as absolute number of disordered 272 amino acids (p < 2.2*10 -16 and p= 9.44*10 -10 resp., Fisher exact test) (Suppl . Table 11- The gradient boosting classifier requires a fixed input size. Therefore, a series of manually 287 engineered input features were proposed, based on physicochemical properties, sequence 288 length and amino acid frequencies. Ten individual classifiers were trained using different 289 properties, and an ensemble of those was constructed using another gradient boosted 290 classifier taking the outputs of the individual classifiers as input. The deep learning approach 291 involved a CNN taking a one-hot encoding as input, followed by three blocks of convolutional, 292 max pooling and dropout layers. We explored different strategies to deal with the variable input 293 size, as this is not supported by standard CNN architectures. A global max pooling layer 294 yielded the best overall results. This layer is finally connected to a dense layer, followed by an 295 output layer with a softmax. 296 Fragments shorter than 50 amino acids were removed from both the S. cerevisiae and P. 297 pastoris datasets, as those are likely not long enough to properly fold, which mitigates their 298 relevance. Using a restrictive 10-fold cross-validation scheme, where we made sure that 299 protein fragments originating from the same gene were included in the same fold, we compared 300 the classifiers based on the area under the receiver operating characteristic curve 301 (AUROC). Gradient boosting achieved an AUROC of 0.781 and 0.772 on the S. cerevisiae 302 and P. pastoris datasets, respectively, whereas the CNNs achieved AUROCs of the same 303 magnitude, 0.779 and 0.768 (Fig. 4a). Classification results of both classifiers thus confirmed 304 the presence of distinctive features within both secretable and depleted subsets of the data. 305 We observed a strong correlation between the predicted values for the two approaches, with 306 Pearson correlation coefficients of 0.810 and 0.887 on the respective datasets (Fig. 4b), which 307 suggests that the two models learned to use similar distinctive features in the data. 308 Feature importance analysis using attribution methods led to compelling insights in the 309 decisions of the CNN (Fig. 4c-d). Aggregation of individual attribution maps on the amino acid 310 residue level indicated that the influence of individual residues on secretability is largely 311 independent of their position in the sequence. Strikingly, there is a positive bias towards 312 smaller residues, in line with our biophysical predictions that random coils are more readily 313 formed in secretable fragments. Negatively charged residues also seemed to substantially 314 contribute to secretability, confirming the pattern we picked up looking at simple averaged 315 parameters across the full dataset. Similarly, a negative bias was observed towards all 316 hydrophobic amino acids, affirming our earlier observations. 317 318

Secretable fragments are enriched in certain folds and domains 319
Other features that affect folding in the ER might influence secretability. Although increased 320 presence of N-glycosylation sequons or uneven number of cysteines could increase ER 321 retention of fragments, we did not observe clear differences in number of Cys or N-322 glycosylation sequons (Supplementary Fig. 16f-g). Furthermore, secreted and depleted 323 fragments do not substantially differ in their predicted propensity to collapse into folded 324 structure (EFoldMine 50 prediction, Supplementary Fig. 14b). Nonetheless, in select cases 325 where depleted and enriched fragments overlap in sequence on the same protein, the 326 presence or absence of regions that are most likely to fold rapidly often correlated with 327 secretability (Fig. 3e). 328 Beyond these parameters, we noticed clear differences in protein folds and domain 329 architectures represented in both fragment groups. Of those fragments that mapped to known 330 structures, secretable fragments are enriched in distorted sandwich and b complex folds 331 compared to depleted fragments, suggesting that these folds are potentially more stable in the 332 secretory environment, while the opposite is true for proteins with, for example, an a horseshoe 333  Fig. 18). This illustrates that sequence-and fold-contextual patterns of 339 features still contain much information that is not apparent from averaged parameters. 340 341

Protein secretability does not correlate with endogenous secretion 342
