High-Penetrance Rare Variants Underlying Familial Lung Cancer Risk: Insights From Genetic Epidemiology of Lung Cancer Consortium

Abstract

Introduction:

Rare, deleterious germline variants are key contributors to inherited lung cancer (LC) risk. The Genetic Epidemiology of LC Consortium (GELCC) has curated valuable high-risk LC families and is uniquely positioned to uncover rare, high-penetrance variants underlying familial LC (FLC).

Methods:

We performed whole-genome and exome sequencing on germline DNA from 120 high-risk LC families (177 FLC cases, 309 unaffected relatives). We prioritized rare (allele frequency <1% in the genome aggregation database), potentially deleterious variants present in two or more FLC cases. These variants were then validated in 10,085 sporadic LC (SLC) cases and 612,970 controls.

Results:

We identified 118 candidate variants, 28 of which were validated in SLC with strong statistical support. We discovered a novel pathogenic axis of three truncating variants in GALNT6, MUC4, and ERBB3 genes, which are critical regulators of mucin-type O-glycosylation. Nine top hits were mapped to the known 6q23-25 linkage region (ROS1, LAMA2, PRKN, SYNE1). Other candidates were clustered in DNA repair (ATM, BRCA2, MLH1), oncogenic signaling (ERBB3, JAK1, PIM1), and extracellular matrix genes (COL6A3, FLG). Carriers of two or more variant alleles had a strong dose-dependent risk. Furthermore, gene-based burden tests revealed strong associations between RARB, MGMT, and EBF1 with FLC susceptibility.

Conclusion:

Our findings underscore the important role of rare, high-penetrance genetic variants in FLC susceptibility, particularly in mucin glycosylation and DNA repair genes. These findings offer promising targets for early detection and personalized therapies.

Introduction

Lung cancer (LC) remains the leading cause of cancer mortality in the United States, despite declining incidence and mortality due to reduced smoking, earlier diagnosis, and advances in targeted and immunotherapies.¹ Although tobacco exposure is the primary environmental risk factor, inherited genetic susceptibility plays a substantial role in LC risk, especially among individuals with a strong family history of LC (FHLC). Since the first report in 1963,² studies have revealed that having a first-degree relative with LC increases risk twofold to fourfold, independent of smoking and other risk factors.^3-12 An early epidemiologic study from the Genetic Epidemiology of LC Consortium (GELCC) revealed that approximately 13% of LC cases had at least one additional affected relative.⁴ Approximately 1% of cases are from high-risk familial LC (HRFLC) families, defined as families with three or more affected individuals, a prevalence comparable to Lynch syndrome in colorectal cancer.¹³

A major finding came from a genome-wide linkage analysis in 52 HRFLC pedigrees collected by GELCC, which identified a significant locus at 6q23-25 (HLOD = 2.79).¹⁴ Subsequent sequencing studies implicated PARK2 (also known as PRKN)^15,16 and RGS17¹⁷ as candidate genes in this linkage region. Further fine mapping and segregation analyses from GELCC and other groups suggested more than 15 additional FLC risk loci.^18-24 However, these studies used multiallelic microsatellite genotypes, which are sparsely distributed across the genome. Moreover, the identified loci collectively account for fewer than 5% of FLC cases. Most HRFLC families lack known large-effect variants, and reported susceptibility loci vary widely among families. Even when combined with common variants identified from genome-wide association studies (GWAS) of sporadic LC (SLC), less than 20% of heritability is explained,²⁵ underscoring the critical need to identify additional inherited risk variants.

Rare variants (minor allele frequency [MAF] < 1%), particularly those that are functionally deleterious variants, often confer large effect sizes,^26-28 as they tend to be evolutionarily recent and under strong negative selection.^29,30 Many familial hereditary cancer syndromes are driven by such variants (e.g., TP53 in Li-Fraumeni syndrome, BRCA1/2 in breast and ovarian cancers, MLH1/MSH2 in Lynch syndrome).^31-33 In FLC, rare coding variants have been reported in TP53, EGFR, HER2, RB1, MET, MAST1, YAP1, and SFTPA/C, although only in a small number of families.^34-47 GELCC^48-53 and other^54,55 studies have also identified rare, moderate-to-high risk variants in ATM, BRCA2, CHEK2, FGF5, and ALKBH2. Despite these findings, the high-penetrance germline variants have not yet been identified for most HRFLC families.

