Estimating protein isoform abundances with PAQu}} Authors: Lorenzo Testa, Lambertus Klei, Alesia Rengle, Anastasia K. Yocum, David A. Lewis, Bernie Devlin, Kathryn Roeder, Matthew L. MacDonald Published: 2026 ABSTRACT SignificanceTranscripts of genes are translated into protein isoforms, which play crucial roles in biological processes; dysregulation of translation can generate disease. In large-scale studies, however, accurately estimating isoform abundances is challenging because only short portions of isoforms, peptides, can be quantified. To address this limitation, we propose PAQu (Protein isoform Abundance Quantification), a method that integrates data from both the peptidome and transcriptome, enabling more accurate estimates of isoform abundance. Importantly, PAQu provides uncertainty quantification and allows testing of hypotheses, such as differential abundance between cases and controls, making it a powerful tool to study diseases at the isoform level. Its superior performance in simulations and suitability for real-world applications, including schizophrenia research, opens avenues to understand complex disease mechanisms. Alternative splicing is a crucial posttranscriptional process (1, 2) that generates multiple protein isoforms from a single gene, thereby contributing to the complexity of eukaryotic organisms. Dysregulation of this process has been linked to various diseases, including cancer (3, 4), neurodegenerative and psychiatric disorders (5–7), and developmental abnormalities (8–10). Despite the significant role of alternative splicing in cellular function (2), accurate and comprehensive quantification of protein isoforms remains a major challenge in proteomics (11, 12). In a typical shotgun proteomics experiment, proteins are first digested into peptides and then analyzed by liquid chromatography tandem mass spectrometry (LC-MS/MS). The resulting spectra are used to identify and quantify peptides in digested samples. Although this approach is effective for analyzing individual peptides, it often falls short when the goal is to characterize protein isoforms (13). In particular, a significant challenge in protein isoform identification and quantification is the presence of shared peptides (11–13), which can map onto multiple isoforms of the same protein or even to different proteins due to sequence similarity. Additionally, peptides unique to individual isoforms are typically less abundant compared to those that are shared. To overcome this challenge, standard practices either collapse proteins with shared peptides into protein groups or ignore shared peptides altogether, which limits isoform characterization (14–16). Several other strategies aggregate peptide abundances into protein isoform estimates by taking summary statistics (e.g., average, sum, maximum) without considering whether peptides are shared or isoform-specific. These approaches are reflected in the default settings in the commonly used informatics programs, such as Proteome Discoverer (17) and MS Fragger (18). As a consequence, standard methods often struggle to distinguish between isoforms with high sequence similarity, leading to underestimation of isoform diversity or potential misinterpretation of biological processes. To improve isoform quantification, recent approaches have attempted to integrate transcriptomic data with proteomic data, using the relationship between gene expression and protein production. These methods offer a promising direction for tackling the challenge of protein quantification. In particular, several studies (12, 19, 20) have introduced techniques that use transcript expression levels as a prior source of information to estimate isoform abundances. However, despite their advances, these methods come with several limitations. For example, Carlyle et al. (19) employ an expectation–maximization (EM) algorithm that does not provide any direct measure of uncertainty quantification. Miller et al. (12) rely on long-read RNA-seq, an expensive technique to obtain transcript data. Bollon et al. (20) design a Bayesian model that operates on each sample individually, lacking knowledge transfer between observations. Moreover, none of these methods provides a direct way to evaluate the presence of differentially abundant protein isoforms between different groups of samples. To overcome these shortcomings, we introduce a method—which we call PAQu (Protein isoform Abundance Quanti-fication)—for accurate Protein isoform Abundance Quantification. PAQu combines the strengths of transcriptomic and proteomic data, exploiting complementary insights from RNA-Seq and mass spectrometry. It offers robust uncertainty quantification for all estimated parameters, integrates external covariate information, elucidates the relationship between the transcriptome and the proteome, and enables simultaneous differential analysis, providing a comprehensive solution for isoform quantification. Results Overview of PAQu. PAQu is a Bayesian supervised factor analysis method designed for Protein isoform Abundance Quantification. It models protein isoform abundances as latent factors and estimates them by learning two key mappings: one from transcript expression levels to protein isoform abundances, and another from protein isoform to peptide abundances. Simultaneously, PAQu assesses the impact of binary conditions on isoform abundances, enabling analysis of differential protein abundance. Our method takes four inputs: a transcript expression matrix from tissue samples from n subjects, a peptide abundance matrix from tissue from the same (or possibly different) subjects, a detectability mask matrix indicating whether a peptide k can come from a protein isoform j, and a binary vector representing a condition for subjects, such as diagnosis (Fig. 1). PAQu assumes that transcript expression levels are proportional to the corresponding protein isoform abundances, and that peptide abundances are related