Most patterns found to be enriched in secretable fragments are not recapitulated in 343 endogenous secretory proteins. We therefore further evaluated whether the human proteins 344 from which secretable fragments were derived, were enriched in secretory proteins. Since 345 many proteins produced both secretable and depleted fragments in our screens, we 346 considered only those proteins for which no depleted fragments were found. In both S. 347 cerevisiae and P. pastoris screens, the proportion of secretory proteins (i.e., with a signal 348 peptide in the endogenous setting) in this set was not significantly higher than the fraction of 349 secretory proteins in the human proteome (Fisher's one-sided exact test, p=0.9183 (S. 350 cerevisiae) and p=0.421 (P. pastoris), Table 1). This disconnect between highly secretable In SECRiFY, yeast surface display screening of these libraries is combined with high efficiency 372 cell sorting and deep sequencing to segregate and identify protein fragments that can 373 productively pass through the yeast secretory system. As such, we demonstrated that the 374 secretability of protein fragments across entire proteomes can be verified experimentally in an 375 efficient, systematic, high-throughput and reproducible manner. Although we here used the 376 human proteome as a testcase, our approach is generic and can be used to screen any 377 eukaryotic, or with minor adaptations, even prokaryotic proteome. Already, the databases we 378 have generated in this work constitute by far the largest resources of such yeast-secretable 379 human protein segments, which will be useful in structural studies (where high levels of 380 proteins are required for crystallization), immunological experiments (where recombinant 381 production is needed for vaccine development or antigen discovery), and biochemical 382 characterizations of particular proteins (which also necessitates a minimal protein amount). 383 Remarkably, using both a gradient boosted method based on feature engineering, and an end-384 to-end trainable convolutional neural network approach, we achieved an AUROC of up to 385 0.790 for S. cerevisiae and 0.777 for P. pastoris. Practically, this means that secretable and 386 depleted fragments have properties that allow for discrimination, even without prior knowledge 387 of their nature. This unbiased approach confirmed our hypothesis-driven observations that 388 biophysical features relating to secondary structure and flexibility affect secretability. 389 Secretability is thus indeed a learnable feature of protein sequences, a finding that will benefit 390 recombinant protein expression and de novo protein design. 391 From a fundamental biology perspective, it is likely that SECRiFY will provide a means to 392 characterize the substrate scope of secretory system processes that regulate secretory protein 393 passage through the eukaryotic secretory system in a proteome-wide manner. This 394 complements existing methods, such as ribosome profiling 54 , which deal with protein 395 biogenesis prior to passage through the secretory system. 396 Our screens, and the combined hypothesis-driven and unbiased data mining of the data, 397 uncovered that secretable fragments tend to be less a-helical, more flexible and more 398 intrinsically disordered than fragments without significant display. Chaperones responsible for 399 secretory system quality control are generally poised to recognized mostly exposed 400 hydrophobic stretches 55-57 , so conceivably, flexible yet polar and charged fragments would 401 avoid these interactions and quickly progress towards cell wall incorporation. Limited 402 proteolysis measurements also recently confirmed the inverse correlation of in-cell thermal 403 stability with intrinsic disorder, a-helical secondary structure, and aspartate content 58 . Although 404 these measurements involved proteins in their endogenous context in lysed cells, and lack of 405 dataset overlap precluded direct comparison of secretability and thermal denaturation, it would 406 be intriguing to investigate the relationship between secretability and stability. Even though our 407 results show a clear correlation between disorder and secretability, for ordered proteins, quality 408 control mechanisms in the secretory system will generally efficiently remove unfolded or 409 unstable proteins. Indeed, recent limited proteolysis screens of known small protein domains 410 present in PDB or Pfam suggest that structurally defined surface-displayed domains have an 411 overall high stability score 59 . 