Merits and challenges of plasma proteomics on association replicability

23 Developments in proteomic platforms have enabled the generation of large-scale high-24 throughput plasma proteomics data [1-3]. With recent breakthroughs in AI modelling, 25 these data have significantly enhanced our understanding of molecular mechanisms 26 underlying human behaviors and diseases [4-6]. However, the replicability of 27 associations between plasma proteomics and phenotypes remains underexplored. Here, 28 we systematically assessed t he replicability of associations with recent plasma 29 proteomics data in the UK biobank. Over 75% of cognitive function and mental health 30 traits demonstrated high overall (proteomics-wide) replicability when brain -related 31 traits were considered as phenotypes. Although mean cortical thickness (CT) as 32 phenotype exhibited clearly reduced replicability, total cortical surface area (CSA) and 33 cortical volume (CV) showed high overall replicability across hemispheres and over 34 twenty brain regions. In comparative multi-omics analyses based on the same cohort of 35 participants, proteomics outperformed genomics across all brain-related traits, and 36 exceeded metabolomics for over half of traits where metabolomics also exhibited high 37 overall replicability. Furthermore, we developed a predictive framework to estimate the 38 replicability for potential future proteomics panels based on the crucial influential 39 factors including dilution level, proportion of samples below the limit of detection 40 (LOD), and sample size. Moreover, we constructed an individual replicability index for 41 proteins and identified eleven proteins with highly replicable associations across 42 cognitive function and mental health traits , which was in line with the recent 43 identifications of pleiotropic proteins in large-scale population studies. Collectively, our 44

revealed the challenges in the association replicability of plasma proteomics 45 under reduced data quality (from “Explore” to “Ex pansion” assay panels), and we 46 further explored how to sustain high replicability in potential future panels. 47 Fundamentally, our findings affirm the merits of plasma proteomics: this molecular 48 omics platform enables highly rep licable associations for mapping biomedical 49 signatures. 50 51 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint

52 The development of high -throughput proteomics platforms and recent 53 breakthroughs in artificial intelligence (AI) have enhanced our understanding of the 54 molecular mechanism s underlying various human behaviors and diseases [1 -6]. In 55 recent years, association analyses remain fundamental in large-scale high-throughput 56 studies, yet replicability challenges arise in these studies [7, 8 ]. For example, the 57 replicability issues of genome-wide association studies (GWAS) have garnered a 58 considerable attention [9]. Furthermore, the replicability of brain-wide association 59 studies (BWAS) was also assessed based on large-scale neuroimaging data in recent 60 studies [8, 10]. Given the continuous expan sion for proteomics-based association 61 studies (PBAS), there is still a lack of systematic replicability investigations. Moreover, 62 the prospect of diagnosing complex conditions through accessible blood tests represents 63 a promising change in health care, which offers an early-stage alternative to costly and 64 invasive procedures [11]. Plasma proteomics has been demonstrated as a cutting-edge 65 approach for the development of blood test s [ 12], although it has not yet been 66 systematically compared to other blood-based platforms (e.g., genomics and 67 metabolomics) in terms of association replicability. 68 Several criteria have been developed to quantify the replicability of associations, 69 generally based on statistical significance [13, 14], rank correlation [15], or directional 70 consistency (DC) [16, 17], etc. Among these, the DC criterion is specifically developed 71 based on the consideration of both high -throughput settings and consistencies of 72 association directions (i.e., overall replicability) , thereby useful in the replicability 73 assessment for association analysis findings [16-19]. Here, we adopted the DC criterion 74 for our systematic assessment of replicability in PBAS. Notably, t he overall 75 replicability (based on directional consistency) may be impacted by influential factors, 76 such as the fraction of missing data, dilution level, proportion of samples below the 77 limit of detection (LOD) and sample size [1, 7, 8, 20]. However, the impact of these 78 factors on the overall replicability of PBAS has not yet been systematically 79 characterized. Moreover, the recent plasma proteomics has been rapidly developed in 80 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint both sample sizes and throughput volumes [21]. A comprehensive understanding of 81 these influential factors is essential for estimating the association replicability of 82 potential future proteomic panels. Here, we intend to address these issues in this study. 83 Recently, the identifications of pleiotropic proteins have been gaining a significant 84 attention to understand systems biology and complex diseases [22]. These proteins can 85 substantially lower the high expenses of broad proteomic screening and facilitate their 86 translation into potential clinical targets [6]. Notably, an individual-level assessment of 87 association replicability for each protein provides an essential contribution to the 88 development of pleiotropic proteins [15, 16]. Therefore, in this study, we also aim to 89 develop a procedure for this purpose. 90 To our knowledge, we were among the first to conduct a comprehensive 91 investigation of the DC-based replicability for associations in plasma proteomic s. 92 Based on this investigation at the individual level for proteins, our study provided an 93 alternative perspective to the discovery of pleiotropic proteins, which could facilitate 94 the development of practically useful biomarkers. Based on the investigation of overall 95 replicability, we enhanced our confidence in the current development and applications 96 of high-throughput plasma proteomics, and highlighted the challenges along with the 97 growth of throughput volume and scale . As the development of future proteomic 98 platforms would be impacted by certain crucial factors, our study provide d a related 99 insight, and also addressed the prediction of association replicability for potential future 100 plasma proteomic platforms. Fundamentally, our study contributed to a systematic 101 understanding of the merits of plasma proteomics on association replicability. (Details 102 of study overview please see Figure 1.) 103 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint 104 Figure 1 | Overview of the study design and analyses. Part 1. Large-scale association 105 studies utilizing UK Biobank plasma proteomics, phenotypes, genomics, and plasma 106 metabolomics. The full list of brain -related traits within physical measures, cognitive 107 function, mental health and brain region names for imaging metrics could be found in 108 Supplementary Table 1. Part 2. Overall (proteomics-wide) replicability assessment for 109 PBAS, sample size calculation and multi-omics comparison. Part 3. Characterization 110 of influential factors on overall replicability and development of a predictive framework 111 for potential future panels. Part 4. Individual (protein-level) replicability assessment 112 with pleiotropic proteins identification and influential factors characterization. 113 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint

114 115 Overall replicability for PBAS 116 The simulation results, summarized in Supplementary Materials and Figure S1, 117 indicated that the MMRA approach could accurately assess the replicability of PBAS 118 results. 119 As shown in Figure 2a and 2b, for all physical measures including body mass index 120 (BMI), height, weight, diastolic blood pressure (DBP), systolic blood pressure (SBP), 121 speech-reception-threshold (SRT) estimate and pulse wave arterial stiffness index (ASI), 122 a median overall irreplicability quantity 𝜌𝐼𝑅< 0.05 was consistently observed for 123 each trait. For all 3 1 traits within cognitive function and mental health, the median 124 overall irreplicability quantity 𝜌𝐼𝑅< 0.05 was observed for 24 traits (77.4%) 125 indicating high levels of overall replicability. For cognitive function, the median 𝜌𝐼𝑅≥ 126 0.05 was observed in two traits . The median 𝜌𝐼𝑅 was 0.429 for the symbol digit 127 substitution test and 0.539 for the paired associate learning test, with corresponding 128 lower- and upper-quartiles (Q1-Q3) of 0.355-0.524 and 0.018-0.703, respectively. For 129 the 24 traits under mental health, only 5 traits exhibited low overall replicability levels, 130 with a median 𝜌𝐼𝑅≥ 0.05. In these five traits, such as the work satisfaction and family 131 relationship satisfaction, the values of median 𝜌𝐼𝑅 were relatively higher (The median 132 𝜌𝐼𝑅 was 0.505 for the work satisfaction and 0.568 for family relationship satisfaction). 133 Please see Supplementary Table 2 for t he details of assessed median 𝜌𝐼𝑅 with 134 corresponding lower- and upper-quartiles (Q1-Q3) for traits within physical measures, 135 cognitive function and mental health . We further evaluated the influence of blood 136 collection season and fasting time on PBAS replicability. Overall, these additional 137 covariates had extremely limited influence on replicability, as shown in Figures S2 and 138 S3. 139 For brain imaging metrics, a s shown in Figure 2c, the median 𝜌𝐼𝑅 values were 140 0.2983 and 0.3492 with the Q1-Q3: 0.0529-0.5326 and 0.0776-0.5346 for mean CT in 141 left and right hemispheres, respectively. However, total CSA and CV showed clearly 142 higher overall replicability than mean CT. The median overall irreplicability quantity 143 𝜌𝐼𝑅 values for total CSA in left and right hemispheres reached 0.0017 and 0.0014 while 144 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint the related Q1-Q3 was 0.0001-0.0743 and 0.0001-0.1071, respectively. For total CV in 145 left and right hemispheres, the median 𝜌𝐼𝑅 values were 0.0017 and 0.0036 while the 146 related Q1 -Q3 was 0.000 2-0.1298 and 0.0002-0.1323, respectively. These results 147 indicated a high level of overall replicability for brain imaging metrics. Please see 148 Supplementary Table 3 for t he details of assessed median 𝜋𝑅, 𝜋𝐼𝑅 or 𝜌𝐼𝑅 for total 149 CSA, total CV and mean CT in left and right hemispheres. 150 No regions were identified for region-wide mean CT with median 𝜌𝐼𝑅< 0.05 151 (Figure 2d). For region-wide CSA versus plasma proteins, when using the threshold of 152 median 𝜌𝐼𝑅< 0.05, twenty-one brain regions were identified in our results (Figure 2e). 153 The top three regions were the right pars orbitalis gyrus ( median 𝜌𝐼𝑅= 0.0024), 154 followed by the left lateral occipital gyrus (median 𝜌𝐼𝑅= 0.0106) and the right rostral 155 middle frontal gyrus ( median 𝜌𝐼𝑅= 0.0113). For C V, twenty-three regions were 156 identified with median 𝜌𝐼𝑅< 0.05 (Figure 2f). T he top three regions were the left 157 medial orbitofrontal gyrus (median 𝜌𝐼𝑅= 0.0037), followed by the right insula gyrus 158 (median 𝜌𝐼𝑅= 0.0050) and the left superior temporal gyrus (median 𝜌𝐼𝑅= 0.0069). 159 (Please see Supplementary Table 4 for the details of median 𝜌𝐼𝑅 for region-wide CSA, 160 CV and CT.) The potential explanations for the observed differences in overall 161 replicability levels between total CSA/CV and mean CT were given in supplementary 162

and Figure S4. 163 164 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint 165 Figure 2 | Overall replicability assessment for brain-related traits. a, Median 166 overall irreplicability quantity (𝜌𝐼𝑅) with lower- and upper -quartiles (Q1-Q3) from 167 1,000 random subsampling times for traits in physical measures, cognitive function, 168 and mental health, alongside corresponding sample sizes. b, Number of traits within 169 physical measures, cognitive function, and mental health with median 𝜌𝐼𝑅< 0.05 170 versus 𝜌𝐼𝑅≥ 0.05. c, Median 𝜌𝐼𝑅 for mean cortical thickness (CT), total cortical 171 surface area (CSA), and cortical volume (CV) in both hemispheres. d–f, Regional 172 median −log10(𝜌𝐼𝑅) for (d) mean CT, (e) total CSV, and (f) total CV. 173 174 log ( ) log ( ) log ( ) < preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint Multi-omics overall replicability comparison 175 As shown in Figure 3a and 3b, compared to pQTLs, association analysis results 176 based on proteomics data exhibited notably lower 𝜌𝐼𝑅 values, indicating higher overall 177 replicability. This comparison was based on the same cohort of participants. For all 31 178 traits within the categories of cognitive function and mental health, we could also 179 observe the median overall irreplicability quantity 𝜌𝐼𝑅< 0.05 on 23 traits (74.2%), 180 indicating high levels of overall replicability for proteomics data (Figure 3b). For 181 genetics-based association analyses, a high replicability level (median 𝜌𝐼𝑅= 0.00082) 182 was observed only for height. For brain-related measures, relatively low overall 183 replicability levels (median 𝜌𝐼𝑅≥ 0.05) were observed in genetic-based analyses. 184 Based on the same cohort of participants , the overall replicability comparison of 185 metabolomics-based association studies versus PBAS was presented in Figure 3 c and 186 3d. Here, we observed the median overall irreplicability quantity 𝜌𝐼𝑅< 0.05 in 7 out 187 of 8 physical measures. Moreover, among the 31 traits related to cognitive function and 188 mental health, 20 traits (64.5%) exhibited the median 𝜌𝐼𝑅< 0.05. For brain structure 189 imaging metrics, the median 𝜌𝐼𝑅< 0.05 was observed for total CSA in both 190 hemispheres and total CV in the left hemisphere. While these results indicated that 191 metabolomics-based association studies also demonstrate high levels of overall 192 replicability, PBAS exhibited even higher overall replicability for all 8 physical 193 measures and over half of brain-related phenotypes (i.e., 17 out of 31 traits related to 194 cognitive function and mental health shown in Supplementary Table 5). Further details 195 of overall replicability comparison s of genetics versus proteomics and metabolomics 196 versus proteomics were also displayed in Supplementary Table 5. 197 198 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint 199 < preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint Figure 3 | Comparative overall replicability across multi -omics. Median overall 200 irreplicability quantity (𝜌𝐼𝑅) and the lower- and upper-quartiles (Q1-Q3) from 1,000 201 random subsampling times are shown. a, b , Comparison between genetics -based 202 associations from pQTLs and PBAS. c, d, Comparison between metabolomics -based 203 associations and PBAS. Traits are grouped into physical measures, cognitive function, 204 mental health, and brain imaging metrics. 205 206 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint Influential factors for overall replicability 207 The proteomics data quality difference between Panels 1 (“Explore 1,536 assay 208 panels”) and 2 (“Expansion 1460 assay panels” ) may lead to different overall 209 replicability for PBAS. Panel 2 showed a clearly higher missing rate: t he median 210 missing rate was 0.169 while Q1-Q3 was 0.152-0.175. Panel 1 displayed impressively 211 low missing rate: t he median missing rate was 0.0274 while the related Q1-Q3 was 212 0.0231-0.0423. Moreover, the category with least -abundant (1:1) dilution section and 213 more than 50% of samples below LOD in Panel 2 contained clearly more proteins than 214 that in Panel 1 (see S upplementary Table 6 for details) . When phenotypes fluid 215 intelligence (Figure 4 a) and neuroticism (Figure 4 b) were considered , we observed 216 clearly lower overall replicability for Panel 2. 