to their respective protein isoforms. This naturally leads to a two-layer model. In the first layer, the transcript expression matrix T is transformed into the latent protein isoform matrix I through a diagonal weight matrix W. In the second layer, the peptide abundance matrix P is modeled as the product of the protein isoform matrix I and the detectability matrix Z (whose entries can be zero according to the detectability mask matrix provided as input). Moreover, for differential protein abundance analysis, the first layer also incorporates the effect of the binary condition vector A on protein isoform abundances I through the vector of effects D. Conceptually, the first layer performs multivariate linear regression of protein isoforms I on transcripts T and condition A, while the second layer applies factor analysis to P. By constraining, in the first layer, the direction of the protein isoform factors within the space spanned by the transcripts, PAQu supervises, in the second layer, the estimation of the protein factors. Finally, if available, PAQu can also take advantage of the information coming from external covariates, easily inserting them into the previous model (Materials and Methods). Summary of PAQu model. The model takes as input a transcript expression matrix T, a binary condition vector A, a detectability mask matrix M, and a peptide abundance matrix P. The output includes the effects of transcripts on protein isoforms W, the difference in protein isoform abundances across conditions D, and the estimated associations between isoforms and peptides Z. Although this two-layer problem is not identifiable in the traditional frequentist sense, imposing prior distributions on the unknowns allows estimation of the quantities of interest. Specifically, for the PAQu model, we use Gaussian priors for the protein isoform abundances I, the weights in W, the effects in D, and truncated Gaussian priors for the detectability factors in Z. For situations with a high signal-to-noise ratio, users can opt for a sparser spike-and-slab prior (21), which offers more conservative inference by selectively shrinking smaller effects. We implement a Gibbs sampling algorithm to generate posterior samples of the model parameters (22). The significance of any effect can be measured by the local false sign rate (LFSR) metric (23), a summary of the posterior distribution akin to local false discovery rate (Materials and Methods). Similarly, for parameters with sparse priors, posterior inclusion probability (PIP) can be easily computed. PAQu generates several outputs: estimated associations between transcripts and protein isoforms, inferred associations between protein isoforms and peptides, estimated abundances of protein isoforms, and a list of differentially abundant protein isoforms, which can be interpreted, for instance, through gene ontology (GO) enrichment analysis of genes associated with each isoform. All of these estimated quantities are accompanied by a measure of their uncertainty. Simulation Results Provide Strong Motivation for PAQu}}. We evaluate the performance of PAQu in two main simulation settings. In the first setting, which we refer to as “easy”, we generate data without shared peptides. Specifically, we sample one transcript (q=1) from a Normal distribution, map it into a protein isoform with random conversion weight, add an effect due to a binary condition, and then split it into two peptides (r=2) using randomly assigned detectability scores (Materials and Methods). This setting serves as a reference point to compare PAQu with conventional methods for protein quantification. We run each simulation experiment with three different sample sizes (n=100,200,500), three effect sizes for the binary condition (Dj=0.33,0.66,1), and with 25 different seeds. In the second, more challenging scenario—referred to as the hard setting—we generate data that include shared peptides. Here, we sample five transcripts (q=5) from a multivariate Normal distribution, map them to their corresponding protein isoforms with random conversion weights, and introduce a binary condition effect to some isoforms. These isoforms are split into 10 peptides (r=10) using random detectability scores, with 30% of the detectability matrix entries being nonzero. Each simulation is run across three sample sizes (n=100,200,500), three effect sizes (Dj=0.33,0.66,1), three numbers of differentially abundant isoforms (|Dact|=1,2,3) and 25 different random seeds. We exploit simulated data to assess the performance of various prior distributions in differential isoform abundance tests. In our simulations, the effects inferred by using the spike-and-slab prior are overly conservative, particularly when the signal or the sample size is small. In contrast, the Normal prior exhibits much less bias (SI Appendix, Figs. S3 and S4), which supports our decision to adopt the Normal prior as default. We then evaluate the ability of PAQu to recover and quantify differentially abundant protein isoforms. PAQu accurately estimates these effects, showing a slightly conservative behavior in the simulation setting hard in terms of recovery of differentially abundant isoforms at LFSR ≤ 0.05 when the sample size and the signal are small (SI Appendix, Figs. S6 and S7). This leads to a modest increase in false negatives, as expected due to the conservative prior used and the extreme weakness of the signal. To assess the performance of PAQu in inferring the latent protein isoform factors, we measure the absolute correlation between the inferred and true isoform abundances. In all scenarios, the estimated protein factors show a strong correlation with the true factors (Fig. 2 and SI Appendix, Fig. S5). As expected, the quality of the estimates increases with the sample size; at the same time, estimates tend to slightly worsen with increasing number of parameters to estimate. Simulation results. (A) Distributions of the absolute correlation