To address this deficiency, we conducted whole-exome sequencing (WES) and whole-genome sequencing (WGS) on DNA from 120 GELCC families spanning three generations with multiple affected members. We hypothesized that these families represent a genetically enriched subset in which rare, high-penetrance variants are more detectable and contribute to greater LC heritability. Defining the role of rare genetic variation to FLC risk will expand our understanding of disease etiology and may ultimately inform more effective precision strategies for genetic risk assessment, early detection, and prevention.

Methods

Study Subjects and Sample Collection in the GELCC Study

GELCC is a U.S.-based research network studying FLC. For more than 25 years, we have enrolled more than 1000 families with a strong FHLC. Recruitment was conducted by institutions such as the University of Cincinnati, University of Colorado, Karmanos Cancer Institute/Wayne State University, Louisiana State University Health Sciences Center, Mayo Clinic, and The University of Toledo. Recruitment methods and participant details are described previously.²² FLC cases were defined as individuals with a histologically confirmed diagnosis of LC (International Classification of Diseases, Ninth or Tenth Revision codes: 162.0–162.9)²¹ and at least one first- or second-degree relative with LC. The diagnosis of LC in both the index case and relatives was verified with medical records, tumor registries, and death certificates when possible. Unaffected relatives included first-, second-, or third-degree blood relatives without LC diagnosis, although some had other cancers. For this study, families were selected based on the following criteria: (1) a well-characterized, multigenerational pedigree spanning three generations; (2) the presence of at least three members affected with LC; and (3) availability of archival samples (blood or saliva) for DNA sequencing. Never smokers were defined as those who smoked fewer than 100 cigarettes in their lifetime.

WES, WGS, and Variant Calling

We performed WES and WGS using DNA extracted from white blood cells of participants. WES was performed on 318 subjects (122 affected FLCs, 196 unaffected relatives) from 80 families using three exome capture platforms: NimbleGen SeqCap (Roche), SureSelect version 5 supplemented with a custom capture (Agilent), and VCRome 2.1 (Roche). Mean coverage was 52×, with more than 97% of targeted bases covered at more than or equal to 100× depth. In addition, WGS (polymerase chain reaction free, Illumina NovaSeq) was performed on 168 samples (55 affected FLCs, 113 unaffected relatives) from 46 families at the National Intramural Sequencing Center at the National Human Genome Research Institute, to an average coverage of 30×.

Sequence reads were aligned to the human reference genome (GRCh38) using the Burrows-Wheeler Aligner. Joint variant calling for both WES and WGS data was performed using the Illumina DRAGEN (Dynamic Read Analysis for GENomics) pipeline,⁵⁶ which has demonstrated superior accuracy than GATK for detecting single-nucleotide variants (SNVs) and insertions/deletions (indels). For WES, analyses were restricted to target regions defined by a common browser extensible data (BED) file covering all three exome capture kits. This ensured direct comparability of variant detection across kits and minimized capture-specific biases. WGS alignments used the entire genome, but analyses were to the exomes to ensure compatibility with the WES analyses. The quality control (QC) procedures include the following: (1) excluding variants located in low-complexity regions or segmental duplications; (2) filtering variants with Phred quality score less than 30, depth less than 10, or allelic balance (AB) less than 0.2 for heterozygous calls; (3) requiring a variant call rate more than or equal to 0.85; and (4) removing samples with abnormal heterozygosity, sex discordance, completion rates less than 95%, or unexpected relatedness (identity by state >10%).

Variant Filtering and Functional Annotation

Variants were first annotated using ANNOVAR and then filtered by a three-tiered strategy: (1) allele frequency: retain rare variants with MAF less than 1% in the global population from the Genome Aggregation Database (gnomAD v.4, n = 807,162 individuals); (2) variant class: focus on protein-altering variants (SNVs and small indels ≤ 21 bp), including missense, stop gained/lost, truncating frameshift indels, and splice donor/acceptor variants; (3) variant effect: retain only variants with a scaled Combined Annotation Dependent Depletion (CADD) Phred like score more than or equal to 20, representing top 1% most deleterious substitutions genome wide.^57,58 Last, we used Genome Browser visualization (read-depth and pile-up) to confirm candidate variants and exclude those with low-confidence variants.^59-62

