Leveraging protein language models to identify complex trait associations with previously inaccessible classes of functional rare variants

Summary

Protein language models (PLMs) improve variant effect predictions, but their role in gene discovery for complex traits remains unclear. We introduce an allelic series-based regression test that uses PLM-derived variant effect predictions as proxies for effect sizes, identifying ∼46% more associations than standard burden tests. Extending this to isoform-level analysis, we find 26 gene-trait pairs with stronger associations in non-canonical versus canonical transcripts, highlighting isoform-specific effects. Finally, we identify evolutionary plausible variants (EPVs), missense variants assigned higher likelihoods than the wild-type alleles by PLMs, representing 0.45% of missense variants. EPVs show higher allele frequencies than synonymous variants, consistent with differential selection pressures, and are linked to nine traits, including protective associations with low-density lipoprotein (LDL) and bone mineral density. Together, our results demonstrate how PLMs can enhance rare-variant interpretation and gene-trait association discovery in exome data.

Graphical abstract

Highlights

• •Gene-based tests with PLMs find ∼50% more associations than burden tests
• •Isoform-aware variant predictions capture transcript-specific associations
• •Novel associations are found from mutations predicted more likely than wild type

Jang et al. demonstrate that protein language models expand coding-variant discovery in complex traits by improving gene-based tests and revealing transcript-level associations as well as mutations that are predicted to be evolutionarily more likely than the wild type.

Introduction

Two challenges confront the use of exome sequencing to identify genes influencing complex traits: first, variants in coding sequence are mostly rare, making it hard to test their phenotypic associations due to lack of power. Second, the impact of variants on gene structure and function, and hence on phenotypes, is often hard to predict. The release of large population-scale whole-exome sequencing (WES) data with sample sizes approaching 1 million individuals¹ is addressing the first challenge, while the deployment of algorithms for variant effect prediction (VEP) to estimate the consequences of protein sequence changes helps address the second by prioritizing risk variants.

A recently introduced class of VEP uses protein language models (PLMs) and has outperformed existing VEPs in distinguishing pathogenic from missense variants in experimental and curated datasets of mutations.^2,3,4,5 PLMs also potentially offer more flexible modeling of protein isoforms compared to methods that require multiple sequence alignments such as PolyPhen-2,⁶ SIFT,⁷ and EVE.⁸ They provide functional impact metrics across the full mutational spectrum—from variants less likely than the wild type to those more plausible from an evolutionary perspective. However, their use as tool for detecting protein sequence variants that contribute to complex disease remains under-explored. In this study, we present novel frameworks to exploit the unique properties of PLMs for the analysis of exome sequencing and provide real-world benchmarks for the state-of-the art PLMs.

First, we propose an isoform-wide gene-based test using the variant sets specific to isoforms together with the scores PLMs produce for the same variants in different isoforms (isoform-specific VEPs). The use of transcripts in most WES analyses to date is limited. Annotations typically rely on a single transcript, most often the canonical transcript, and multiple transcripts are only considered when determining which variants are potentially pathogenic.^9,10,11 Yet isoform usage is widespread and involved in disease: alternative splicing is estimated to affect approximately 90% of the multi-exon human genes,^12,13,14 and many disease associations with spliced protein isoforms have been reported.^{15,16,17,18,19} PLMs infer the likelihood of each residue based on the whole protein sequence, and their VEPs may be sensitive to the isoform context and can be used to interrogate complex trait association. Using ESM1b, Brandes et al.² detected about 1.8 million isoform-sensitive variants across 9,000 genes, a majority of which lie near the splice-disrupted domains. For example, removal of five amino acids from the major isoform of MEN1 protein (O00255-1) changes predicted protein structure and the variants near this excision site are now predicted to have a greater deleterious impact in the alternative isoform (O00255-2).

Second, we use PLMs to identify missense variants predicted to be more likely to appear in the protein than the wild-type variant and test for their associations with complex traits. Historically, VEPs focus on distinguishing pathogenic from benign variants where the evaluation of pathogenicity incorporates the extent of depletion of genetic variations in a given region.^20,21,22 PLMs offer a full spectrum of variant-impact prediction, from variants highly depleted to those more likely to be observed than the wild type in protein sequences. Recent work has termed these evolutionary plausible variants (EPVs) and demonstrated that they often alter protein function in a range of settings.²³ Investigating EPVs may offer insights into fitness, evolution, and phenotypic associations missed in the previous efforts. So far, it is unknown how often this class of variants is observed in humans and how they relate to various human phenotypes.

Finally, we introduce a gene-based test using a regression framework (PLM-R), which serves as a basis method for the isoform and EPV analyses. PLM-R relates PLM-based VEPs of missense variants to the phenotypes of the carriers in a regression framework and was motivated by the evidence that VEP scores are correlated with variant effect sizes within a gene.^4,24 This naturally corresponds to an allelic series model in which a spectrum of variants exist in a gene with graded or distinct phenotypic consequences.²⁵ We examine how this approach compares with existing approaches, including burden tests with and without VEP-based variant filtering and SKAT-O.²⁶

We apply our methods to 82 traits in the UK Biobank (UKB), with replication in All of Us (AoU) when phenotypes are available. We show that PLM-R provides a simple framework that does not require arbitrary cutoffs, has increased power over standard burden test with and without variant filtering by VEPs, and augments discoveries from SKAT-O. Following this, we show that isoform-wide PLM-R identifies more discoveries than the same test using only canonical transcripts and that multiple gene-trait pairs show significantly higher effect sizes for an alternative transcript than the canonical transcript. Finally, we found that testing EPVs within a gene can lead to identification of novel gene-trait associations. Our replication analysis suggests that discoveries from isoform-wide and EPV analyses are generally replicable and connected to trait biology. We conclude that the proposed PLM tests will lead to novel gene-trait associations and biological insights, especially associations with previously ignored classes of variants and isoforms.