values between true protein isoforms and the estimates inferred by PAQu under the easy (Left) and hard (Right) scenarios. Different colors represent different sample sizes, ranging from 100 to 500. (B) Box plots of absolute effect sizes estimated by PAQu under the easy (Left) and hard (Right) scenario. Different colors represent different sample sizes. For each box, the Center line represents the median; the Lower and Upper hinges correspond to the first and third quartiles; the Upper and Lower whiskers span 2 times the interquartile range. (C) Comparison between the proportion of detected effects Dj under the easy scenario between PAQu (LFSR≤0.05), t test on the average of peptide abundances, t test on the sum of peptide abundances, t test on the maximum peptide abundance (P−value≤0.05). All simulations were repeated 25 times. Note that LFSR controls for directional error, making it a structurally more conservative threshold than an uncorrected nominal P-value (see Materials and Methods for a detailed comparison). We then compare the performance of PAQu in detecting differentially abundant isoforms with other commonly used tools: t test on the average peptide abundance, t test on the sum of peptide abundance, and t test on the maximum peptide abundance. PAQu outperforms all these methods in all simulation settings (Fig. 2 and SI Appendix, Figs. S6–S8). These results demonstrate not only the performance gains PAQu can provide, but also the value of integrating data from both the transcriptome and the peptidome to achieve more precise isoform-level estimates. Moreover, to further challenge PAQu and measure its performance under misspecification of the model, we evaluate simulations in which we assume that there is no link between transcript expression levels and protein isoform abundances (i.e. W=0). Remarkably, even when transcripts carry no information about protein isoforms, PAQu can still provide meaningful inference, significantly outperforming competitors (SI Appendix, Figs. S10–S12). To evaluate the contribution of information from transcript expressions, we performed simulations in which we compared the estimates of PAQu without transcripts, average values between samples only, and full transcript expression levels. As the amount of transcript expression increases, so does the quality of isoform abundance estimates, measured by the absolute correlation between true and estimated abundances (SI Appendix, Fig. S13). Finally, we assess computational feasibility. Although T and P can be large, because Z is block diagonal, computations via the Gibbs sampler can proceed in parallel by block, making them efficient (SI Appendix, Fig. S14). PAQu}} Reveals Cortical Protein Isoform Differences Associated with Schizophrenia. We applied PAQu to the most deeply characterized proteomic dataset of human brain tissue for a psychiatric disorder. Quantitative proteomics using TMT labeling and deep offline fractionation was applied to postmortem dorsal anterior cingulate cortex gray matter from 56 individuals diagnosed with schizophrenia and 56 matched controls with no history of psychiatric or neurological conditions. While “protein” expression alterations have previously been observed in schizophrenia, here we report differentially abundant protein isoforms. Furthermore, we have complementary data on transcripts, based on RNA-seq measurements (24), in adjacent tissue sections from a subset of the same subjects. After collecting and preprocessing data, which included adjustment for covariates such as postmortem interval (see Materials and Methods and SI appendix), we have protein and transcript data from n=103 subjects, 53 of whom are affected by schizophrenia, and consisting of 9,169 isoforms and r=79,040 peptides. Of 9,169 isoforms, 3,338 shared exactly the same measured peptides, and therefore were not uniquely estimable. We rolled them up into 1,440 isoform groups. Of the 1,440 isoform groups, 1,396 represented isoforms from a single gene; the remaining 44 represented multiple genes. PAQu calculations were performed using the combined q=7,271 singleton and grouped isoforms. We ran PAQu with 10 different MCMC chains, each consisting of 3,000 iterations (the first 2,000 are burn-in), both to measure PAQu consistency and to gauge the stability of our findings. For differential case–control abundance, SI Appendix, Fig. S15 shows the detection frequency results across chains at LFSR≤0.05. Out of 7,271 isoforms (SI Appendix, Table S1), PAQu never selected 6,211 isoforms (85.4%) as differentially abundant, and always detected 541 isoforms (7.4%). The remaining 519 (7.1%) isoforms are selected at least once, but only 179 of them are detected more than 5 times. The high degree of consistency among runs indicates that PAQu converges to estimates that are robust to Monte Carlo variability. While this consistency does not establish the biological correctness of the findings, it increases confidence that the reported results are reproducible. One of the most critical open questions in the biology of schizophrenia is the mechanism by which genetic risk contributes to decreased synaptic connectivity, which is believed to drive multiple symptom types (25, 26). In part, genetic risk has been traced to transcript-level alterations directly relevant to disease pathology (6), in that distinct protein isoforms directly and differentially regulate synaptic structure and function. PAQu confirmed some of these alterations also affect protein isoform abundances (SI Appendix, Table S1). For example, RTN4A, decreased in schizophrenia at the transcript level, regulates neurite outgrowth and other aspects of excitatory synaptic activity (27) through domains not present in the other RTN-A isoforms. Specifically, the “Nogo-66 Loop” facilitates interactions with NRG1 and p75NTR, activating RhoA/ROCK signaling, initiating growth cone collapse. PAQu confirms this reduction at the isoform level. Similarly, SLC1A2 