Variant Prioritization Based on Clinical and Functional Relevance

We applied an evidence-based framework to prioritize rare variants across three tiers: Tier 1 (pathogenicity): Variants with clinical or experimental evidence of pathogenicity, including ClinVar annotations (pathogenic, likely pathogenic, or drug response). Tier 2 (oncogenic relevance): Variants located in curated cancer driver genes (OncoKB, COSMIC), including protooncogenes, tumor suppressors, and genes associated with LC phenotypes, as identified by Human Phenotype Ontology (HPO) or the NHGRI-EBI GWAS Catalog. Tier 3 (functional impact): Variants in genes predicted to have gain or loss of function effects at the gene level, based on LoGoFunc or GoFCards. This tiered approach integrates established cancer biology with novel candidates, allowing us to capture both known and putative susceptibility genes.

Single Variant Association Test in the Discovery Set

To further prioritize the candidate variants, we retained rare variants present in more than or equal to two affected FLC cases and compared variant allele counts between FLC cases and gnomAD controls. Unaffected relatives within GELCC families were excluded from association testing. Association analyses were conducted using Fisher’s exact test under an additive genetic model. Covariate adjustment was not performed due to the limited number of variant carriers. To address instability of effect size estimates (odds ratio [OR]) for ultra-rare variants (MAF < 0.001%) with sparse carrier counts, we applied an Empirical Bayes shrinkage approach,⁶³ which stabilizes estimates by shrinking extreme values toward the overall distribution of effect sizes. Multiple testing was accounted for by calculating false discovery rate (FDR)–adjusted p values.

Candidates Co-Occurring, Gene-Based Burden, and Network Analysis in the Discovery Set

To assess variant clustering, we evaluated the distribution of candidate variants within GELCC families and carrier demographic characteristics. To test gene-level association, we applied the Combined Multivariate and Collapsing (CMC)⁶⁴ method to genes containing three or more rare, predicted deleterious variants that passed the three-tier filtering pipeline. Variants were collapsed into bins by MAF and tested with Hotelling’s T², with FDR-adjusted p values to account for multiple comparisons.

Gene-gene interactions were evaluated using STRING (Search Tool for the Retrieval of Interacting Genes/Proteins),⁶⁵ which integrates experimental evidence, curated databases, and text mining.

Study Populations in the Validation Sets and Meta-Analyses

Candidate variants were validated in five independent data sets, including two national biobanks (All of Us [AOU] and UK Biobank [UKB]), one population-based consortium (International LC Consortium, ILCCO), and two clinically enrolled cohorts (Baylor College of Medicine [BCM] and The Cancer Genome Atlas [TCGA]).

The ILCCO data set included participants from Harvard School of Public Health (HSPH), the International Agency for Research on Cancer (IARC), the University of Liverpool, and Mount Sinai Hospital–Princess Margaret Hospital (MSH–PMH) in Toronto. It comprised 1094 SLC cases and 885 non-cancer controls, primarily of European descent, enriched for individuals with a first-degree family history (FHLC) or early-onset (<55 y).⁵⁰ WES was performed using Agilent SureSelect XT and Whole Exome v.5 capture kits. The BCM cohort included 398 hospital-enrolled SLC cases, with WES performed using NimbleGen VCRome 2.1 capture and 30× WGS. The TCGA study included germline DNA from 1015 patients with SLC, with WES performed using Agilent SureSelect and NimbleGen SeqCap capture platforms. From the AoU cohort, 2775 SLC cases and 268,332 cancer-free controls of European and African American ancestries were selected, all with 30× WGS data. The UKB data set consisted of 4902 SLC cases and 401,453 cancer-free controls, predominantly of European ancestry, with WES data generated using the Axiom Array platform.

Meta-analyses were conducted using validation data sets alone and in combination with discovery data, applying an inverse-variance weighted fixed-effects model, assuming a common effect size across studies. To assess robustness and potential heterogeneity, we also conducted sensitivity analyses stratified by key characteristics, including sex, race, and smoking status.