Results

Use of VEPs in burden testing

Burden tests are widely used to detect gene-trait associations from WES data.²⁷ It is not clear whether using VEPs to filter potentially pathogenic variants increases the power of burden tests and how different VEPs compare in their performance. To answer this question, we first conducted a burden test with a set of likely damaging variants defined by various VEPs with categorical predictions (Mutation Taster, LRT, PolyPhen-HumVar, PolyPhen-HumDiv, SIFT, ESM1b, and Alpha Missense) (see STAR Methods). In all approaches, we used rare variants (allele frequency <1%) and applied minor allele frequency (MAF) weights (i.e., beta (1,2)). We used 82 phenotypes (43 quantitative, 39 binary) encompassing various diseases, biomarkers, and behavioral phenotypes in ∼150,000 UKB exomes (see STAR Methods and Table S1 for detailed description of the phenotypes). Compared to 203 discoveries from using all missense variants without VEP-based variant filtering, VEP-based variant filtering yielded a similar number or fewer discoveries with SIFT yielding the greatest number of associations (n = 210) followed by ESM1b (n = 205) (Figure 1A; Tables S2 and S3). These results suggest that the advantages in discovery power by applying VEPs for variant filtering in burden test is small (<5%) or inconsistent.

Figure 1.

Number of gene-trait associations discovered from rare-variant gene-based test in UKB

(A) The number of significant associations from the burden test using all rare (allele frequency <1%) missense variants versus pathogenic missense variants filtered by various VEPs.

(B) The number of associations from the PLM-R approach using different PLM-based VEPs.

(C) The number of unique and overlapping associations from burden, SKAT-O, and PLM-R ACAT.

All associations reported here are p < 2.78e−6.

PLM-R improves gene discovery

We next present a regression framework for leveraging PLM-based VEPs in exome sequencing studies. Specifically, we regress the phenotypic values of the carriers onto the PLM-based VEP scores of the rare missense variants that each individual carries (i.e., PLM-R, hereafter; see STAR Methods). While a burden test usually aggregates putative damaging variants in one group and the variants are typically weighted by their MAFs, PLM-R uses all available missense variants and assumes that the effect sizes of the variants are approximated by VEP scores, as recently suggested.^24,28 Its power increases when multiple variants exist whose effect sizes align with VEP scores, and thus it is also suited for detecting gene-trait pairs within allelic series. Here, we compare the number of associations from PLM-R using VEP scores trained by recent PLM models, including multiple ESM models (ESM1b, ESM2, and ESM3 with and without structural information during inference) as well as Alpha Missense and PrimateAI-3D. We also tested CPT-1, a transfer learning paradigm based on deep mutational scan data, and CADD v.1.7, which includes features from PLMs, in PLM-R. We apply these methods to 82 phenotypes in ∼150,000 unrelated European (EUR) ancestry individuals in UKB whole-exome sequence data.

Across 82 UKB phenotypes, we found 195–248 significant gene-trait associations from PLM-R depending on specific PLMs used, after multiple testing adjustment (STAR Methods and Figures S1 and S2). VEP from the ESM2-650M model showed the largest number of associations (n = 248), followed by Alpha Missense (n = 244), CPT1 (n = 242), ESM3/ESM1-b (n = 240), and CADD (n = 228). Aggregating the results from PLM-R with different VEPs resulted in 296 discoveries in total, representing a 47% increase in the number of associations compared to that from standard burden test (Figure 1B; Table S4).

We also performed SKAT-O with standard MAF weights (beta (1,25)) on the same dataset. All missense variants, without VEP-based variant filtering, were included to maximize the power, which allows varying effect sizes and directions within a gene, and we found 308 associations (Table S5). This indicates that more associations can be found by relaxing assumptions on the effect sizes and direction of the variants within a gene. Interestingly, discoveries made by PLM-R and SKAT-O only partially overlapped: 167 associations were unique to either PLM-R or SKAT-O, suggesting that these methods tap into different sets of associations (see Figure 1C).

PLM-R+ for incorporating pLoF and splice variants

While PLMs provide effect predictions for missense variants, we assumed that incorporating putative loss of function (pLoF) and splice variants can detect more gene-trait associations. To test this idea, we extended the current regression framework to pLoF and splice variants by assigning the maximum PLM VEP score (i.e., most deleterious score) to the former and the top 33% most deleterious score to the splice variants, based on recent analysis of splice effect sizes.⁴ We computed the omnibus p values between PLM-R (using only missense variant) and the regression using pLoF, splice, and missense variants and refer to this approach as PLM-R+ (see STAR Methods). Since PLM-R+ assumes non-null effects of the pLoF and splice variants, its underlying model is that of an allelic series encompassing different classes of coding variants. Here, we used ESM1b and Alpha Missense for the demonstration and application of PLM-R+. We found a total of 265 and 279 associations from PLM-R+ for ESM1b and Alpha Missense, respectively, compared to PLM-R using missense variants alone (240 and 244 associations for each VEP, respectively). This indicates that incorporating other classes of coding variants in the PLM-R framework can detect further associations with allelic series.

We note that PLM-R and PLM-R+ can be used with many PLMs, and we anticipate performance will improve as PLMs improve. Overall, applying PLM-based VEPs in our regression framework complements popular methods, including burden tests and SKAT-O, and enhances gene discovery from rare coding variants, especially gene-trait associations with allelic series. Indeed, we found associations not reported in a recent full-scale UKB WES analysis that used burden, SKAT, and SKAT-O, despite our smaller discovery sample size. Significant associations were found for PAM and diabetes, CREB3 and melanoma, SAMHD1 and breast cancer, SPG11 and educational attainment, OTOP1 and birth weight, ANGPTL4 and triglycerides, HMCN1 and FEV1/FVC ratio, SF3B1 and red blood cell count, and OMA1 and platelet count (Table S6).

Isoform-wide PLM-R to detect complex trait associations and their isoform contexts

Despite rich catalogs of transcripts and research showing its involvement in gene regulation, isoforms are typically underutilized in rare-variant association studies.^9,11 Here, we integrated isoform-specific annotations and VEPs in our PLM-R framework and conducted isoform-level associations for 82 phenotypes in ∼150,000 unrelated European ancestry individuals in UKB with replication in the remaining UKB (n = ∼190,000) and AoU (n = from 24,293 to 104,939) samples. For this analysis, we used ESM1b scores, which released VEP scores trained in a wide array of protein isoforms.