encodes multiple EAAT2 isoforms, which clear the majority of glutamate from the excitatory synaptic cleft. However, each isoform, defined by toggling terminal sequences, has different localization and internalization rates. PAQu again confirms these observations at the isoform level, finding decreases in EAAT2a protein levels. Perhaps the most impactful finding facilitated by PAQu is the estimation of significant differential abundance between C4A and C4B, which represent one of, if not the, strongest links between genetic risk for schizophrenia and synaptic connectivity impairments. Indeed, the largest signal for genetic association of common variation to schizophrenia emanates from the MHC region of chromosome 6 (28), and earlier research has shown that the signal can largely be explained by the risk variant’s impact on C4 isoforms, especially strong upregulation of C4A but not C4B (29). To understand what drives this association, we evaluated estimated differential expression of C4 transcripts and C4 peptides (Fig. 3). Remarkably, in our data transcripts for C4A and C4B are both estimated to be greater in dACC tissue from cases compared to controls (Fig. 3), contrary to expectation. This could be due to chance fluctuation, but it more likely is due to the challenge of differentiating the two transcripts from short-read sequencing—their expression levels are estimated by likelihood methods, they are not measured directly. Whatever the source of upregulation for C4B, transcript abundances cannot be driving the case–control difference for C4A and not C4B isoforms. To the contrary, the pattern of differential peptide abundances tells the tale (Fig. 3). Among the 38 peptides of C4 measured for our samples, 36 map to both isoforms and 2 map exclusively to C4B. Of the 36, 35 are significantly greater in cases than in controls, whereas both peptides exclusive to C4B show no case–control differences. Therefore PAQu overrides the information in transcript abundance, favoring the strong evidence from peptides that C4A is more abundant in cases and C4B is not. (A) Interval plot showing Welch t-statistics (SCZ vs. Control, n=53 vs. n=50) for two C4A/C4B transcripts, 36 shared peptides, and 2 C4B-unique peptides. Error bars represent 95% CIs. 35 out of 36 shared peptides and both transcripts show significant upregulation in SCZ, while the two C4B-unique peptides do not reach significance. (B) Distribution of PAQu estimated effect sizes D^ for C4A and C4B across 10 independent model runs. Boxes span the interquartile range, with the median indicated by the vertical line and whiskers extending to the minimum and maximum estimates. C4A shows consistently positive D^ across all runs (0.3 to 0.6), while C4B estimates straddle zero, showing that upregulation of C4B detected in panel (A) does not translate into a significant effect at protein isoform level as captured by PAQu. We next grouped isoforms on the basis of the number of times they are selected across MCMC chains. Then, for each level of minimum agreement factor γ among PAQu runs (Materials and Methods), we map protein isoforms back to their associated genes, and then perform Gene Ontology (GO) enrichment analysis (SI Appendix, Fig. S16). Notably, for more lenient thresholds for significance, synaptic and neuronal biology emerge from the gene set, whereas themes involving energetics (mitochondrial) and protein synthesis (ribosomal) are always overrepresented. PAQu}} Sheds Light on the Relationship Between Transcriptome, Proteome, and Peptidome. PAQu also provides estimates of the links between transcripts and protein isoforms. Using the same data as described above, we analyze the posterior distributions of the conversion coefficients W (Fig. 1). Interestingly, the distribution of the coefficients related to transcripts and protein isoforms, which are detected as significant in their respective domains, is shifted rightward (Fig. 4 A and B). This suggests that in the presence of biological variation in both the transcriptome and the proteome, statistical relationships arise naturally. We decided to investigate this further because the relationship between gene expression measured in tissue and the amount of corresponding protein found in the same tissue has been an intriguing and important scientific puzzle (30, 31). In some experimental settings, gene expression and protein abundance are highly correlated (32–34); in survey studies, the correlation is often low, including most studies evaluating brain tissue (35, 36). While our analyses bring the scale down to RNA transcripts and their protein isoforms, the parallels to previous work are obvious. Relationship of transcripts and protein isoform abundance within individuals. (A and B) Transcripts and protein isoforms that show biological variation (γ>5): (A) for all, (B) detected by the t test on transcripts and PAQu. (C and D) Per-isoform correlation of isoform and transcript abundances over individuals: (C) for all transcripts and their isoform abundances; and (D) for transcript and isoform abundance split by whether there is a unique or ambiguous mapping of peptides onto transcripts/isoforms. Of the 7,271 isoforms, 7,107 show limited variability between MCMC chains (Materials and Methods). For each of these reliably estimated isoforms, we computed the per-isoform correlation of isoform and transcript abundances over individuals. We obtain a roughly symmetric distribution centered at 0.067 but showing right skew (Fig. 4C). If the true transcript/isoform correlation were zero, then the observed distribution would be normally distributed about zero and only 0.2% would exceed |r|^>0.30, about 7 observations exceeding each bound. Instead, we found 417 transcripts/isoforms with correlations >0.3, 5.9% of the pairs. We also observed 19 transcripts/isoforms with unusually low correlations, below −0.3 (0.3%). Note that isoforms can be split into two distinct categories: isoforms