Results

Demographics and Characteristics of Study Subjects

This GELCC study included 486 individuals from 120 families, comprising 177 FLC cases and 309 relatives who were either unaffected or had unknown LC status. Families had 3 to 14 affected members, with 42 families (33%) having more than or equal to five affected individuals. Of the 120 families, 55 were biospecimen limited (only one FLC sequenced per family),⁴⁹ whereas 65 families had two or more sequenced members. Among the 177 FLC cases, three also had breast cancer and one had prostate cancer. Among the relatives, 28 (9%) were married-in controls and 52 (17%) had other cancers, including breast (n = 15), prostate (n = 9), and others (Supplementary Table 1). Demographics are summarized in Table 1. Of the participants, 13% self-reported as African American (38 FLCs and 37 unaffected relatives). Smoking history was reported in 80% of FLC cases versus 43% of unaffected relatives. Regarding histology, lung adenocarcinoma accounted for 58% of FLC cases, followed by squamous cell carcinoma (30%), and large cell and small cell carcinoma (approximately 15%). For comparison, 807,162 gnomAD controls were included, with approximately 17% non-European and approximately 5% African ancestry.

Table 1.

Basic Characteristics of the Study Participants

Study Phase	Discovery			Validation
Characteristics	GELCC		GnomAD v4	ILCCO-BCM-TCGA		UK Biobank		All of Us
N. (%)	FLC 177	Relative 309	Controls 807,162	SLC 2408	Control 885	SLC 4902	Control 343,753	SLC 2775	Control 268,332
Race
European (%)	143 (81)	272 (88)	622,057 (77)	2071 (86)	814 (92)	4755 (97)	323,128 (94)	2270 (82)	191,662 (72)
African (%)	32 (18)	34 (11)	37,545 (4.7)	241 (10)	44 (5)	22 (0.4)	3057 (0.8)	505 (18)	76,670 (28)
Others (%)	2 (1)	3 (1)	147,560 (18.3)	96 (4)	27 (3)	125 (2.5)	17,568 (5)	0	0
Sex
Male (%)	73 (41)	130 (42)	400,897 (49.7)	1421 (59)	513 (58)	2696 (55)	158,276 (46)	1197 (43)	105,394 (40)
Female (%)	103 (58)	176 (57)	406,265 (50)	987 (41)	372 (42)	2385 (45)	185,477 (54)	1533 (56)	159,979 (59)
Age, y
Mean (range)	63 (28-86)	61 (16-95)	54 (20-89)	63 (32-89)	61 (20-80)	62 (40-70)	56 (37-72)	67 (20-89)	52 (29-89)
≤55 y (%)	35 (20)	99 (32)	206,661 (25.6)	650 (27)	195 (22)	588 (12)	151,251 (44)	380 (14)	140,397 (52)
Smoking history
Ever (%)	141 (80)	139 (45)	n.a.	1999 (83)	576 (65)	4310 (88)	200,514 (58)	1818 (66)	120,058 (45)
Never (%)	35 (20)	127 (41)	n.a.	361 (15)	301 (34)	546 (11)	141,681 (41)	896 (32)	148,274 (55)
LC family historya
Yes (%)	177 (100)	281 (91)	n.a.	699 (29)	71 (8)	980 (20)	41,250 (11)	300 (11)	24,563 (8.2)
No (%)	0 (0)	28 (9)a married-in	n.a.	1109 (46)	487 (55)	3823 (78)	295,627 (86)	2475 (89)	243,769 (91)
Histologyb
Adenocarcinoma (%)	92 (52)	-	-	1216 (50)	-	1965 (40)	-	n.a.	-
Squamous (%)	55 (31)	-	-	734 (30.5)	-	1128 (21)	-	n.a.	-
Other (%)	30 (17)	-	-	458 (19)	-	1809 (37)	-	n.a.	-

^aMarried-in relatives in GELCC families who were not biologically related to the FLC cases.

^bOther histologies include large cell carcinoma, SCLC, mixed, and unspecified types.

AOU, All of Us research program; FLC, familial lung cancer; GELCC, Genetic Epidemiology of Lung Cancer Consortium; GnomAD, genome aggregation database v.4; ILCCO-BCM-TCGA, combined data set of International Lung Cancer Consortium (ILCCO), Baylor College of Medicine (BCM), and TCGA (The Cancer Genome Atlas); n.a., not available; SLC, sporadic lung cancer; UKB, UK Biobank.

The validation phase included 10,085 SLC cases and 612,970 controls (Table 1). Ethnic composition and smoking prevalence varied across cohorts. Most participants were of European ancestry, though the AOU cohort contributed substantial diversity, with African Americans representing 18% of SLC cases (n = 505) and 28% of controls (n = 76,670). Smoking prevalence was consistently higher in cases than controls across all the data sets: 83% versus 65% (ILCCO-BCM-TCGA), 88% versus 58% (UKB), and 66% versus 45% (AOU), all revealing statistically significant differences in smoking rates (p = 0.0058, 2.48 × 10⁻⁶, 0.0043, respectively). Age distributions also differed, with cases generally older than controls, and early onset cases (≤55 y) enriched in GELCC (20%) and ILCCO (27%).