We first removed transcripts for which there was low evidence, defined as transcript support level (TSL) less than three in the ENSEMBL (v.109) database, resulting in 57,368 transcripts across 17,959 genes. Of these, 12,918 genes had more than two transcripts giving a total of 52,725 transcripts. After adjusting for multiple testing by the effective number of tests (n_effective = 28,882; see STAR Methods), we found 566 significant transcript-trait associations, which amounts to 256 associations (234 quantitative, 22 binary traits) at the gene level, higher than the number of discoveries using canonical transcript only (n = 240). Of note, 24 of these associations were only detected from non-canonical transcripts, including SPG11-202 and education, CLPTM1-205 and C-reactive protein and low-density lipoprotein (LDL), MAP3K11-210 and urate, and SMIM29-208 and height.

We conducted post hoc analysis of whether a given gene-trait pair has associations of different strength across isoforms. To do this, we calculated the difference of the standardized effect sizes from canonical and non-canonical transcripts for the gene-traits that showed at least one significant transcript-level association and had more than two transcripts (175 gene-trait pairs), and we computed p values for the difference between effect sizes using permutation (see STAR Methods). A total of 33 isoform-trait pairs (26 gene-trait pairs) showed higher effect sizes in non-canonical than canonical transcripts (see Table 1; Table S7). For example, we found higher effect size from non-canonical transcripts for JAML and leukocyte count (JAML-202, beta = 0.01, SE = 0.005 versus JAML-204, beta = 0.03, SE = 0.005), GCK and glucose (GCK-204, beta = 0.045, SE = 0.006 versus GCK-202, beta = 0.053, SE = 0.006), and platelet count and SH2B3 (SH2B3-201, beta = 0.019, SE = 0.001 versus SH2B3-202, beta = 0.037, SE = 0.002).

Table 1.

Top 20 gene-trait pairs showing stronger associations in non-canonical transcripts

	–	$\beta$ (SE)		Log10 p value
Phenotype	Gene	Canonical	Non-canonical	Canonical	Non-canonical
LDL_custom	CLPTM1	0.003 (0.003)	0.029 (0.005)	0.45	9.57
C_reactive_protein	CLPTM1	0.006 (0.003)	−0.035 (0.005)	1.20	13.31
neutrophil_count	JAML	0.007 (0.005)	0.027 (0.005)	0.91	8.12
LDL_custom	APOB	−0.002 (0.002)	−0.026 (0.005)	0.82	5.90
leukocyte_count	JAML	0.009 (0.005)	0.031 (0.005)	1.35	10.77
Glucose	AGBL5	0.005 (0.003)	0.024 (0.005)	1.29	6.27
Height	SMIM29	0.007 (0.003)	0.029 (0.006)	1.29	5.94
eosinophill_count	SH2B3	0.003 (0.001)	0.012 (0.002)	1.76	8.23
Urate	MAP3K11	−0.010 (0.003)	−0.029 (0.005)	3.62	9.66
IGF_1	PARPBP	−0.032 (0.005)	−0.076 (0.008)	8.83	20.50
monocyte_count	GPSM3	0.021 (0.006)	0.039 (0.008)	3.30	6.50
age_menopause	HAGHL	−0.032 (0.009)	−0.059 (0.012)	3.37	6.06
leukocyte_count	AGER	0.009 (0.002)	0.013 (0.002)	4.17	7.38
Platelet_count	SH2B3	0.019 (0.001)	0.037 (0.002)	37.67	69.48
HDL_custom	SH2B3	−0.008 (0.001)	−0.014 (0.002)	6.94	11.40
red_blood_cell_count	FCGRT	0.012 (0.003)	0.017 (0.003)	3.89	6.20
leukocyte_count	SH2B3	0.009 (0.001)	0.016 (0.002)	8.88	13.59
Lipoprotein A	MAP3K4	0.019 (0.003)	0.024 (0.003)	9.55	13.21
Glucose	GCK	0.045 (0.006)	0.053 (0.006)	13.65	17.47
monocyte_count	CCR2	−0.016 (0.004)	−0.019 (0.004)	5.00	6.16

We present gene-trait pairs showing higher effect sizes in non-canonical transcript in the discovery sample. Effect sizes (β), standard errors (SE), and p values are presented. The list is ordered by the magnitude of the percentage increase in the absolute effect sizes. $\beta$ here is non-standardized (i.e., raw ESM1b scores were used). When there are multiple transcript pairs for a given gene trait, only the non-canonical transcript with the largest increase in the effect size is presented. A full list with more detailed information, such as transcript IDs, can be found in Table S7.

Among the 33 isoform-level associations, 30 of them were replicated after multiple testing adjustment (p < 0.05/33) in a held-out UKB WES sample, with 18 of them showing significantly higher effect sizes in the non-canonical than in the canonical transcript with the same direction of effect as in the initial discoveries. We also tested 18 isoform-level associations for a subset of phenotypes existing in AoU (i.e., height, glucose, LDL, HDL, triglycerides, platelet/leukocyte/red blood cell/eosinophil/monocyte counts, and alkaline phosphatase) and found that nine of them showed significant associations after multiple testing adjustment (14 at nominal significance level). Two showed significant differential association between canonical and non-canonical transcripts, including GCK-204 and GCK-202 for glucose (beta = 0.015, SE = 0.004 versus beta = 0.016, SE = 0.004). These results can be found in Table S7.

Sources of differential isoform associations

We interrogate two potential sources for differential associations between isoforms: the inclusion of different sets of variants and isoform-specific VEP scores despite similar variant sets. First, we categorized variants into those existing in both canonical and non-canonical transcripts and those unique to each transcript. We then conducted burden tests and SKAT-O separately on these variant sets to isolate these two sources (see Tables S8 and S9). In most cases (e.g., 30 out of 33 associations), variants in overlapping regions showed higher test statistics than variants unique to the canonical or non-canonical transcripts (see Figures S4 and S6). For example, in SKAT-O, 119 variants lying in the overlapping region of the two transcripts of MAP3K11 yielded strong association with urate (${\chi }^2$ = 94,9, p = 1.99e−22), while 166 variants specific to canonical transcript showed virtually no signals (${\chi }^2$ = 0.16, p = 0.69). Similarly, 41 variants in overlapping regions of the transcripts of PARPBP showed strong association with insulin growth factor (IGF)-1 (${\chi }^2$ = 113.87, p = 1.39e−26), while 132 variants specific to canonical transcripts showed little association (${\chi }^2$ = 3.19, p = 0.07). In some cases, we found strong association from variants unique to non-canonical transcripts but not from variants in overlapping regions or in regions unique to the canonical transcript. This includes the association between CLPTM1 and C-reactive protein, and LDL, and the association between SMIM29 and height (see Figures S5 and S7). Overall, the results suggest that incorporating multiple transcript models narrows down signals from the gene-based test.