with a unique mapping of peptides onto isoform (unique) and those for which the measured abundance of a peptide could trace to two or more isoforms (ambiguous). For the 4,056 unique transcript/isoform pairs (Fig. 4D), their distribution is centered at 0.087, with 316 (7.8%) in the positive tail and only 6 (0.1%) strongly negative, as previously defined. For the 3,051 in the ambiguous set (Fig. 4D), the distribution is centered at 0.041 and it has 101 (3.3%) in the positive tail and 13 (0.4%) in the negative tail. Thus, when there is greater uncertainty in how peptide abundances inform isoform abundances, PAQu is somewhat less accurate, leading to fewer strongly positive correlations. Next, we focus on extreme positive correlations. As we develop in SI Appendix, Table S2, these correlations must arise from substantial biological variation among samples for both transcripts and isoforms. For example, if cases and controls differed in their means by one half of a SD for both transcript and corresponding isoform—an unusually large difference—the expected correlation induced would only be r≈0.053. To determine what variation predicts the observed correlation for each (isoform, transcript) pair, we build a linear model to predict the estimated correlation using the following five predictors: mean case–control difference for isoform or transcript abundances; coefficient for the effect of age on isoform or transcript abundances; and impact of genetic variation on the gene. For this last predictor, we use results from GTEx (37) analysis of gene expression. Specifically, we use the variation of gene expression explained by the eGENE genetic variant (Materials and Methods). Using these predictors, we fit the model for two classes of outcomes: The 412 transcript/isoform pairs with strong positive correlation values, r>0.3 (5 others did not map uniquely to a single gene); and 412 randomly chosen transcript/isoform pairs, conditional on their correlation being in the range of |r|<0.05. Using all 824 outcomes, the model has low predictive value, R2=0.061 (SI Appendix, Table S2), but the impact of age on transcript abundance, diagnosis-based mean isoform differences, and eGENE variation explain a portion of the variation. If the low correlation set is analyzed, no predictor explains a significant amount of the variation; when the strong positive correlation set is analyzed, age and diagnosis for both transcripts and isoforms are predictors, whereas genetic variation is not. The model accounts for a slightly larger portion of the variability of these higher correlation outcomes (R2=0.092); SI Appendix, Table S2). The distribution of correlations (Fig. 4 C and D) and our modeling results suggest biological variation among samples is limited, which is consistent with the literature on adult human biology. Human variation in gene (24, 38) and protein (39) expression is modest, regardless of genetic, disease, or environmental origin, even as adults age (40). We noted a complementary feature of the genes encoding transcripts/isoforms in the high correlation set, namely they were highly enriched for terms related to neuronal and synaptic biology. We used ReviGO (41) to summarize these terms into coherent themes (SI Appendix, Fig. S17). This result is consistent with prior studies that show tight scaling of transcription and translation based on synaptic activity (42, 43), although this tight coupling potentially diminishes with increasing age (44, 45). PAQu also returns estimates of the detectability matrix Z, which projects peptide abundances onto isoform abundances. Peptides vary in their detectability by the mass spectrometer and spectral matching algorithms for a host of reasons, such as their amino acid composition and how those amino acids are chemically altered in vivo, i.e., posttranslational modifications (PTMs). For the latter, while the peptide itself has the same amino acid sequence, PTMs change its mass/charge signature and thus the detectability as the same peptide (46). In turn, PTMs can induce variability in measured abundances of peptides from the same isoform and their relationship to the isoform. Discussion Advanced transcriptomics approaches facilitate quantification of transcript alterations in disorders like schizophrenia that could contribute to disease pathology. However, confirming these observations at the isoform level has proven challenging due to bottlenecks in proteomic data acquisition and analysis. PAQu stands as an effective tool for addressing several of these challenges: It provides uncertainty quantification (via highest posterior density credible intervals) for all estimated parameters, integrates external covariate information, elucidates the relationship between the transcriptome and the proteome, highlights the potential presence of posttranslational modifications, and allows users to perform differential abundance analysis. Our analyses of transcriptome and proteome from postmortem dorsal anterior cingulate cortex (dACC) gray matter reveal close integration of certain neuronal and synaptic transcripts and isoforms (Fig. 4 and SI Appendix, Fig. S17). Meaningful biological variation, including changes in abundance with aging, based on case-status, and genetic variation determine some of this integration (SI Appendix, Table S3). For instance, among the 412 highly correlated transcripts and isoforms (r>0.3), 47 are classified under the term regulation of trans-synaptic signaling (GO:0099177), and these show the greatest enrichment for Gene Ontology’s Biological Processes. Our results on case–control differences for transcripts and isoforms confirm a critical observation made in schizophrenia, first published by Sekar et al. (29) 10 years ago. Elevated levels of the Complement Component 4 transcript C4A, but not C4B, are a genetically based risk factor for schizophrenia, directly correlated to structurally diverse risk alleles. This increase in C4A is widely