Identification of Rare, Deleterious Variants

From 2,182,753 variants detected in 486 GELCC subjects, stepwise filtering identified 53,306 rare, potentially functional variants: 79% missense, 15% frameshift indels, 4% stop-gain/loss, and 2% splice acceptors/receptors. Of these, 2850 were present in more than or equal to 2 patients with FLC. Single-variant allelic association testing identified 118 candidate variants. Detailed variant annotations, location, rs ID, MAF, and carrier counts are provided in Supplementary Table 2.

Validation and Combined Meta-Analyses

We assessed 118 candidate variants in external validation data sets, of which two PRDM9 truncating variants (p.N815X and p.K816Gfs*124) were absent, leaving 116 variants for further analysis. Among these, 60 had a significant association in meta-analyses. However, 14 had protective effects that were inconsistent with discovery results. In the combined discovery-validation meta-analysis, 77 variants reached statistical significance, with 48 having consistent signals in both validation and combined analyses.

Table 2 reveals the top 28 hits, including 17 missense variants, 10 protein-truncating variants (six frameshift and four stop-gain), and one splice acceptor variant. Notably, 11 variants had a CADD score more than or equal to 30, ranking within the top 0.1% of the most deleterious substitutions in the human genome. Supplementary Figure 1 presents effect sizes across individual studies and the combined meta-analysis. Among the validations, ILCCO-BCM-TCGA displayed the smallest OR with the widest confidence intervals (CIs), whereas AoU demonstrated higher ORs. These effect sizes should be interpreted with caution due to the limited number of variant carriers. In addition, estimates from the combined meta-analysis may be upwardly biased because of selection in the discovery phase rather than independent replication.

Table 2.