In addition to different variant sets, we examined how isoform-specific scores contribute to differential associations across isoforms by performing PLM-R with the same set of variants overlapping between the transcripts. The associations between JAML and leukocyte and neutrophil counts (e.g., beta = 0.017, SE = 0.003, p = 1.60e−11 versus beta = 0.013, SE = 0.003, p = 3.50e−7 for leukocytes) showed more than 30% higher effect sizes when using ESM1b scores of the non-canonical than canonical transcripts. This suggests that PLM VEPs trained in individual protein isoforms contribute to differential isoform associations (Figure 2).

Figure 2.

Isoform-specific ESM1b scores drive differential association between isoforms

(A) Correlation between ESM1b scores from canonical (JAML-202, x axis) and non-canonical (JAML-204, y axis) transcripts. Red dotted line shows identity line. ESM1b scores trained in isoforms show associations of varying strengths with leukocyte in JAML gene (B). x axis shows ESM1b scores (note that higher scores here indicate higher pathogenicity), and y axis shows inverse-rank normalized phenotype values. Standardized beta, SEs, and p values from the PLM-R are shown. Blue lines show the regression between ESM1b scores and mean leukocyte counts for each missense variant.

EPV associations with human traits

PLM models identify some variants as more likely to be observed in the protein than the wild type.² Recent evidence suggests that this class of variants exists in the opposite functional space to the pathogenic variants typically studied in exome analyses,^23,28 and their frequency and associations with complex human phenotypes are not known. We hypothesized that EPVs may relate to phenotypes in a different way than pathogenic variants do. To test this idea, we identified carriers of EPVs in unrelated European ancestry individuals in UKB WES (n = 348,290) and tested their associations with 82 phenotypes.

We first identified 21,764 rare missense variants (0.45% of all missense variants in UKB) with higher evolutionary likelihood (defined as variants with ESM1b score greater than zero) in UKB WES. Despite their relative paucity, these variants showed higher allele frequencies than non-EPV missense variants (allele frequency = 9.52e−05 versus 3.88e−05, t = −14.16, p < 2.2e−16; Figure S3) and synonymous variants (allele frequency = 9.52e−05 versus 5.34e−05, t = 10.48, p < 2.2e−16). This suggests that, while EPV mutations rarely occur, once they do, they tend to rise in population allele frequency relative to other variant classes. Out of 17,914 protein-coding genes, we identified 694 genes with at least 100 carriers of the EPVs. To characterize these genes, we performed an enrichment test with Gene Ontology pathways. We found significant enrichment of 23 gene sets from genes with EPV carriers, including keratin filament (p = 6.02e−9), homophilic cell adhesion via plasma membrane adhesion molecules (p = 6.79e−5), detection of chemical stimulus (p = 2.78e−4), and structural constituent of postsynaptic actin cytoskeleton (p = 1.86e−3) (see Table S10).

Next, we performed PLM-R and burden test in carriers of wild type and EPVs in the UKB WES. Carriers of pLoF and other missense variants were dropped to isolate the effect of EPVs. An omnibus p value was computed between PLM-R and burden test. After correcting for multiple testing (p < 4.8e−5), nine EPV-phenotype associations were found with 0.06–0.44 SD changes associated with the carrier status. Several gene-phenotype associations were not reported previously: DNM2 associated with lower levels of LDL (beta = −0.116, SE = 0.025), PRORP with higher bone mineral density (beta = 0.43, SE = 0.1), and ZNF99 with later age of menopause (beta = 0.44, SE = 0.1). These associations are shown in Figure 3. We also replicated known associations between ACAN, TARS2, and height; AMN with gamma glutamyl transferase; and LPA with lipoprotein A, indicating the presence of allelic series in this gene, extended to the variants stronger than the wild type (R Table S11). We tested significant EPV associations in four phenotypes (height, LDL, platelet count, basophil count) that exist in the AoU replication cohort (N ≈ 24,046–103,313). Two of the four (DNM2 ∼ LDL and TARS2 ∼ height) gene-trait pairs were replicated (beta = −0.203, SE = 0.089, p = 0.012 and beta = 0.053, SE = 0.018, p = 0.001 from the burden test, respectively) after multiple testing adjustment and another association (ACAN-height) was nominally significant (beta = −0.094, SE = 0.045, p = 0.025 from the burden test). The associations between MUC6 and platelet and basophil count were not replicated in the AoU sample.

Figure 3.

Gene-trait pairs with significant associations with EPV load

x axis shows the ESM1b scores of the missense variants and y axis indicates mean phenotypic values (inverse-rank normalized) of the carriers. Black, red, and blue dashed lines indicate mean phenotypic value of the carriers of the wild-type variants and variants with negative (deleterious) and positive ESM1b scores (EPVs), and triangle shows that of the pLoF carriers. Size of the circle and triangle indicates the number of carriers. Note that the original ESM1b scores, without flipping the sign, were used here for the ease of interpretation (mean, SE).

Discussion

Large protein language models show promise as predictors of variant impact in clinical and experimental benchmarks.^2,3,4,24 However, whether and how PLM models facilitate discovery and interpretation of complex trait associations with rare variants are not well understood. In this study, we presented a regression framework (PLM-R) to incorporate VEPs without arbitrary cutoffs and compared the performance of state-of-the art PLM models on rare-variant association tests, showing that PLM-R outperforms standard burden tests and enhances gene discoveries.