believed to contribute to impaired cognitive and sensory processing in schizophrenia via its promotion of dendritic spine pruning (47). As these isoforms are almost identical, antibodies cannot distinguish between the two, and thus the findings could not be confirmed at the protein level. Our data capture 38 peptides of C4, 36 mapping to both isoforms and 2 mapping exclusively to C4B. Our results show that both C4A transcript and protein isoform levels are significantly increased in schizophrenia. By contrast, the abundance of C4B did not change with case status. We find that several other transcript alterations with relevance for glial and synaptic processes, implicated together in schizophrenia, are realized at the isoform level. For example, CD44 isoform P16070 is also significantly elevated in cases, which is consistent with recent results showing upregulation of reactive astrocyte marker CD44 in the presence of elevated levels of C4A (48). However, not all synaptic transcript changes result in concurrent isoform alterations, as is the case in CASKIN2, a key molecule for excitatory synaptic function (49). Levels of the CASKIN2 transcript ENST00000321617 are decreased whereas its isoform, Q8WXE0, is significantly elevated, possibly due to a compensatory mechanism to maintain synaptic transmission. Collectively, our findings in schizophrenia demonstrate the power of PAQu to provide confirmation that transcript level alterations manifest at the protein level, providing a direct link from genetic risk to protein function while also showing that not all such changes are translated. PAQu has several limitations. First, it can currently only work in scenarios when q≤r. Second, PAQu performs differential isoform expression (DIE) testing, identifying isoforms whose abundance differs between conditions. DIE can reflect either changes in overall gene expression or changes in relative isoform usage; therefore, significant findings should be interpreted as evidence of altered isoform abundance rather than direct evidence of altered splicing. Third, as is typical for large-scale proteomic studies, the real-data schizophrenia analysis lacks a ground-truth measurement of protein isoform abundance, making it impossible to directly assess the accuracy of inferred isoform abundances or differential effects in vivo. Future work could address these limitations through more scalable inference procedures and explicit modeling of posttranslational modifications and isoform usage. In conclusion, in this paper we propose PAQu, a powerful framework for the analysis of mass spectrometry data that takes advantage of additional information about the transcriptome to quantify the abundance of protein isoforms. By implementing a version of Bayesian supervised factor analysis, we believe that PAQu has the potential to help researchers make full use of mass spectrometry data. Materials and Methods Notation and PAQu}} Model. We define the n×q matrix of transcript expression data as T, with Tij representing its (i,j)-th element. Similarly, the n×r matrix of peptide abundance data is denoted by P, where Pik is the (i,k)-th element. The vector A is an n×1 binary indicator describing the condition of sample i, with Ai∈{0,1}. The indices i=1,⋯,n, j=1,⋯,q, and k=1,⋯,r refer to the samples, isoforms, and peptides, respectively. Our objective is to infer the parameters that link these quantities, as described by the following model:[1]I=I0+TW+ADT+EI,P=IZ+EP. Here, I is an n×q matrix where Iij represents the unobserved abundance of isoform j in sample i, and I0 is an n×q matrix of intercepts, with Iij0 indicating the baseline abundance of isoform j across samples (Iij0 is constant for every i). The diagonal matrix W has elements Wjj, representing the conversion weight between transcript j and its corresponding protein isoform. The vector D is q×1, where each Dj describes the effect of the binary condition A on the abundance of protein isoform j. The matrix Z is a q×r matrix of detectability coefficients, with Zjk=0 if peptide k is not associated with isoform j, and otherwise unknown. The terms EI and EP are random heteroskedastic Gaussian noise terms affecting isoform and peptide abundances, respectively. Note that knowledge of T and A is not strictly necessary: PAQu can operate as long as P and a detectability mask are provided. We include here these additional components in the model for the sake of completeness. In SI Appendix, section 1, we provide details and in SI Appendix, section 2, we describe how to use PAQu in cases where information about T and A is limited or absent, or in the presence of external covariates. Estimation and Inference with PAQu}}. Prior distributions. The parameters in the model described in Eq. 1 are not identifiable. Thus, standard optimization procedures can lead only to locally optimal solutions. A viable alternative is provided by the Bayesian framework, which requires specification of prior distributions on parameters. We therefore assume that any conversion weight between a transcript j and its protein isoform follows a Gaussian distribution:[2]Wjj∼N1,τjW,j∈{1,⋯,q}, where τjW denotes the prior variance. The prior mean is set to 1, which follows the theoretical assumption of perfect proportionality between transcript expression levels and corresponding protein isoform abundances. We impose a similar Gaussian prior on condition effects D:[3]Dj∼N0,τjD,j∈{1,⋯,q}, where τjD denotes the prior variance, and the prior mean is set to 0 to reflect conservative prior knowledge about potential differential isoform abundances. The prior distribution for detectability scores depends on the compatibility between protein isoforms and peptides. We write k≺j if a peptide k is compatible with an isoform j. Therefore, we can write the prior distribution as a mixture between a truncated Gaussian and a Dirac mass at 0: [4]Zjk∼1{k≺j}N¯l¯u¯1,τjZ+(1−1{k≺j})δ0,j∈{1,⋯,q},k∈{1,⋯,r}, where N¯ denotes a