217 We then investigated the impact of proteomics data quality for overall 218 replicability and the results were also demonstrated in Figure 4 . Consider ing fluid 219 intelligence and neuroticism as examples, we observed that the least -abundant (1:1) 220 dilution section exhibited relatively higher overall irreplicability quantity 𝜌𝐼𝑅 values 221 compared to the moderate -abundant (1:10) and more -abundant (1:100 to 1:100000) 222 dilution section. Across four categories stratified by the proportion of samples below 223 the LOD (50%), overall replicability declined as the 224 proportion of samples below the LOD increased. A downward trend was observed for 225 overall irreplicability quantity 𝜌𝐼𝑅 with increasing sample size. 226 For additional traits, a clear decline in overall replicability was observed in Panel 227 2, lower dilution levels, higher proportions of samples below the LOD and smaller 228 sample size (Figure S5-S8). Based on Panel 2, only 51.6% of traits related to cognitive 229 function and mental health showed high overall replicability (median 𝜌𝐼𝑅< 0.05), 230 compared to 80.6% based on Panel 1 (Figure 4c). Based on the least-abundant dilution 231 section (1:1), 64.5% of cognitive and mental health-related traits exhibited high overall 232 replicability, compared to 80.6% and 77.4% of traits based on the moderate-abundant 233 (1:10) and more-abundant (1:100 to 1:100,000) dilution sections, respectively (Figure 234 4d). Based on the categories of proteins with more than 50% of samples below the LOD, 235 only 3.2% of traits demonstrated high overall replicability ; this percentage was 236 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint substantially lower than those based on the other three LOD -based categories (Figure 237 4e). When the sample size was reduced to approximately 50% and 25% of the original 238 dataset, 71% and 41.9% of traits respectively exhibited high overall replicability , 239 compared to 77.4% based on the original dataset (Figure 4f) . Detailed results were 240 provided in Supplementary Table 7. 241 242 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint 243 Figure 4 | Characterization of influential factors for overall replicability. a, b, The 244 overall replicability (𝜌𝐼𝑅) for (a) fluid intelligence and (b) neuroticism with different 245 panels, dilution levels, the proportion of samples below LOD and study sample size . 246 a. Fluid intelligence Panels Dilution levels Proportion of samples below LOD Sample size . Neuroticism Panels Dilution levels Proportion of samples below LOD Sample size log ( ) log ( ) c. . e. . preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint Each box plot represents the – log10 (𝜌𝐼𝑅) from 1,000 random subsampling times. The 247 center line in each box represents the median; the lower and upper hinges represent the 248 25th and 75th percentiles, respectively; the whiskers represent 1.5 × the lower- and 249 upper-quartiles. c-f, Percentage of traits with median 𝜌𝐼𝑅< 0.05, summarized 250 separately for (c) panels, (d) dilution levels, (e) the proportion of sample below the 251 LOD, and (f) sample sizes. 252 253 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint Replica ility or potential uture panels 254 When the fluid intelligence and neuroticism were considered as phenotypes in our 255 logistic regression model, larger sample size and lower proportion of samples below 256 the LOD showed positive contributions while negative for lower abundant of dilution 257 level (Figure S9); and AUC-ROC of 0.62 and 0.58 were achieved, respectively (Figure 258 S10). According to the difference between Panel 2 versus Panel 1, we would consider 259 a potential future panel with lower dilution level and higher proportion of samples 260 below the LOD. For the hypothetical examples presented in Figure S10, increasing 261 sample size would generally improve the predicted probability of overall replicability, 262 though the improvements may be modest for certain phenotypes such as neuroticism. 263 264 265 266 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint Pleotropic proteins identification 267 The top ten proteins with the highest individual replicability were identified as 268 being highly replicable in their associations with each trait. The full list of top ten 269 proteins for each PBAS scenario could be found in Supplementary Table 8. GDF15 270 was identified as a pleotropic protein (See Methods section for details) in both mental 271 health (in 14 traits) and cognitive function (in 3 traits). The proteins ASGR1, PIGR, 272 PLAUR, and PRSS8 were identified as pleotropic proteins in mental health, appearing 273 in 10 to 14 traits ; and the proteins APOF, CCL20, CDCP1, GGH, MZB1 and TFF1 274 were identified as pleotropic proteins in cognitive function (Figure 5a). 275 As shown in Figure 5b, an FDC > 0.99 (fraction of directional consistency, see 276

section for details) was observed in all 1,000 random subsampling for at least 277 one trait in the cognitive function and mental health for each pleotropic protein. Many 278 traits exhibited FDC > 0.95 for all these pleotropic proteins (such as reaction time, fluid 279 intelligence in the cognitive function, as well as mood swings, fed -up feelings, 280 loneliness, depressed mood, unenthusiasm, tenseness, tiredness, health satisfaction and 281 financial situation satisfaction in the mental health). For certain traits, such as worry 282 and worry after embarrassment, many pleiotropic proteins showed FDC < 0.05. 283 Nevertheless, the protein MZB1 demonstrated high FDC values of 0.99 and 0.90 for 284 worry and worry after embarrassment, respectively. In summary, for every trait 285 considered, there was at least one pleotropic protein exhibited an FDC > 0.9. 286 To support the significance of eleven pleotropic proteins for cognitive function and 287 mental health, we randomly selected eleven proteins from the remaining 2,909 proteins 288 and calculated the FDCs. For each set of randomly selected eleven proteins, we 289 calculated the proportion of entries (total 11 × 24 proteins∙traits entries) with FDC < 290 0.05, indicating relatively high individual irreplicability. This selected procedure was 291 randomly repeated 1,000 times. Similar results were also calculated for the proportion 292 of entries with FDC > 0.95. The corresponding distributions were shown in Figure 5c. 293 Compared to the random selections, the set of eleven pleotropic proteins exhibited 294 significantly fewer proportion of entries with FDC 0.95. 296 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint 297 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint Figure 5 | Individual-level replicability assessment a, Pleiotropic proteins and their 298 observed frequency of being ranked among the top 10 across traits in cognitive function 299 and mental health. b, FDC for each pleotropic protein across traits. Blank cells indicated 300 FDC > 0.99 . c, Proportion of entries with extreme FDC values (upper: FDC 0.95) for pleiotropic proteins versus randomly selected proteins. Dashed 302 lines indicated the observed proportions of pleiotropic proteins. d, Median values of 303 median individual irreplicability quantity (𝛾𝑘) between Panel 1 and Panel 2 for each 304 trait within cognitive function and mental health. 305 306 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint Influential factors for individual replicability 307 The individual replicability level for P anel 2 was clearly lower than P anel 1 308 (Figure 5d). For example, for mood-related phenotypes such as depressed mood and 309 unenthusiasm, we observed low median 𝛾𝑘 values of 0.001 and 5.24× 10-4 for all the 310 proteins in Panel 1; however, these values were 0.443 and 0.596 for Panel 2. Please see 311 Supplementary Table 9 for the details of median 𝛾𝑘 for Panels 1 and 2. In addition, the 312 individual irreplicability quantity 𝛾𝑘 of PBAS was also assessed for sub -panels in 313 Panels 1 and 2, including Cardiometabolic, Inflammation, Neurology and Oncology. 