Building on PLM-R and taking advantage of isoform-specific PLM VEPs, we showed that isoform-wide association has higher power to detect gene-trait associations than the test based solely on canonical transcripts, and it detects genes with differential effect sizes across isoforms (i.e., 26 gene-trait pairs showing higher effect sizes in non-canonical than canonical transcript). Finally, we showed that PLMs can be used to identify protein coding variants with higher evolutionary plausibility and detects new associations, including those with potential therapeutic values if experimentally validated (e.g., higher bone mineral density, later age of menopause, and lower LDL).

We found that VEPs from PLMs showed more associations when used as proxies for variant effect sizes in PLM-R than when used as a binary classifier of pathogenic variants in burden test. For example, Alpha Missense detected 156 associations from the burden test when used as a binary classifier, while it detected 244 associations from PLM-R. State-of-the art PLM models found generally comparable numbers of associations (n ≈ 240–248) with ESM2-650M detecting the largest number of associations. They detected non-redundant set of associations as we observed a substantial increase in the number of associations (n = 296) in omnibus test of PLM-R with different VEPs.

To date, only a few studies have examined the usage of PLMs in biobank-scale WES data. Fiziev et al.²⁴ compared burden test with and without using variant stratification based on PrimateAI-3D scores across 90 UKB phenotypes and reported higher power for the former. Parry et al.²⁹ compared PrimateAI-3D and Alpha Missense in their correlation strength with protein abundance measures in the UKB proteomics dataset (n = 41,836) and reported stronger correlation for the test using PrimateAI-3D scores. While informative, these studies included a limited set of PLMs, non-PLM-based VEPs, and gene-based test methods. Furthermore, approaches presented in these studies involve computationally intensive permutation procedures²⁴ or lack inflation-adjusted methods for binary traits.²⁹

Our study presents a simple alternative way of using PLMs that has more power than standard burden tests with and without variant filtering. PLM-R does not require selecting potentially damaging variants and is suited to detecting gene-traits with allelic series, with a relatively small computational burden compared to some of the existing methods (e.g., SKAT and SKAT-O). PLM-R can be further extended to incorporate pLoF and splice variants for the discovery of allelic series, an approach in line with a recently published method that assigns varying weights to different classes of variants.²⁵ Rather than assigning arbitrary weights,²⁵ PLM-R+ uses PLM-based VEPs in weighting variants.

From PLM-R, we detected associations not reported in recent full-scale UKB WES,¹⁰ including diabetes (PAM), education (SPG11), birth weight (OTOP1), lung function (HMCN1), and triglycerides (ANGPTL4). OTOP1 encodes a transmembrane protein, OTOPETRIN1, expressed in adipose tissue, and is involved in metabolic dysfunction in obesity (e.g., attenuating obesity-induced inflammatory response in adipose tissue).³⁰ We found that deleterious missense mutations in this gene predicted greater birth weight. ANGPTL4 plays a key role in regulating lipid metabolism, particularly in the inhibition of lipoprotein lipase, which breaks down triglycerides into free fatty acids for tissue uptake. Despite its known association with triglyceride level, missense mutations in this gene were not associated with triglycerides in recent UKB -based WES analysis, suggesting that PLM-R can identify associations missed by existing approaches.

Human mRNA and protein isoforms are abundant, and they serve important roles in trait and disease biology.³¹ The use of isoform information in WES studies to date is limited; typically only canonical transcripts are considered for annotations or multiple transcripts are considered when determining the variant’s most severe consequence.^9,32 We presented a multi-transcript approach using isoform-specific variant sets and PLM VEP scores in PLM-R and showed that it has more power (n = 256) than using canonical transcripts alone (n = 240). It also detected 26 gene-trait pairs showing higher effect sizes in non-canonical than the canonical transcript where 16 of them were replicated in a held-out UKB sample, suggesting that differential isoform association is a replicable phenomenon with potential relevance to disease biology.

There are several interpretations for higher effect sizes from non-canonical than canonical transcripts in isoform-wide PLM-R. First, transcripts span varying regions of the genome, and some non-canonical transcripts may include causal variants missed by canonical transcripts or cover a higher percentage of causal variants. For example, in follow-up analyses, variants lying in the region overlapping between the transcripts most often showed higher test statistics than the tests using variants specific to each transcript. However, in some cases, variants lying in the region specific to non-canonical transcripts showed higher effect sizes than those in the overlapping region. For example, variants in an exonic region of CLPTM-010, which correspond to the 3′ UTR of CLPTM-001, predicted the level of LDL and C-reactive protein, but variants in other regions in this gene did not. This suggests that differential isoform associations can be leveraged to narrow down the source of signals to smaller regions of the gene.

Second, differential isoform associations may reflect the role of alternative isoforms involved in trait/disease biology. Protein isoforms are known to have various functions, including compensating for the loss or malfunctioning of major isoforms,³³ exacerbating the disease or contributing to distinct pathological features,³⁴ or exerting tissue-specific effects.³⁵ Isoforms showing higher association with a trait may inform transcript biology relevant to the disease. Note that these possibilities are not mutually exclusive. For example, in discovery and replication cohorts, missense variants in GCK consistently showed higher effect sizes for blood glucose level from GCK-202, known as the hepatic form of GCK, compared to GCK-204, known as the pancreatic form. While follow-up study is required to pinpoint what drives this difference, several possibilities coexist: the GCK-202 isoform may be better capturing sub-regions of this gene that affect blood glucose level; mutations in GCK-202 isoform have a significant role in glucose biology (e.g., disrupting liver GCK activity and in turn affecting blood glucose levels).

EPVs represent a previously unexplored space of mutations for population and quantitative genetic analyses. Importantly, it is not straightforward to identify this subset of variants using existing VEPs, because existing models have usually focused on discerning damaging from non-damaging variants. PLMs such as ESM1b learn and predict the likelihood of each amino acid in a protein sequence and provide a simple metric to identify such variants by contrasting the log likelihoods of wild type and mutation. We found that, while representing a small subset of missense mutations, these variants show higher allele frequency than the rest of the missense variants and synonymous variants, consistent with the idea that these variants may collectively be under different selective pressures than other classes of variant.