truncated Normal distribution with prior mean 1, prior variance τjZ, and lower and upper truncation parameters l¯=P¯jmin/P¯jmax and u¯=P¯jmax/P¯jmin with P¯jmin=mink≺jn−1∑i=1nPik and P¯jmax=maxk≺jn−1∑i=1nPik. We employ this truncated prior specification for Z to guarantee that estimates in I are consistent with observed values at the peptide level. We also specify a Gaussian prior for the intercept I0:[5]Ij0∼N0,τjI0,j∈{1,⋯,q}, where τjI0 denotes the prior variance. Notice that in an experiment where transcript expression levels have not been measured, but the scientist has access to average expression levels from another study, these may be used as prior means for the intercept (see SI Appendix, section 2 for details). As an alternative to the Gaussian prior, PAQu offers the possibility of implementing a spike-and-slab prior. For example, the spike-and-slab prior for D reads as[6]Dj∼θjN0,τjD+(1−θj)δ0,j∈{1,⋯,q}, where θj is the prior probability that a given isoform j is differentially abundant across conditions. As supported by our simulation study, we prefer to use Gaussian priors as they lead to more sensible results. The last fundamental prior distribution, which links the first and second layers (Fig. 1), is the one for I. Again, we assume a Gaussian prior, this time informing it with knowledge about the other covariates:[7]Iij∼NIj0+TijWjj+AiDj,1,i∈{1,⋯,n},j∈{1,⋯,q}. Intuitively, the prior distribution in the previous Equation guarantees that the protein isoforms approximately lie in the space spanned by transcripts. This allows PAQu to supervise the factor analysis of peptides in the second layer in a smooth and unified manner. The prior distributions of the other parameters in the model are specified in SI Appendix, section 1. Posterior sampling. We approximate the posterior distribution of the model parameters using Gibbs sampling, which is a natural choice since we can derive closed-form expressions for the full conditionals. Intuitively, given the values of (W,D), I, or Z, the remaining unknown parameters can be easily determined by linear regression. This approach simplifies the complexity of estimating a multivariate posterior distribution by breaking it down into several manageable univariate problems. The analytical posterior distributions of all parameters in the model—using both Gaussian and spike-and-slab priors—are specified in SI Appendix, section 1. Our choice of default parameters for the Gibbs sampling procedure includes 3,000 MCMC iterations and a burn-in period of 2,000 iterations (50). As a result, each MCMC chain approximates its corresponding conditional posterior distribution using 1,000 samples. Measuring significance. Although posterior distributions contain all the information about the parameters in the model, we use the Local False Sign Rate (LFSR) proposed by ref. 23 to summarize them in a compact and interpretable measure for hypothesis testing. Intuitively, the LFSR measures confidence in the sign of the effect. We are particularly interested in evaluating when the effect Dj of a condition A on an isoform j is significantly different from 0. We thus compute the LFSR as [8]LFSR(Dj)=minPDj≥0|·,PDj≤0|·, where PDj≥0|· and PDj≤0|· denote the posterior probabilities that Dj is greater or smaller than 0, respectively. We similarly define the LFSR for the other parameters. Importantly, the LFSR serves as a Bayesian counterpart to the local false discovery rate, but it operates strictly within a range of [0,0.5]. Because it takes the minimum of the two directional posterior probabilities, a value of 0 indicates absolute certainty in the sign of the effect, while a value of 0.5 represents complete ambiguity. While classical FDR controls for the probability that a null hypothesis is falsely rejected, the LFSR strictly bounds the probability of assigning an incorrect directional sign to an effect. Consequently, evaluating PAQu at an LFSR threshold of 0.05 introduces a higher statistical bar than using an uncorrected nominal P-value threshold of 0.05 (as used in the benchmark t tests in Fig. 2). Simulation Study. We simulate transcript, protein isoform, and peptide data under the two-layer linear model in Eq. 1. We simulate data in two different settings. In the first easy setting, we generate data without shared peptides. Specifically, we sample one transcript (q=1) from a Normal distribution (i.e. Tij∼N3,1), map it into a protein isoform Iij with a random conversion weight (Wjj∼N1,0.5), add an effect due to a binary condition (Dj∈{0.33,0.66,1}), a random intercept term (Iij0∼N10,0.5, constant for all i) and random noise (EijI∼N0,1), and then split it into two peptides (r=2) using randomly assigned detectability scores (Zjk∼N¯−0.50.51,1), again adding random noise (EikP∼N0,1). We define Ai=1 for the first half of the sample, i.e. for i=1,⋯,n/2, and Ai=0 otherwise. In symbols:[9]Iij=Iij0+TijWjj+AiDj+EijI,i∈{1,⋯,n},j=q=1,Pik=IijZjk+EikP,i∈{1,⋯,n},k∈{1,r=2}. In the second hard setting, we generate data that include shared peptides. Here, we sample five transcripts (q=5) from a multivariate Normal distribution with identity covariance matrix (i.e. Tij∼N3,1), map them to their corresponding protein isoforms Iij with random conversion weights (Wjj∼N1,0.5), add an intercept term (Iij0∼N10,0.5), and introduce a binary condition effect (Dj∈{0.33,0.66,1}) to the first |Dact|∈{1,2,3} isoforms, i.e. Dj=Dj1{j≤|Dact|}. The isoforms are then split into 10 peptides (r=10) using random detectability scores, with s=30% of the detectability matrix entries being nonzero, i.e. Zjk∼sN¯−0.50.51,1+(1−s)δ0. As before, we define Ai=1 for the first half of the sample, i.e. for i=1,⋯,n/2, and Ai=0 otherwise. In symbols:[10]Iij=Iij0+TijWjj+AiDj+EijI,i∈{1,⋯,n},j∈{1,⋯,q=5},[11]Pik=∑j=1qIijZjk+EikP,i∈{1,⋯,n},k∈{1,⋯,r=10}. We run each simulation experiment with three different sample sizes (n=100,200,500), and with 25 different seeds. Relationship to Other Methodology. We want to