314 The assessment details can be found in the Supplementary Materials. For these four 315 sub-panels, higher levels of individual replicability and a lower data missing rate were 316 also observed in Panel 1 (Figure S11 and Supplementary Table 10). 317 When analyzing the impact of different dilution levels for individual replicability 318 performance, we observed that the least -abundant dilution section (1:1) exhibited 319 relatively higher median individual irreplicability quantity 𝛾𝑘 values for a ll traits 320 compared to the moderate -abundant (1:10) and more-abundant (1:100 to 1:100000 ) 321 dilution section (Figure S 12). Detailed individual replicability assessments for each 322 dilution section were provided in Supplementary Table 11. Additionally, we 323 investigated the impact of the proportion of samples below the LOD on individual 324 replicability. For each trait related to cognitive function and mental health, we found a 325 significant negative relationship between individual replicability (median 326 − log10 (𝛾𝑘)) and the proportion of samples with measurements below the LOD 327 (Spearman’s 𝜌 < −0.1399, P <  −8; Figure S13). 328 329 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint

330 With the development of high-throughput proteomics platforms and AI modeling, 331 the replicability of associations serves as a foundation for both biological discover ies 332 and clinical translations [1-6]. While the advantages of plasma proteomics have been 333 widely acknowledged [1 -3], its association replicability remains a critical yet 334 underexplored area. In order to further understand the merits and challenges in plasma 335 proteomics, we provided a comprehensive investigation of the DC-based replicability 336 for associations. Our work highlighted three key insights: (1) We demonstrated the high 337 overall association replicability of plasma proteomics, which underlined its advantages 338 versus genomics and metabolomics platforms; (2) We assessed crucial influential 339 factors for association replicability, and developed a predictive framework according to 340 potential future challenges along with the growing throughput of proteomics; and (3) 341 Based on an individual-level replicability index and its related evaluation procedure, 342 we identified eleven replicable pleiotropic proteins for cognitive function and mental 343 health. 344 A key strength of our study lies in its comprehensive evaluations, which provide 345 the depth and breadth in assessing the association replicability of proteomics. Across a 346 diverse range of phenotypes, including physical measures and brain-related traits 347 (cognitive function and mental health traits) , we demonstrated the broad utility of 348 plasma proteomics. For brain imaging measures, high overall replicability could also 349 be observed for total CSA and CV in both hemispheres and over twenty brain regions. 350 For mean CT in the proteomics -based association study, the proportion of no signals 351 (i.e., true null hypotheses) was relatively higher than that for total CSA and CV (Figure 352 S4). As the DC-based replicability assessment was influenced by the proportion of 353 positive/negative signals and no signals, this could explain the lower overall 354 replicability observed for mean CT. Likewise, a recent genomic study highlighted a 355 comparatively larger genetic architecture for total CSA than that for mean CT [23]. 356 Moreover, multi -omics data have emerged as a promising foundation for the 357 development of blood tests which could enhance disease screening rates and facilitate 358 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint early diagnosis [11, 24]. For example, molecular phenotyping based on genomic data 359 facilitates early prediction and more accurate characterization of disease progression 360 [25]. Metabolomics has emerged as a powerful approach for the identification of pre -361 disease states [26]. Furthermore, the prediction modeling based on plasma proteomics 362 enables reliable estimation of 7-year to 10-year incidence risk for various common and 363 rare diseases [27, 28]. By demonstrating the advantages of proteomics versus genomics 364 and metabolomics in association replicability within the same cohort of pa rticipants, 365 our study strengthened the confidence in both ongoing large-scale proteomic 366 association studies and the desirable development of reliable blood-based diagnostic 367 tools. Notably, to our best of knowledge, a large sample size and multi-phenotype 368 dataset that contain ed matched plasma proteomics and transcriptomic s was not 369 available currently. The comparison between plasma proteomic and transcriptomics 370 based on the same cohort of participants was not conducted. 371 While our findings affirm ed the merits of plasma proteomics, we also face d 372 challenges in replicability with the growth of throughput volume. To address this issue, 373 we assessed the factors influencing the replicability of association findings (including 374 missing rate, LOD, dilution level and sample size ), and also developed a predictive 375 framework for estimating the replicability of potential future assay panels. A crucial 376 observation from our work is the concerning trend of declining data quality (proteins 377 with higher missing rate, lower abundant dilution level and higher proportion of 378 samples below LOD) in 'Expansion 1460 assay panels' versus 'Explore 1,536 assay 379 panels'. In high-throughput association studies, higher rates of missing data were shown 380 to reduce the replicability of findings [7]. Furthermore, a recent study demonstrated the 381 decrease of number of identified pQTLs versus the increase of dilution level or 382 proportion of samples below the LOD [1]. Another study also shown the influence of 383 these two factors on the correlations between the proteomics data generated from two 384 different platforms [20]. In line with these previous studies, our results revealed the 385 impact of these three factors (i.e., missing rate, LOD and dilution level ) on the 386 association replicability, and showed that the Expansion assay panel s exhibited a 387 relatively lower replicability than the Explore assay panels. Moreover, in our results, 388 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint sample size was also identified as a crucial influential factor for replicability. Based on 389 our predictive framework, we further demonstrated that increasing sample size can be 390 a practically feasible way to sustain replicability for potential future proteomics panels. 391 Adequate sample sizes were essential to ensure replicable results in high -392 throughput omics association studies [8, 12]. For instance, the results of GWAS with 393 relatively small sample sizes generally failed to be replicated [ 29], while millions of 394 samples were often required [ 30]. Similarly, growing concern s were raised o n the 395 sample size requirement for brain-wide association studies (BWAS), where thousands 396 of samples are suggested to achieve satisfactory replicabilities [8]. In our recent study, 397 a desirable overall replicability was achieved for physical measures when the sample 398 size reached several hundred to a few thousand [17]. In this study , for plasma 399 proteomics data, our results demonstrated that thousands of samples were sufficient to 400 achieve a high overall association replicability for physical measures. For cognitive 401 function, mental health, and brain imaging measures, a high overall association 402 replicability was also achieved in most traits (21 out of 37 with sample size <10,000) 403 when the sample size reached several thousands (Supplementary Table 12). 