Several novel gene-trait associations were identified with EPVs. Here, we highlight several associations not reported in earlier analyses of UKB WES.^9,10 Carriers of EPVs in PRORP (N carriers = 90) had an approximately 0.44 SD higher bone mineral density than those without the EPVs. PRORP, also known as MRPP3, encodes a protein processing pre-tRNAs within mitochondria and found to be overexpressed in bone. Missense variants in this gene reduce mitochondrial calcium level and contribute to insulin resistance,³⁶ and variants within and near PRORP were associated with postpartum hypocalcemia,³⁷ indicating its potential role in intracellular calcium regulation. Another association we highlight is between ZNF99 and age of menopause. We found individuals carrying EPVs (n = 90) in this gene reported menopause about 2.22 years later than wild-type carriers. Zinc-finger family protein serves diverse molecular functions from transcription regulation to tissue development and differentiation. While few studies have examined this particular gene, ZNF genes have been implicated in reproductive timing, such as age of menarche^38,39,40 and age at menopause (e.g., ZNF728 located 0.2 MB upstream of ZNF99⁴¹) from several genome-wide association studies (GWASs) and in vivo experimental study with primates.⁴² Overall, the results demonstrate that variants in the opposite spectrum of depletion can be identified by PLM and lead to new phenotypic associations.

Limitations of the study

This study has several limitations. First, while we used raw scores from PLMs (e.g., log likelihood ratios in the case of ESM1b), they can be further tuned to improve for complex trait association testing (e.g., adjusting scale, fine-tuning). Improvements may also be obtained by training models over different protein datasets. Second, the differential functions of the isoforms for most genes are not currently well understood, limiting our ability to leverage isoform-specific discoveries for interpretation. Third, the comparison of discovery power across various VEPs and methods can be influenced by the selection of phenotypes. We selected phenotypes covering various domains from anthropometric, metabolic, hematological, behavioral, and cancer phenotypes, and the performance reflects overall trend across multiple domains. However, certain VEPs/methods may work better than others in a subset of phenotypes or phenotypes not included here. The comparison results can also be influenced by specific cutoffs for VEPs. Similarly, VEPs may have differential accuracy across genes. Lastly, the extent to which PLMs enhance gene discovery when integrated into other statistical approaches, such as deepRVAT,² should be explored in future work. Despite these limitations, the frameworks presented here can facilitate discovery of allelic series and isoform target for transcript biology and expand the scope of the coding variants contributing to phenotypic spectrum of complex traits. Current study also offers a systematic evaluation of state-of-the-art PLMs and their alternatives in complex trait gene discovery.

Resource availability

Lead contact

Further information and requests for resources should be directed to the lead contact, Noah Zaitlen (nzaitlen@g.ucla.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •UKB whole-exome sequence is available through Research Analysis Platform (https://ukbiobank.dnanexus.com/landing).
• •All of Us whole-exome sequence is available through AoU research hub (https://www.researchallofus.org/).
• •ESM models can be accessed at https://github.com/facebookresearch/esm and https://github.com/evolutionaryscale/esm.
• •PrimateAI-3D scores can be obtained by applying through https://primateai3d.basespace.illumina.com/download.
• •CADD, ESM1b, and Alpha Missense scores can be obtained through dbNSFP (https://www.dbnsfp.org/).
• •CPT-1 scores can be obtained at https://zenodo.org/records/8137108.
• •PLM-R script is available at https://github.com/sunkjang/PLM-R and has been archived in Zenodo (https://zenodo.org/records/15664945) with the DOI https://doi.org/10.5281/zenodo.15664944.

Acknowledgments

This work was supported by the National Institute of Mental Health under R01MH130581. This research was conducted using the UK Biobank Resource under application 33127. We gratefully acknowledge the participants of UK Biobank and All of Us for their contributions, without whom this research would not have been possible.

Author contributions

N.Z., J.F., and S.-K.J. conceived the idea and designed the study. S.-K.J., Z.W., R.B., U.A., D.T., V.N., A.W., and S.S. prepared materials. S.-K.J. performed the analysis. S.-K.J., N.Z., and J.F. wrote the manuscript. All authors reviewed and approved the paper.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
UKB whole exome sequence	UK Biobank	https://ukbiobank.dnanexus.com/landing
All of Us whole exome sequence	All of Us	https://www.researchallofus.org/
ESM1b scores	Brandes et al.2	https://github.com/facebookresearch/esm
ESM2 scores	Lin et al.3	https://github.com/facebookresearch/esm
ESM3 scores	Hayes et al.43	https://github.com/evolutionaryscale/esm
CADD scores	Rentzch et al.21	https://cadd.gs.washington.edu/
Alpha Missense scores	Cheng et al.5	https://github.com/google-deepmind/alphamissense
PrimateAI-3D scores	Parry et al.29	https://primateai3d.basespace.illumina.com/download
CPT-1 scores	Jagota et al.44	https://zenodo.org/records/8137108
VEPs annotated for missense variants	Liu et al.45	https://www.dbnsfp.org/
Software and algorithms
REGENIE v3.3	Mbatchou et al.46	https://github.com/rgcgithub/regenie
PLM-R	This study	https://github.com/sunkjang/PLM-R (doi: https://doi.org/10.5281/zenodo.15664944)

Method details

Phenotype and genetic data

We included whole exome sequencing (WES) data of unrelated European (EUR) ancestry individuals in UK Biobank (UKB) in this study. UKB WES analysis protocols and QC process are documented in Szustakowski et al.⁴⁷ We used self-report and genomic PC clustering to subset individuals with EUR ancestry. Specifically, after excluding samples with excess missing genotypes, mismatch between self-reported and genetic sex, and excess heterozygosity (>5 SD above mean), we estimated relatedness using KING and removed 629 individuals with ten or more potential 3rd degree relatives, then used the `maximal_independent_set` algorithm in the python module `networkx` to selectively prune away individuals and maximize the number of unrelated individuals, resulting in 348,290 individuals in total. We used a subset of 153,823 individuals from the 2^nd UK biobank WES release for validating PLM-R and isoform-wide analysis and used the whole sample (N = 348,290) for EPV associations. For all analyses, we used 82 complex phenotypes (43 continuous and 39 binary) from UK Biobank. We selected a wide range of phenotypes encompassing blood, lipid, anthropometric, behavioral, cardiovascular, and cancer phenotypes, so that the results reflect general patterns observed in multiple phenotypic domains and are relevant to diverse disease areas in human genetics. For continuous phenotypes, we applied inverse-rank-based normalization for all analyses. Details of the phenotypes, including phenotype codes and additional QC process, can be found in Table S1.