412 Fragment detectability effects may also contribute to the observed enrichment of aspartates 413 and glutamates. Phosphomannose proteins of the yeast cell wall, which confer it a net negative 414 charge, may be repelled by negatively charged displayed fragments, leading to more efficient 415 antibody staining due to increased accessibility. Our observation that more fragments with a 416 high overall positive charge (Lys + Arg content) are not secreted could relate to translational 417 ribosomal stalling at polybasic stretches 60 . Although proline-rich proteins are generally 418 localized in the nucleus or cytoplasm 61 , and are associated with ribosome slowdown 62 , it is 419 possible that the intrinsic ability of prolines to 'lock' conformations or reduce conformational 420 freedom enhances secretory protein stability and hence, display and secretion. Polyproline or 421 pro-rich stretches are also known motifs for binding to a wide variety of other proteins 29,63,64 ; 422 as such, the human oxidoreductase ERp57 binds calnexin and calreticulin at their pro-rich 423 motif, and peptidyl-prolyl isomerases often act in complexes associated with multiple 424 chaperones. Prolines are also used as gatekeeper residues against aggregation 65 , and 425 perhaps by extension, also against degradation and therefore display. 426 Secretable fragments are not enriched in secretory proteins. SECRiFY detects secretion at the 427 fragment level, potentially causing some features affecting full-length protein secretion to be 428 missed. Nonetheless, the absence of correlation also underlines that SECRiFY assesses 429 secretability, i.e. the capacity to be secreted, rather than actual endogenous cellular 430 localization to the secretory system. Indeed, just like most proteins are only marginally stable, 431 endogenous secretory proteins evolved for function, not secretability. 432 Arguably, our method also has its limitations. In the current SECRiFY setup, secretability was 433 measured in the sequence context of the a mating factor prepro sequence at the N-terminus, 434 and the Sag1 cell wall protein at the C-terminus. While results from our and other labs have 435 indicated that for several single proteins, display efficiency correlates with relative secretion 436 levels, it cannot be excluded that, at least for certain fragments, both leader sequence and the 437 +/-300 amino acid Sag1 anchor might differentially influence fragment folding, solubility, or 438 stability. In E. coli, fusion to large proteins such as SUMO, the T. harzanium cellulose binding 439 domain (CBD), or to maltose binding protein (MBP) is an often used strategy to promote 440 'passenger solubilization', although again, effects vary depending on the protein 66,67 . 441 Considering the vectorial nature of translation, a C-terminal fusion, as is the case in our setup, 442 is nevertheless generally deemed less perturbing than an N-terminal fusion, although this is 443 not absolute. Sag1 is also a GPI-anchored protein, affecting the entry pathway into the ER 68-444 70 . Similarly, the prepro leader sequence, with its multi-step processing and preference for 445 14 posttranslational translocation 71-73 , may bias secretability of certain fragments. It remains to be 446 determined whether similar patterns will emerge with different secretory leaders, anchors, 447 promoters, untranslated regions, or growth conditions. 448 Display also imposes limitations on the dynamic range of the method, as there is an upper limit 449 to the number of molecules that the cell wall can accommodate. Generally, this is in the range 450 of about 10 4 molecules per cell 74,75 . Thus, perturbations affecting secretion efficiency in these 451 higher ranges may be missed. 452 In all, with SECRiFY, we here show that our fragment sequence library is proof-of-concept for 453 massively parallel assessment of passage through the secretory system, providing the 454 opportunity the learn which features influence secretability, and what rules sequences must 455 abide by for successful transit through the yeast secretory system. We anticipate that our 456 method and its next-generation derivatives will be of great value in both protein engineering 457 and fundamental studies of the secretory system.   Enriched fragments have a lower helical content (p= 2.95*10 -124 ) and a higher random coil (p= 1.99*10 -127 ) propensity, which is confirmed further by mapping representative fragments to known structures in PDB (a-helix p= 2.35*10 -8 , random coil p= 1.24*10 -4 ) (b). Beta sheet differences are not as pronounced (p= 1.26*10 -3 for Dynamine prediction and p= 0.02 for PBD mapping). (c) Enriched consolidated fragments are also predicted to be more dynamic than depleted ones (p= 5.08*10 -118 ). (d) Similarly, the predicted disorder content in the total set of enriched fragments is significantly higher than in depleted fragments (p= <2.2*10 -16 ). Twosided Mann-Whitney-Wilcoxon tests in (a-d). * p < 0.05, ** p <0.01, **** p<0.0001. (e). Two overlapping fragments of the human protein EDIL3 differ in secretability outcome. Early folding (EF) propensity predictions suggest that for the depleted fragment regions E2, T3/E3 and R4 are likely the regions driving folding of the depleted fragment, and lack of these regions in the enriched fragment result in a change in secretability.