404 Our investigation included an individual level replicability assessment, which 405 could also identify proteins with highly replicable associations across a scope of 406 phenotypes. Eleven pleiotropic proteins were identified for cognitive function and 407 mental health , including GDF15, ASGR1, PIGR, PLAUR, PRSS8, APOF, CCL20, 408 CDCP1, GGH, MZB1 and TFF1. Among them, plasma GDF15 [31-35], PLAUR [36-409 38], CDCP1 [22, 39] and TFF1 [22] were identified as potential response biomarkers 410 associated with cognitive function and mental health, and these proteins were further 411 highlighted on the importance of pleiotropy in complex traits [22]. Moreover, ASGR1, 412 PIGR, PRSS8 [36] and APOF [40] were identified as important protein biomarkers for 413 mental diseases including depression, neurodegeneration and schizophrenia. 414 Furthermore, CCL20 [41], GGH [42] and MZB1 [43] were associated with Alzheimer’s 415 disease and reduced cognitive functions during the process of ageing. As pointed out 416 by Topol [6], identifying a relatively small set of proteins with replicable associations 417 could facilitate the development of targeted, low-cost assay panels for the proteomics-418 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint driven clinical translations. 419 Our study had several limitations. First, our analysis was primarily focused on a 420 range of brain-related phenotypes. Although these traits are diverse and complex, the 421 replicability levels and influential factors may not be directly generalized to other 422 disease domains. Further investigations are warranted to a broader spectrum of human 423 diseases. Second, our investigation was conducted based on the UK Biobank data. 424 Consequently, our findings may not be representative to other populations. To address 425 this limitation and the generalization of proteomic discoveries, it is necessary to conduct 426 further investigations for diverse cohorts. Third, our results were based on the Olink 427 high-throughput proteomics platform. Nevertheless, similar results may still be 428 observed from proteomics data generated by other technologies. 429 Despite these limitations, our study provides evidence s for the advantages of 430 plasma proteomics in large -scale association studies. In summary, this study was 431 among the first to provide a comprehensive DC-based assessment of the association 432 replicability. Our study included the assessment at the overall and individual levels for 433 association replicability. Our work further revealed the challenges for the future 434 developments, and our analyses on influential factors and association replicability 435 prediction provided a valuable contribution. Fundamentally, our findings affirmed that 436 plasma proteomics was replicable in association analyses. 437 438 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint

439 [1] Sun B B , Chiou J, Traylor M, et al. Plasma proteomic associations with genetics 440 and health in the UK Biobank. Nature, 2023, 622(7982): 329-338. 441 [2] Eldjarn G H, Ferkingstad E, Lund S H, et al. Large -scale plasma proteomics 442 comparisons through genetics and disease associations. Nature, 2023, 622(7982): 348-443 358. 444 [3] Dhindsa R S, Burren O S, Sun B B, et al. Rare variant associations with plasma 445 protein levels in the UK Biobank. Nature, 2023, 622(7982): 339-347. 446 [4] Carrasco-Zanini J, Pietzner M, Davitte J, et al. Proteomic signatures improve risk 447 prediction for common and rare diseases. Nature Medicine, 2024: 1-10. 448 [5] Carrasco-Zanini J, Pietzner M, Koprulu M, et al. Proteomic prediction of diverse 449 incident diseases: a machine learning -guided biomarker discovery study using data 450 from a prospective cohort study. The Lancet Digital Health, 2024, 6(7): e470-e479. 451 [6] Topol E J. The revolution in high -throughput proteomics and AI. Science, 2024, 452 385(6716): eads5749. 453 [7] National Academies of Sciences, Engineering, and Medicine. Reproducibility and 454 replicability in science. National Academies Press, 2019. 455 [8] Marek S, Tervo -Clemmens B, Calabro F J, et al. Reproducible brain -wide 456 association studies require thousands of individuals. Nature, 2022, 603(7902): 654-660. 457 [9] Lin X. Learning lessons on reproducibility and replicability in large scale genome-458 wide association studies. Harvard Data Science Review, 2020, 2(4): 10.1162/99608f92. 459 33703976. 460 [10] Liu S, Abdellaoui A, Verweij K J H, et al. Replicable brain–phenotype associations 461 require large-scale neuroimaging data. Nature Human Behaviour, 2023, 7(8): 1344 -462 1356. 463 [11] Chabon J J, Hamilton E G, Kurtz D M, et al. Integrating genomic features for non-464 invasive early lung cancer detection. Nature, 2020, 580(7802): 245-251. 465 [12] Liotta L A, Ferrari M, Petricoin E. Clinical proteomics: written in blood. Nature, 466 2003, 425(6961): 905-905. 467 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint [13] Goodman SN . A comment on replication, p -values and evidence. Statistics in 468 Medicine, 1992, 11: 875-879. 469 [14] Heller R, Bogomolov M, Benjamini Y. Deciding whether follow-up studies have 470 replicated findings in a preliminary large -scale omics study. Proceedings of the 471 National Academy of Sciences, 2014, 111: 16262-16267. 472 [15] Li Q, Brown JB, Huang H, Bickel PJ , et al. Measuring reproducibility of high -473 throughput experiments. The Annals of Applied Statistics, 2011, 5: 1752-1779. 474 [16] Zhao Y, Sampson M G, Wen X. Quantify and control reproducibility in high -475 throughput experiments. Nature Methods, 2020, 17(12): 1207-1213. 476 [17] Jiao Z, Lai Y, Kang J, et al. A model-based approach to assess reproducibility for 477 large-scale high-throughput MRI-based studies. NeuroImage, 2022, 255: 119166. 478 [18] Li Y, Lei H, Wen X, et al. A powerful approach to identify replicable variants in 479 genome-wide association studies. The American Journal of Human Genetics, 2024, 480 111(5): 966-978. 481 [19] Lai Y, Adam B, Podolsky R, et al. A mixture model approach to the tests of 482 concordance and discordance between two large -scale experiments with two -sample 483 groups. Bioinformatics, 2007, 23(10): 1243-1250. 484 [20] Wang B, Pozarickij A, Mazidi M, et al. Comparative studies of 2168 plasma 485 proteins measured by two affinity -based platforms in 4000 Chinese adults. Nature 486 Communications, 2025, 16(1): 1869. 487 [21] UK biobank news. Launch of world ’s most significant protein study set to usher 488 in new understanding for medicine. 2025 (https://www.ukbiobank.ac.uk/). 489 [22] Deng Y T, You J, He Y, et al. Atlas of the plasma proteome in health and disease 490 in 53,026 adults. Cell, 2025, 188: 253-271. 491 [23] Grasby K L, Jahanshad N, Painter J N, et al. The genetic architecture of the human 492 cerebral cortex. Science, 2020, 367(6484): eaay6690. 493 [24] Dong Z, Li P, Jiang Y, et al. Integrative Multi‐Omics and Routine Blood Analysis 494 Using Deep Learning: Cost‐Effective Early Prediction of Chronic Disease Risks. 495 Advanced Science, 2025: 2412775. 496 [25] Bell J. Predicting disease using genomics. Nature, 2004, 429(6990): 453-456. 497 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint [26] Pereira P R, Carrageta D F, Oliveira P F, et al. Metabolomics as a tool for the early 498 diagnosis and prognosis of diabetic kidney disease. Medicinal Research Reviews, 2022, 499 42(4): 1518-1544. 500 [27] Hä llqvist J, Bartl M, Dakna M, et al. Plasma proteomics identify biomarkers 501 predicting Parkinson’s disease up to 7 years before symptom onset. Nature 502 Communications, 2024, 15(1): 4759. 503 [28] Carrasco-Zanini J, Pietzner M, Davitte J, et al. Proteomic signatures improve risk 504 prediction for common and rare diseases. Nature Medicine, 2024, 30(9): 2489-2498. 