Variant annotation

We used ENSEMBL vep (v.109) to annotate variants with their functional consequences (e.g., loss of function, splice, and missense variants). pLoF variants were identified as those defined as ‘high confidence’ by LOFTEE algorithm⁴⁸ and splice variants were identified by spliceAI (delta score >0.5).⁴⁹ Missense variants were annotated with widely-used VEP scores, including PolyPhen2-HumVar/HumDiv,⁶ SIFT,⁷ Mutation Taster,⁵⁰ and LRT.⁵¹ We also annotated missense variants with their predicted pathogenicity from PLM-based VEPs, including four recent ESM models: ESM1b,² ESM2 (650 million parameters),³ ESM3 (ESM3 without structural information during inference),⁴³ ESM3-structure (ESM3 with structural information during inference;ESM3-sst hereafter),⁴³ along with PriamteAI-3D⁴ and Alpha Missense.⁵ As an alternative to PLM, we also included VEP score from CPT-1, a transfer learning paradigm using deep mutational scanning⁴⁴ as well as CADD-v.1.7 which incorporated ESM 1-v scores in training. For ESM models (see URLs), we extracted log likelihoods of wildtype and mutation residues and calculated the log likelihood ratio (LLR) between the two. Predicted AlphaFold⁵² structures were used to extract VEP scores from ESM3-sst. We obtained CADD, CPT-1, and PrimateAI-3D scores from the developer for non-commercial use. For Alpha Missense and non-PLM-based VEPs, predicted pathogenicity scores were extracted from dbNSFP database (see URLs). Annotations from all available transcripts in ensemble database were collected for isoform analysis. For all VEPs, we adjusted the signs of the VEP scores so that higher scores indicate higher pathogenicity unless mentioned otherwise.

Quantification and statistical analysis

Burden test

REGENIE (v3.3) was used to conduct standard burden test.⁴⁶ We performed whole genome regression (step 1) using 956,669 variants (MAF >1%) from UKB WES data after linkage disequilibrium (LD) pruning (--indep-pairwise 1000 100 0.6) to generate leave-one-chromosome-out genetic predictors for step 2. We performed burden test (step 2) with MAF weights drawn from beta (1,25) distribution using either all available missense variants or putative pathogenic missense variants filtered by various VEPs. To define putative pathogenic variants, we used categories or cut-offs suggested by the developer for PolyPhen2 (“probably damaging”), SIFT (“deleterious”, “deleterious low confidence”), ESM1b (<-7.5), and Alpha Missense (“likely pathogenic”). Variant predictions from canonical transcript were used unless they are available only for noncanonical transcripts. We used two-sided p - values and adjusted for the number of genes tested (p <0.05/17,971).

PLM-R

Based on the observation that PLM-based VEPs approximate effect sizes of the variants in our previous work (ref), we use following regression framework for testing the association between rare variants in a given gene and a phenotype of interest.

$$\begin{array}{c}y\sim X\alpha +E\beta +\epsilon \end{array}$$ (Equation 1)

where y is a (n x 1) phenotype vector, $X$ is a (n x p) matrix of covariates, and $E$ is a (n x 1) vector of PLM VEP scores of the missense variants one carries within a gene, and $\epsilon$ is a random noise drawn from $N\left(0,1\right)$. We tested following null hypothesis: H₀: $\beta =0$. When a person carries multiple missense variants, we used the minimum (most deleterious) score and when a person carries no missense variants, we assigned them a score of zero. When there are multiple annotations available for the same variant from different transcripts, we took two approaches: 1) keeping the score from canonical transcript and 2) keeping the minimum ESM1b score across the transcripts, to check the sensitivity of the results to the choice of transcript. We also present PLM-R+ where one incorporates pPLoF and splice variants into this regression. Since pLoF and splice variants do not have VEP scores, we assign the minimum and the top 33% of the VEP scores to the pLoF and splice variants, respectively.⁴

We perform PLM-R for the associations between 17,971 protein coding genes and 82 phenotypes. We use linear regression with two-sided Wald test for continuous traits. Testing the association of rare variants with rare binary traits with case control imbalance has been found to have increased false positive rates.⁵³ We performed the n-of-1 permutation to determine gene-based test methods for binary traits with proper control of type I error rate and appropriate p-value threshold. First, we created a set of permuted phenotypes (total 82 sets). Then we performed linear regression for quantitative traits and six different tests for binary traits. Tests for binary traits include 1) logistic regression with Wald test, 2) logistic regression with Likelihood Ratio Test (LRT), 3) firth logistic regression with Wald test using firth-corrected effect estimates, 4) approximate firth logistic regression with penalized likelihood ratio (pLRT) test as formulated in Mbatchou et al.,⁴⁶ 5) firth logistic regression with pLRT as implanted in logistf package in R, and 6) Saddle Point Approximation (SPA) test⁵³ as implemented using SPAtest package in R. At p-value <0.05/17,971, we found one false positive out of 735,608 tests for 43 quantitative traits (Figure S1). For binary traits, we found 9, 54, 21, 3, 4, and 1 false positives for 39 binary traits from logistic regression with Wald test, logistic regression with LRT, firth logistic regression with Wald test, firth logistic regression with pLRT from Mbatchou et al.,⁴⁶ firth logistic regression with pLRT from logistf, and SPA, respectively (Figure S2). Thus, we reported p-values from SPA and firth-corrected effect estimates for binary traits, chose multiple testing threshold at p < 0.05/17,971.