Top 28 Candidates Identified in GELCC Discovery and Validation Sets

Variant Annotation					Discovery: 177 vs. 807,162		Validation: 10,085 vs. 612,970		Combined Meta-Analysis
Gene	Varianta	CADDb ClinVar	Located Domain	MAF% in GnomAD	n. Carriers in Families (FLC ∣ Relative)	OR (95% CI)c	n. Carriers in SLC	OR (95% CI)c	OR (95% CI)c	p Value
Tier 1: ClinVar pathogenic (clinical evidence)
LAMA 2 6q23-25	p.R2383X*	44 ⚠	Laminin G	0.0008	2 (2 ∣ 2) SMK	86 (29.1-298)	0	6.32 (1.31-30.5)	60.3 (25.56-179)	1.4 × 10−13
	p.G2472V	28		0.02	2 (2 ∣ 0) SMK	25.1 (6.2-101.5)	40	47.5 (26.1-86.6)	38.9 (24.73-75)	1.3 × 10−40
	p.A3054E	23		0.009	2 (2 ∣ 0) SMK	61 (14.9-214)	29	45.4 (24.1-85.5)	41.6 (26.17-83)	5.6 × 10−39
PRKN ⌧6q23-25	p.Q34R	23	Ubiquitin ligase	0.2	3 (4 ∣ 0) AA	7.5 (2.8-20.3)	31	2.23 (1.54-3.22)	2.58 (1.83-3.65)	7.6 × 10−8
	R275W	26 ⚠	RING1 core	0.3	4(5 ∣ 3)	4.8 (1.9-11.7)	87	1.94 (1.52-2.37)	2.01 (1.62-2.48)	2.0 × 10−10
FLG	p.R3009X*	27 ⚠	Filaggrin repeats	0.004	4 (3 ∣ 4) AA	58 (13.7-255)	22	63.7 (31.5-128)	68.2 (47.1-158)	8.0 × 10−48
	p.Y3105X*	33		0.001	2 (4 ∣ 2) AA	81 (27-503)	4	30.1 (1.9-202)	68.2 (26.3-424)	1.6 × 10−14
	p.R3404Gfs*51	30		0.00001	5 (7 ∣ 2) AA	73 (20-369)	3	9.91 (1.7-56.7)	59.5 (13.4-210)	7.8 × 10−8
ASXL1 ⌧ +	p.G646Wfs*12	34 ⚠	N-terminal	0.04	3 (2 ∣ 5)	13.5 (3.3-54.5)	30	3.99 (2.5-6.31)	4.49 (2.91-6.94)	1.1 × 10−11
BRCA2 ⌧	p.N986Kfs*2	25 ⚠	BRC repeat	0.0002	1 (2 ∣ 0) AA, F, EO	52.5 (13-449)	22	94 (16-351)	68.2 (22.8-205)	1.9 × 10−23
CFTR ⌧ +	p.R74W	29 ⚠	N-terminal	0.09	4 (4 ∣ 2) AA	12.5 (4.6-33.8)	30	1.54 (1.04-2.27)	2.48 (1.77-3.48)	1.2 × 10−7
Tier 2: Oncogenic relevance (known)
ROS1 ☑6q23-25 +	Splice acceptor	33	Extracellular	0.004	1 (2 ∣ 0) AA, F, EO	67 (16.8-319)	20	19.4 (12.1-31.2)	23.5 (14.99-37)	2.4 × 10−43
	p.F1433S	29		0.07	2 (3 ∣ 1) SMK	12.5 (3.99-39.2)	34	2.25 (1.58-3.2)	2.61 (1.86-3.66)	2.6 × 10−8
PIM1☑ +	p.E124Q	26	Kinase	0.3	3 (3 ∣ 1) SMK	3.28 (1.05-10.3)	84	3.03 (2.4-3.8)	3.04 (2.44-3.77)	1.4 × 10−23
JAK1 ☑ +	p.V651M	31	Kinase	0.1	3 (2 ∣ 3) F	4.60 (1.14-18.6)	45	4.01 (2.8-5.6)	4.04 (2.92-5.59)	4.5 × 10−17
ERBB3☑ +	p.A1131T	28	C-terminal	0.05	2 (3 ∣ 1) AA	18.1 (5.7-56.5)	19	1.67 (1.04-2.68)	2.36 (1.52-3.65)	1.1 × 10−4
ATM ⌧ −	p.D1853V	26	Kinase	0.5	4 (4 ∣ 0)	2.37 (0.98-6.4)	83	1.75 (1.5-2.06)	1.76 (1.50-2.07)	4.6 × 10−12
MLH1⌧ −	p.V180G	27	ATPase	0.005	2(2 ∣ 5)	55 (17.4-260)	3	5.45 (1.7-17.5)	18.6 (7.58-45.7)	1.7 × 10−10
SYNE 1 6q23-25	p.R4673Q	25	Spectrin repeat	0.009	2 (3 ∣ 2) SMK	61.0 (18.3-292)	15	10.2 (4.25-24.5)	19.6 (9.8-39.4)	4.9 × 10−17
	p.S6763L	23		0.02	3 (3 ∣ 1) AA	36.9 (11.8-117)	17	2.99 (1.85-4.85)	4.37 (2.81-6.82)	7.5 × 10−11
LRP1B ⌧	p.S1014P	22	β-propeller	0.002	1 (2 ∣ 4)	46.7 (15.3-310)	1	4.07 (1.04-15.9)	27.9 (10.43-75)	3.6 × 10−11
Tier 3: Functional impact (novel)
COL6A 3	p.R2251W	33	C-terminal	0.024	3 (3 ∣ 1) AA	37.6 (11.8-115)	83	18.4 (14.4-23.5)	18.9 (14.94-24.1)	1.9 × 10−129
CHD2 ⌧	p.K1245Nfs*4	33	Helicase	0.008	5 (6 ∣ 0)	55 (22.9-237)	22	77.4 (15.4-430)	68.2 (39.9-165)	1.8 × 10−48
GALNT6 +	p.R195X*	36	Catalytic	0.0008	2(4∣ 1)	67 (20-339)	9	77.5 (26.9-223)	68.2 (39.9-165)	1.3 × 10−47
MUC4 +	p.D4327Tfs*100	33	Tandem repeat	0.0007	2 (2 ∣ 3) SMK	78 (25-461)	2	9.88 (2.4-40.6)	62.9 (27.7-175)	4.4 × 10−29
WNK1 +	p.K583Sfs*11	33	Kinase	0.008	4 (4 ∣ 0) SMK	52 (19-272)	23	52.3 (13.09-408)	68.2 (33.1-181)	5.9 × 10−27
KIF26B +	p.R174W	24	Motor	0.008	2 (3 ∣ 0)	55.2 (18.3-174)	2	4.90 (0.9-40.3)	20.7 (8.45-50.8)	3.6 × 10−11
	p.V865I	26		0.015	2 (4 ∣ 0)	37 (13.5-210)	6	2.84 (1.2-6.7)	11.6 (6.05-22.3)	1.7 × 10−13

^⚠Pathogenic (ClinVar)

^☑Known proto-oncogenes

^⌧Known tumor suppressor

⁺GoF

⁻LoF.