505 [29] Border R, Johnson E C, Evans L M, et al. No support for historical candidate gene 506 or candidate gene-by-interaction hypotheses for major depression across multiple large 507 samples. American Journal of Psychiatry, 2019, 176(5): 376-387. 508 [30] Lee J J, Wedow R, Okbay A, et al. Gene discovery and polygenic prediction from 509 a genome-wide association study of educational attainment in 1.1 million individuals. 510 Nature Genetics, 2018, 50(8): 1112-1121. 511 [31] Guo Y, You J, Zhang Y, et al. Plasma proteomic profiles predict future dementia 512 in healthy adults. Nature Aging, 2024, 4(2): 247-260. 513 [32] Fuchs T, Trollor J N, Crawford J, et al. Macrophage inhibitory cytokine ‐1 is 514 associated with cognitive impairment and predicts cognitive decline - the Sydney 515 Memory and Aging Study. Aging Cell, 2013, 12(5): 882-889. 516 [33] Frye M A, Nassan M, Jenkins G D, et al. Feasibility of investigating differential 517 proteomic expression in depression: implications for biomarker development in mood 518 disorders. Translational Psychiatry, 2015, 5(12): e689-e689. 519 [34] Lemmelä S, Wigmore E M, Benner C, et al. Integrated analyses of growth 520 differentiation factor-15 concentration and cardiometabolic diseases in humans. Elife, 521 2022, 11: e76272. 522 [35] Shchelochkov O A, Farmer C A, Chlebowski C, et al. Intellectual disability and 523 autism in propionic acidemia: a biomarker -behavioral investigation implicating 524 dysregulated mitochondrial biology. Molecular Psychiatry, 2024: 1-8. 525 [36] Kang J, Yang L, Jia T, et al. Plasma proteomics identifies proteins and pathways 526 associated with incident depression in 46,165 adults. Science Bulletin, 2024 , 70(4): 527 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint 573-586. 528 [37] Eagleson K L, Campbell D B, Thompson B L, et al. The autism risk genes MET 529 and PLAUR differentially impact cortical development. Autism Research, 2011, 4(1): 530 68-83. 531 [38] Zhu Y, Bujo H, Yamazaki H, et al. LR11, an LDL receptor gene family member, 532 is a novel regulator of smooth muscle cell migration. Circulation Research, 2004, 94(6): 533 752-758. 534 [39] Walker R M, Chong M, Perrot N, et al. The circulating proteome and brain health: 535 Mendelian randomisation and cross-sectional analyses. Translational Psychiatry, 2024, 536 14(1): 204. 537 [40] Li M, Yang X, Sun L, et al. Decreased serum apolipoprotein A4 as a potential 538 peripheral biomarker for patients with schizophrenia. Journal of Psychiatric Research, 539 2021, 137: 14-21. 540 [41] Llansola M, Izquierdo -Altarejos P, Montoliu C, et al. Role of peripheral 541 inflammation in minimal hepatic encephalopathy. Metabolic Brain Disease, 2024: 1 -542 11. 543 [42] Xu H, Jia J. Immune -related hub genes and the competitive endogenous RNA 544 network in Alzheimer ’s disease. Journal of Alzheimer ’s Disease, 2020, 77(3): 1255 -545 1265. 546 [43] Huo S, Clocchiatti -Tuozzo S, Magid -Bernstein J, et al. Integrated Genomic and 547 Proteomic Drug Target Discovery for Vascular Dementia and Alzheimer's Disease (P3-548 3.015). Neurology, 2025, 104(7_Supplement_1): 4123. 549 550 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint

551 Stu y participants 552 This study included data from the UKB and integrated multiple data sources, 553 including blood collection, imaging data, and various self-reported questionnaires [44, 554 45]. All participants provided explicit, written informed consent to the UKB. The UKB 555 cohort received approval from the NHS National Research Ethics Service North West 556 (reference number: 16/NW/0274). 557 Blood samples for proteomic analysis were collected and processed at Olink 558 Analytical Services using the antibody -based Olink Explore ™ Proximity Extension 559 Assay. A total of 2,923 distinct proteins were measured, with stringent quality control 560 procedures applied as outlined in the previous studies [1, 3] . Additional details 561 regarding sample selection, processing, and quality control procedures were available 562 in previous publications [1, 3] . The reported normalized protein expression (NPX) 563 values from Olink were utilized. After further data processing , our study included a 564 total of 52,632 individuals and 2,920 unique proteins ( details see Supplementary 565 Materials). 566 Genotype data were available for 502,493 participants in the UK Biobank v3 567 imputation. Detailed genotyping and quality control procedures performed by UK 568 Biobank were described in a previous publication [46]. Our study excluded SNPs with 569 call rates < 9 5%, minor allele frequency < 0.5% and deviation from the Hardy –570 Weinberg equilibrium with P < 1 ×  −6. Participants with less than 5% missing rates, 571 not outliers in heterozygosity, who had no sex chromosome aneuploidy, of British 572 ancestry, and who had no m ore than ten putative third -degree relatives in the kinship 573 table were selected. 574 We used the nuclear magnetic resonance (NMR) metabolomics data in the UKB, 575 which were recently released by Nightingale Health, containing around 292,000 576 individuals [47]. Here, a total of 249 metabolomics biomarkers were directly provided 577 in the UKB and the details of these biomarkers could be found in Supplementary Table 578 13. More details for data processing were available in previous publications [47]. 579 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint We used the brain structur e imaging measures derived from T1 imaging data, as 580 processed by WIN FMRIB on behalf of UK B [ 48]. The detailed preprocessing 581 information was provided in Supplementary Materials. Both cerebral hemispheres and 582 66 regions defined by the Desikan-Killiany (DK) atlas for total CSA, total CV and mean 583 CT were estimated [49]. 584 585 Calculation o z-scores 586 The general linear model was used to test the association between the 587 proteomics/genomics/metabolomics data and brain-related traits. The effects of certain 588 covariates (i.e. sex and age for plasma proteomics/metabolomics; sex, age and the first 589 20 genetic principal components for genomics) were regressed out. Then, we obtained 590 an upper-tailed p-value. For each p-value, we performed a transformation based on the 591 inverse normal cumulative distribution function (c.d.f.) into a z-score: 592 𝑧=Ф−1( −𝑝), 593 where Ф−1 was the inverse function of the standard normal cumulative distribution 594 function. 595 596 Mixture model-based replicability assessment (MMRA) 597 For two lists of z-scores: {[𝑧𝑘 (1),𝑧𝑘 (2)]: k = 1, 2, ..., m}, where m was the number of 598 common units (i.e., proteins, SNPs and metabolites) from two different datasets, we 599 considered a nine -component normal -mixture model for the joint distribution (see 600 above for z-score calculation): 601 𝑓[𝑧(1),𝑧(2)]= ∑ ∑ 𝜋𝑖𝑗 2 𝑗=0 2 𝑖=0 𝜙𝜇𝑖,𝜎𝑖 2[𝑧(1)]𝜙𝜈𝑗,𝜏𝑗 2[𝑧(2)], 602 where 𝜙𝜇,𝜎2 was the normal probability distribution function with mean 𝜇 and 603 variance 𝜎2. We used the first component (index 0) to represent the null (no change/ 604 association) feature component. Then, 𝜇0 = 𝜈0 = 0 and 𝜎0 2 = 𝜏0 2 = 1. The second 605 and third components (indices 1 and 2) were used to represent negative and positive 606 associations. Their corresponding parameters (means and variances) were estimated 607 from the paired z-scores with the following constraints: 𝜇1,𝜈1 ≤ and 𝜇2,𝜈2 ≥ . 608 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint 𝜋𝑖𝑗 was the proportion for component 𝑖 in the first association study and component 609 𝑗 in the second association study, and ∑ 𝜋𝑖𝑗𝑖𝑗 = . 610 This model was termed as partial concordance/discordance (PCD) model [17, 19]. 611 Then, we defined 𝜋𝑁𝑢𝑙𝑙, 𝜋𝑅 and 𝜋𝐼𝑅 as the proportions of the non-signals, replicable, 612 and irreplicable signals, respectively. These three latent categories were represented by: 613 𝜋𝑁𝑢𝑙𝑙 = 𝜋00, 614 𝜋𝑅 = 𝜋11+ 𝜋22, 615 𝜋𝐼𝑅= − ∑ 𝜋𝑖𝑗𝑖=𝑗 . 