Isoform-wide PLM-R

153,823 unrelated, EUR ancestry individuals in UKB were used for the discovery stage of isoform-wide associations and differential isoform test. We adopted the two-step procedure to identify gene-trait pairs with higher effect size in non-canonical than canonical transcript. First, we performed PLM-R across transcripts from 17,951 protein coding genes using VEP scores trained in individual isoforms. In this analysis, we used ESM1b as ESM1b scores trained in protein isoforms are publicly available. Given the large number of transcripts in ensemble database (N = 276,949), we limited the search space to the transcripts with TSL greater than three to increase the chance that the transcripts are indeed expressed in human tissues, increase the signal to noise ratio, and to decrease multiple testing burden. This resulted in 57,368 transcripts from 17,951 protein coding genes where 12,918 genes had more than two transcripts. We annotated missense variants with ESM1b scores from each of these transcripts and performed PLM-R separately for transcript-phenotype pairs. To adjust for multiple testing, we computed the number of independent test (M_e) as these transcripts are highly correlated. To do this, we calculated the correlation matrix of the ESM1b scores of the missense variants from all transcripts within a gene and obtained eigenvalues. Then, we calculated $M_{e,i}=M-\sum _{j=1}^M\left[I\left({\lambda }_j>1\right)\left({\lambda }_j-1\right)\right]$ for the i-th gene,⁵⁴ where M is the number of transcripts within a gene and ${\lambda }_j$ are the j-th eigenvalue. We summed $M_{e,i}$ across the genes, resulting in M_e = 28,882.

After screening significant transcript-phenotype pairs at p < 0.05/28,882, we computed $\delta {=beta}_2-{beta}_1$ for these pairs, where ${beta}_1$ and ${beta}_2$ are the effect size estimates from canonical and non-canonical transcript, respectively. For continuous traits, effect sizes estimates are standardized betas. For binary traits, SPAtest only accepts the dosage values between 0 and 2, we applied min-max scaling first so that the scores have the desired range, and equalized their variance afterward. Next, we conducted permutation (1,000 replicates) where we constructed the null distribution of ${\delta }_{perm}{=beta}_{2,perm}-{beta}_{1,perm}$. We calculated 2-sided p-value as the proportion of ${\delta }_{perm}$ which is more extreme than $\delta$ in the direction where non-canonical transcript shows higher effect sizes than canonical transcript and vice versa. Finally, we applied Bonferroni correction (0.05/the number of isoform-level associations tested) to the p-value when there are multiple pairs of transcripts to compare within a given gene.

Follow-up analysis of differential isoform association

We conducted follow-up analyses to identify potential sources driving the differential association. We divided variants into those overlapping between the transcripts, and those unique to each transcript. We performed burden test and SKAT-O separately for these variant sets to see if the differential association is driven by the inclusion of different sets of variants. Other than the different variants included in each transcript, it is also possible that the difference in VEP scores themselves contributes to the differential associations between transcripts even though these transcripts share the same set of variants. To confirm this possibility, we performed PLM-R using isoform-specific ESM1b scores of the variants overlapping between the transcripts. Using this same set of variants with potentially different ESM1b scores, we identified gene-trait pairs showing at least 20% higher effect size in non-canonical transcript than in canonical transcript.

Replicating isoform associations in UK biobank held-out sample and All of Us

Once we identified significant isoform-phenotype pairs with higher effect sizes in non-canonical transcripts, we attempted to replicate these pairs in held-out UK biobank WES sample (N = ∼194,467). We performed PLM-R using isoform-specific ESM1b scores with age, sex, and the top 15 genetic PCs as covariates, and reported one-sided pp-value. Then we calculated pp-values for the difference of the two effect size estimates (i.e., ${{\delta }_{perm}={\beta }_2-\beta }_1$) using the permutation procedure used in the main isoform analysis (see STAR Methods). We also attempted replication for significant isoform-phenotype pairs in AoU in case corresponding phenotypes exist in AoU. This includes 15 isoform-trait pairs involving glucose, LDL, HDL, triglycerides, platelet counts, and leukocyte/red blood cell/neutrophil/eosinophil/monocyte counts, and alkaline phosphatase (N = 24,293 to 104,939). Details of the phenotype definition and harmonization can be found in Table S1. We adjusted for age, ageˆ2, sex, age∗sex, and the top 15 genetic PCs as covariates, followed by permutation to detect differential isoform association in AoU.

EPV analysis

Among missense variants found in the UKB WES of unrelated EUR ancestry individuals, we identified 21,764 rare missense variants whose ESM1b score is greater than zero (here we refer them as Evolutionary Plausible Variants; EPV), meaning that its protein product is predicted to be in higher evolutionary likelihood. When there are multiple ESM1b scores from different transcripts for the same variant, we kept the most deleterious ESM1b score. For this analysis, signs of ESM1b scores were not flipped for easier interpretation. We calculated mean allele frequency of these variants and compared it to allele frequency of the non-EPV missense variants and synonymous variants using t-test. We reserved a set of genes having at least 100 carriers for gene set enrichment test and phenotypic scan. We examined the characteristics of these genes using gene set enrichment test in FUMA (hypergeometric test; ENSEMBL version 102, FDR multiple testing correction) with gene ontology pathways. We then performed PLM-R and burden test using only wildtype and EPV carriers (excluding the carriers of missense variants with ESM1b score lower than zero and pLoF variants) to isolate the effect of EPVs from other classes of variants. We calculated omnibus p-values from the PLM-R using an aggregated Cauchy association test (ACAT)⁵⁵ to increase statistical power and refered it to PLM-R ACAT (two-sided test). We set multiple testing threshold at p < 4.8e−5 which showed appropriate type I error rate in the n-of-1 permutation (one false positive detected out of 56,908 tests).

Replicating EPV associations in AoU

For a subset of phenotypes that exist in AoU (LDL, height, platelet count), we attempted to replicate the four out of nine EPV associations using the WES of unrelated EUR sample in AoU (N = 24,046–103,313). European ancestry was inferred by relatedness estimates and predicted ancestry released by AoU (i.e., predicted probability of EUR ancestry >0.9, probability of other ancestries <0.05). We inverse-rank transformed phenotypes and tested the association with EPVs using PLM-R and burden test after which omnibus p-values of the two tests were calculated. We adjusted for age, sex, top 15 PCs, and BMI and conducted one-sided tests with multiple testing adjustment by the number of total gene-trait pairs tested. Details of the phenotype definitions and QC process can be found in Table S1.

Published: November 19, 2025