point out that, from a technical standpoint, our approach shares some similarities with refs. 50 and 51, as both employ a variant of Bayesian factor analysis. However, our method differs significantly in several key aspects. Specifically, we adopt different prior distributions, implement a distinct sampling algorithm, and target a different application. These differences result in a more tailored and efficient solution to the problem of protein isoform quantification, setting PAQu apart from previous approaches. Moreover, by developing a tool for supervised factor analysis, our work has some connections to refs. 52–54. However, these works only provide inference for latent factors (in our case, protein isoforms). Instead, PAQu provides estimates for a full class of parameters, shedding light on the biological processes behind protein formation. Sample Preparation, LC-MS/MS Measurement, and Database Search for Proteins. See SI Appendix, section 4, for full details. In brief, brain samples were obtained from the University of Pittsburgh brain bank. A tissue sample of similar size was collected from each brain. Each sample was homogenized, its proteins were then solubilized, digested into peptides, TMT labeled, fractionated, and analyzed with an Orbitrap Tribrid Eclipse mass spectrometer. Samples were processed in batches of 14 samples, which will be described as a plex. RNA transcripts observed in the same brains (24), as measured by RNA-seq, were used to inform which isoforms were likely to be found in these samples (isoform matching). Data Preprocessing. Peptide abundance data. We first normalize raw expression values of samples to account for differences in total expression levels across samples (see SI Appendix, section 4, for details). After log2-transforming normalized values, effects of batch (i.e., plex), diagnosis, age, and postmortem interval are modeled using ordinary least squares (OLS) regression. Effects of plex were subtracted from the data, leaving residuals that capture biological variation. Transcript expression data. Transcript expression data were produced by the CommonMind Consortium (24) and were downloaded from https://www.synapse.org, specifically syn29442530, which was accessed on February 4, 2025. These data have been moved to the NIMH Data Archive (NDA Collection C5032). We first normalize transcript expression counts, transform them to counts per million, set them on a log2 scale, and remove the effect of technical covariates (see SI Appendix, section 4, for details). PAQu}} Analysis of Schizophrenia Dataset. We exploit the block-diagonal structure in Z to divide the original problem into 5,241 independent blocks. We then combined isoforms represented by the same set of peptides into isoform groups. This reduces the number of isoforms from 9,169 to 7,271. Of these, 5,831 are singleton isoforms while 3,338 isoforms were grouped into 1,440 isoform groups with 2 or more isoforms. Out of the 1,440 isoform groups, 1,396 represented isoforms from a single gene; the remaining 44 represent multiple genes. From here on, we will refer to the combined set of singleton isoforms and isoform groups as isoforms. Analyzing this dataset using 10 independent MCMC chains requires an estimated wall time of approximately 1 h per chain on 32 cores, yielding a total analysis time of roughly 10 h (SI Appendix, Fig. S18). To quantify the stability and robustness of our differential abundance findings across independent simulations, we define the selection agreement factor, γ, for each isoform. This factor is a discrete frequency count representing the total number of independent MCMC chains in which a given isoform was identified as significantly differentially abundant at local false sign rate threshold of 0.05. Because our study utilizes a total of 10 independent MCMC chains, the agreement factor γ ranges from a minimum value of 0 to a maximum value of 10. An agreement factor of 0 indicates that the isoform was never selected as differentially abundant in any run, while a value of 10 indicates absolute consensus across all parallel chains. GO Enrichment Analysis. We perform GO enrichment analysis using the g:GOSt tool for functional profiling contained in the g:profiler toolkit (55), version e111_eg58_p18_f463989d, with Benjamini–Hochberg significance threshold set at 0.05. All ENSG isoforms employed in the analysis are treated as background; isoforms with LFSR≤0.05 are treated as foreground. Analysis of Transcript and Corresponding Isoform Correlations. Of the 7,271 isoforms, we selected for analysis those isoforms with low variability of estimated abundance across MCMC chains. Specifically, for each isoform and MCMC chain, we computed the mean abundance over 103 samples. Next, we computed the variance of the mean over the 10 MCMC chains. If the variance fell below one, the isoform was retained (N=7,107); otherwise, it was removed (N=164). We then calculated the Pearson correlation, over samples, for each of the 7,107 transcript/isoform pairs. To predict the observed correlation of transcript/isoform pairs, we estimated the effects for diagnosis and age from the estimated isoform abundances and adjusted transcript expression. It is well known that gene expression and protein abundance can vary by genotype. To quantify the impact of genetic variation, we gathered information on 26,134 eGenes for the cortex from the GTEx website (56). (By definition, an eGene is the genetic variant that explains the greatest variance in gene expression.) From this set, we selected 16,491 protein-coding genes and recorded the estimate of the slope (b) and allele frequency (f). The estimated variance explained by each eGene was then determined as 2f(1−f)b2, and this value was used to predict the observed correlation of transcript/isoform pairs. Transcript/isoform groups mapping to multiple genes were excluded from this analysis. Supplementary Material Appendix 01 (PDF) Dataset S01 (CSV)