616 We reported an overall irreplicability quantity 𝜌𝐼𝑅 to measure the relative proportion 617 of irreplicable signals in non-null signals: 618 𝜌𝐼𝑅= 𝜋𝐼𝑅 𝜋𝐼𝑅+𝜋𝑅 . 619 Additionally, for each unit k in an association study, we also defined the posterior 620 probability of replicability as follows: 621 𝑆𝑘 − =Pr(𝑢𝑛𝑖𝑡 𝑋𝑘 𝑖𝑠 𝑟𝑒𝑝𝑙𝑖𝑐𝑎𝑏𝑙𝑒 negative association|[𝑧𝑘 (1),𝑧𝑘 (2)]) =622 𝜋11𝜙𝜇1,𝜎12[𝑧𝑘 (1)]𝜙𝜈1,𝜏12[𝑧𝑘 (2)] ∑ ∑ 𝜋𝑖𝑗2 𝑗=0 2 𝑖=0 𝜙𝜇𝑖,𝜎𝑖 2[𝑧𝑘 (1)]𝜙𝜈𝑗,𝜏𝑗 2[𝑧𝑘 (2)] , 623 𝑆𝑘 + =Pr(𝑢𝑛𝑖𝑡 𝑋𝑘 𝑖𝑠 𝑟𝑒𝑝𝑙𝑖𝑐𝑎𝑏𝑙𝑒 positive association|[𝑧𝑘 (1),𝑧𝑘 (2)])=624 𝜋22𝜙𝜇2,𝜎22[𝑧𝑘 (1)]𝜙𝜈2,𝜏22[𝑧𝑘 (2)] ∑ ∑ 𝜋𝑖𝑗2 𝑗=0 2 𝑖=0 𝜙𝜇𝑖,𝜎𝑖 2[𝑧𝑘 (1)]𝜙𝜈𝑗,𝜏𝑗 2[𝑧𝑘 (2)] . 625 This estimated probability 𝑆̂𝑘 ± of 𝑆𝑘 ± could be calculated by plugging-in the estimated 626 parameters in the PCD model. Similarly, we also reported an individual irreplicability 627 quantity 𝛾𝑘 to measure the relative proportion of irrep licable positive/negative 628 association and no association for each unit k in an association study: 629 𝛾𝑘 = −𝑆𝑘 −−𝑆𝑘 +. 630 In this context, a lower 𝛾𝑘 value indicated a higher probabilit y of replicable 631 positive/negative association for unit k. 632 633 Simulation design 634 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint Our simulations were designed based on the proteomics data from the UK Biobank. 635 First, we split the original data into two subsets (referred to as Data 1 and Data 2 based 636 on the order of subject number) with equal sample sizes. Then, we partitioned each 637 subset randomly into two further subsets (referred to as Data 1A, Data 1B, Data 2A and 638 Data 2B). Before the analysis, we ensured that sex, age was statistically similar between 639 Data 1A versus 2A as well as Data 1B versus 2B (t test for age and chi-square test for 640 sex, P > 0.05). Otherwise, we repeated the random data partition until this similarity 641 requirement was satisfied. For each feature, there was no statistically significant 642 differences in distribution between Data 1A versus 2A nor Data 1B versus 2B. 643 To generate upward or downward changes, a protein set was randomly chosen and 644 an adjustment of 0.0 123-0.0369 standard deviations of all the protein expression 645 (corresponding to approximately 1-3 effect sizes in z-scores) were randomly added to 646 or subtracted from the expression levels of the chosen protein set for each subject in 647 Data 1A and Data 1B. This procedure was repeated 1,000 times. For each repetition, 648 we obtained two lists of z-scores: one by protein-wisely comparing Data 1A versus Data 649 2A and the other by comparing Data 1B versus Data 2B. Z-scores were calculated based 650 on the traditional two-sample t-test. A pair of z-scores were obtained for each protein. 651 The replicability between two lists of z-scores was assessed by the MMRA approach. 652 The following three simulations were considered. 653 (a) No un-replicable signal . According to our random data partition, there were no 654 statistically significant differences between Data 1A versus 2A or Data 1B versus 2B. 655 We modified the 100% of null (no change) to 60% null, 20% upward changes and 20% 656 downward changes as follows. We randomly selected two protein sets, each with 20% 657 of the total proteins. To simulate 20% upward changes, for each protein in the first set, 658 we randomly added a value equivalent to 1 -3 effect sizes in z-scores to each subject's 659 protein expression levels in Data 1A and repeated this process in Data 1B to ensure 20% 660 replicable upward changes. For each protein in the second set, we randomly subtracted 661 a value equivalent to 1-3 effect sizes in z-scores from each subject's protein expression 662 in Data 1A and repeated this process in Data 1B to achieve 20% replicable downward 663 changes. 664 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint (b) Moderate level of un-replicable signal. We randomly selected four protein sets. The 665 first two sets each comprised 15% of the total proteins, and the upward changes and 666 downward changes were simulated as described in (a). The next two sets each 667 comprised 5% of the total proteins. For each protein in the third set, we randomly added 668 a value equivalent to 1 -3 effect sizes in z-scores to each subject's protein expression 669 level in Data 1A (but not in Data 1B). Then, we had 5% discordant changes (up versus 670 null). Similarly, for each protein in the fourth set, we subtracted a value from each 671 subject's protein expression level in Data 1A (but not in Data 1B) so that we had 5% 672 discordant changes (down versus null). 673 (c) High level of un-replicable signal. Considering that the replicability levels may vary 674 across different studies, we randomly selected four protein sets, each with 10% of the 675 total proteins. The replicable upward/downward changes (the first/second set) and un-676 replicable upward/downward changes (the third/fourth set) were simulated similarly as 677 described in (b). 678 679 Overall replicability assessment 680 To investigate the overall replicability (𝜌𝐼𝑅) for proteomics-based association study 681 (namely PBAS) results with different phenotypes , we considered a random 682 subsampling approach which randomly split the whole data into two subsets with 683 (approximately) equal sample sizes for 1,000 times . Due to constraints from missing 684 observations, different sample sizes were included in different PBAS scenarios. Aside 685 from proteomics data, we used the following four types of measures in our study: 686 physical measures, cognitive function, mental health and brain imaging metrics (details 687 for each trait see Supplementary Table 1). The brain structure imaging metrics included 688 total cortical surface area (CSA), total cortical volume (CV) and mean cortical thickness 689 (CT). We calculated the overall replicability ( 𝜌𝐼𝑅) for these brain metrics in both 690 cerebral hemispheres ( N = 5,623). We also assess the overall replicability for these 691 metrics based on the Desikan –Killiany (DK) atlas which included 66 regions. Using 692 the above data, we obtained 1,000 pairs of z-scores and assessed the median 𝜌𝐼𝑅 for 693 each PBAS scenario. 694 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint 695 Effects of season and fasting time at blood collection 696 To assess the effects of season and fasting time on replicability, we further included 697 these factors as additional covariates in the PBAS. The season of blood collection was 698 categorized as summer/autumn (June to November) or winter/spring (December to 699 May), based on the blood collection date. Fasting time was determined from 700 participant-reported fasting duration prior to blood collection. 701 702 Multi-omics comparison 703 The protein quantitative trait locus (pQTL) identified in a previous study [1] were 704 adopted to compare the overall replicability level between genetics -based association 705 analyses and proteomics -based association analyses. After quality control, 43,685 706 participants with both genetics and proteomics data were included for this analysis. 707 Each participant had 2,920 proteins and their corresponding 6,386 pQTL-related SNPs 708 in UKB [1] . Moreover, to compare the overall replicability level between 709 metabolomics-based association analyses and proteomics -based association analyses, 710 we then included 30,079 participants with both metabolomics and proteomics data. The 711 PBAS was conducted based on the same procedure as mentioned in the previous section. 712 The same procedure was also employed to analyze the associations between the 713 genetics/metabolomics data and each brain-related measure (within physical measures, 714 cognitive function, mental health and brain imaging metrics). Then, we obtained 1,000 715 pairs of z -scores based on 1,000 random subsampling times; and we calculated the 716 median 𝜌𝐼𝑅 for each association analysis scenario. 717 The PBAS scenarios considered here excluded subjects without British ancestry 718 (see Study Participants Section in Methods for details ) when compare the overall 719 replicability level between genetics -based association analyses and proteomics -based 720 association analyses. 721 722 Influential factors 723 According to the details of UKB proteomics data collection 724 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint (http://biobank.ndph.ox.ac.uk/ukb/ukb/docs/Olink_proteomics_data.pdf), two versions 725 of the assay panel were employed by UKB to collect proteomics data which included 726 “Explore 1,536 assay panels” and “Expansion 1460 assay panels” (referred as Panel 1 727 and Panel 2, respectively). 728 Here, we stratified all plasma proteins into three categories based on their dilution 729 levels: 1,941 proteins at least-abundant (1:1), 524 proteins at moderate-abundant (1:10) 730 and 455 proteins at more-abundant (1:100 to 1:100,000). Moreover, we further 731 stratified the proteins into four categories according to the proportion of samples below 732 the LOD: 50% (657 proteins). Then, we counted the number of proteins across twelve 734 categories (i.e., three dilution level categories × four proportion of samples below the 735 LOD categories) in Panel 1 and Panel 2. To evaluate the impact of sample size on 736 overall replicability, we randomly selected two settings with approximately 50% and 737 25% sample size from the original dataset. Then, we split each setting into two subsets 738 with (approximately) equal sample sizes. This process was repeated 1,000 times. We 739 then calculated the overall irreplicability quantity 𝜌𝐼𝑅 based on each of these subsets. 740 741 Pre iction mo eling 742 The proteomics technologies are currently under active developments, and more 743 proteomics panels to include additional proteins may be made available in the near 744 future. Here, we assumed that a total of 1,460 proteins would still be included in a future 745 proteomics panel, and we intend ed to estimate the related overall replicability of 746 associations (based on simulations). For this analysis, given a phenotype and a 747 simulated panel, we defined the term “overall replicable” as 𝜌𝐼𝑅< 0.05 (i.e., binary 748 response variable). With the dilution level (divided into three categories), the proportion 749 of samples below the LOD (divided into four categories), and the sample size, we can 750 construct a logistic regression model for this purpose. Please see Supplementary 751

for the related details and comprehensive simulation results. 752 753 Individual replicability assessment 754 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint To assess the individual replicability (𝛾𝑘) of proteins in the PBAS within cognitive 755 function and mental health, we conducted further analyses for each PBAS scenario 756 which exhibited high levels of overall replicability ( median overall irreplicability 757 quantity 𝜌𝐼𝑅< 0.05). Initially, we calculated the median individual irreplicability 758 quantity 𝛾𝑘 for each protein in 1,000 random subsampling times. These median 759 𝛾𝑘 values were then ranked in an ascending order. 760 In this study, we identified the pleotropic proteins for cognitive function and mental 761 health that demonstrated the highest replicability associations with more than 50% of 762 the traits under investigation. The pleotropic proteins were selected based on the 763 following criteria: they must be identified as one of the top ten proteins in at least 10 764 mental health-related traits (more than 50% of total 19 traits) or 3 cognitive function-765 related traits (more than 50% of total 5 traits). The random subsampling approach (See 766 Overall replicability assessment section in Methods) was used to simulate practical 767 study cohorts and their replications. For each random subsampling, we defined 768 individual irreplicability quantity 𝛾𝑘 < 0.05 as the criterion to indicate that a protein 769 exhibited directional consistency. For each trait in the cognitive function and mental 770 health, the fraction of directional consistency (FDC) was defined as the fraction of times 771 that a protein demonstrated directional consistency ( 𝛾𝑘 < 0.05) across 1,000 random 772 subsampling. 773 We further investigated the impact of proteomics data quality on individual 774 replicability for proteins. Here, for each trait under cognitive function and mental health 775 with median overall irreplicability quantity 𝜌𝐼𝑅< 0.05, we calculated median 776 individual irreplicability quantity 𝛾𝑘 for each protein from 1,000 random subsampling 777 times. Then, for all proteins' median 𝛾𝑘 , we reported the median value for Panel 1 778 versus Panel 2. We also analyzed the impact of dilution levels and the proportion of 779 samples below the LOD for individual replicability performance. 780 781 Data availability 782 The data used in the study from the UKB was accessible under restricted access 783 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint (application number 19542). Access can be obtained by submitting an application 784 through the UKB platform (https://www.ukbiobank.ac.uk/). 785 786 Code availability 787 Code for overall/individual replicability assessment and predictive framework for 788 potential future panels is openly shared in GitHub (https://github.com/YixinZhang-789 stat/plasma-proteomics-replicability). R version 4. 3.0 and R package ggseg was used 790 to show median 𝜌𝐼𝑅 for each region in DK atlas. 791 792 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint

793 [44] Sudlow C, Gallacher J, Allen N, et al. UK biobank: an open access resource for 794 identifying the causes of a wide range of complex diseases of middle and old age. PLoS 795 Medicine, 2015, 12(3): e1001779. 796 [45] Bycroft C, Freeman C, Petkova D, et al. The UK Biobank resource with deep 797 phenotyping and genomic data. Nature, 2018, 562(7726): 203-209. 798 [46] Zhang B, You J, Rolls E T, et al. Identifying behaviour -related and physiological 799 risk factors for suicide attempts in the UK Biobank. Nature Human Behaviour, 2024: 800 1-14. 801 [47] Qiang Y X, You J, He X Y, et al. Plasma metabolic profiles predict future dementia 802 and dementia subtypes: a prospective analysis of 274,160 participants. Alzheimer's 803 Research & Therapy, 2024, 16(1): 16. 804 [48] Alfaro-Almagro F, Jenkinson M, Bangerter N K, et al. Image processing and 805 Quality Control for the first 10,000 brain imaging datasets from UK Biobank. 806 Neuroimage, 2018, 166: 400-424. 807 [49] Desikan R S, S égonne F, Fischl B, et al. An automated labeling system for 808 subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. 809 Neuroimage, 2006, 31(3): 968-980. 810 811

This work was partially supported by a start -up fund from the 812 University of Science and Technology of China (to Yinglei Lai). National Key R&D 813 Program of China (2019YFA0709502, 2018YFC1312904 to Jianfeng Feng), Shanghai 814 Municipal Science and Technology Major Project (2018SHZDZX01 to Jianfeng Feng), 815 111 Project (B18015 to Jianfeng Feng), Humboldt Research Award (to Jianfeng Feng). 816 Some image materials in Figure 1 were free acquired from Freepik (www.freepik.com). 817 818 Author contributions Conception: Zeyu Jiao, Yinglei Lai, Yixin Zhang and Jianfeng 819 Feng. Design: Zeyu Jiao, Yixin Zhang, Yinglei Lai, Jujiao Kang and Jianfeng Feng. 820 Data acquisition, processing and analysis: Zeyu Jiao, Yixin Zhang, Yinglei Lai, Jujiao 821 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint Kang, Wei Zhao, Jia You , Wei Cheng and Jianfeng Feng . Manuscript writing and 822 revising: Zeyu Jiao, Yixin Zhang, Yinglei Lai, Jujiao Kang, Liang Ma and Jianfeng 823 Feng. 824 825 Competing interests The authors declare no competing interests. 826 preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for thisthis version posted September 5, 2025. ; https://doi.org/10.1101/2025.09.01.673490doi: bioRxiv preprint