^aDenotes a stop codon (Ter), indicating translation termination due to stop-gain and fs indels.

^bCADD score ≥ 30 ranks in the top 0.1% most deleterious in the human genome, indicating strong potential for functional impact.

^cFixed-effect meta-analysis was conducted to combine information from studies. Empirical Bayes shrinkage applied to stabilize effect size estimates for ultra-rare variants with sparse counts.

AA, African American; CADD, combined annotation-dependent depletion; CI, confidence interval; EO, early onset LC (<55 y old); F, female; FLC, familial lung cancer (n = 177); fs, frameshift; GELCC, Genetic Epidemiology of Lung Cancer; gnomAD, genome aggregation database v.4 (n = 807,162); GoF, gain of function; indels, insertions and deletions; LoF, loss of function; MAF, minor allele frequency (shown as %); OR, odds ratio; SLC, sporadic lung cancer (n = 10,085); SMK, smoker.

Among the top hits, six variants were classified as pathogenic in ClinVar and designated as Tier 1. (1) Laminin Subunit Alpha-2 (LAMA2) p.R2383X: a stop-gain variant disrupting the laminin-211 complex, pathogenic for congenital muscular dystrophy (ClinVar ID 14300). This variant also had the highest CADD score of 44. Two additional missense variants in LAMA2, p.G2472V and p.A3054E, were also observed. (2). PRKN p.R275W: A missense variant affecting the RING1 catalytic core of the E3 ligase. ClinVar classifies it as pathogenic for both FLC and Parkinson’s disease (ClinVar ID 7050). (3) Filaggrin (FLG) p.R3009X: a stop-gain variant truncating the protein by 1053 amino acids. It is reported as pathogenic for atopic dermatitis and allergic disorders (ClinVar ID 1207438). Two additional ultra-rare truncating variants in FLG, p.Y3105X and p.R3404Gfs*51, were also identified and are likely to disrupt protein integrity. (4) Additional Sex Combs Like 1 (ASXL1) p.G646Wfs*12: a truncating variant affecting the conserved N-terminal region essential for chromatin remodeling. It is reported as pathogenic for myeloid malignancies (ClinVar ID 426927). (5) BRCA2 p.N986Kfs*2: A frameshift variant truncating variant resulting in premature termination and loss of all known functional domains. It is classified as pathogenic for familial breast-ovarian cancer (ClinVar ID 51378). This variant is extremely rare, observed in only three of 807,162 individuals in gnomAD. (6) Cystic Fibrosis Transmembrane Conductance Regulator (CFTR) p.R74W: a missense variant in the membrane-spanning domain, associated with drug response (ClinVar ID 196277).

Tier 2 variants include those with known oncogenic relevance, such as proto-oncogenes (ROS1, PIM1, JAK1, and ERBB3) and tumor suppressors (ATM, MLH1, and LRP1B). Both ROS1 and SYNE1 are located within the known FLC linkage region 6q23-25. Five variants mapped to protein kinase domains, impacting kinase activity and known oncogenic hotspots: ATM p.D1853V and p.S2146T, MLH1 p.V180G, CHD2 p.K1245Nfs*4, JAK1 p.V651M, and PIM1 p.E124Q.

Tier 3 variants represent novel candidates with strong predicted functional impact. The most prominent signal was observed for Collagen Type VI Alpha 3 Chain (COL6A3) p.R2251W, a missense variant with a CADD score of 33. Four ultra-rare (MAF < 0.001%) protein-truncating variants with CADD score more than or equal to 30 were identified: Chromodomain Helicase DNA Binding Protein 2 (CHD2) p.K1245Nfs*4, WNK Lysine Deficient Protein Kinase 1 (WNK1) p.K583Sfs*11, GalNAc Transferase 6 (GALNT6) p.R195X, and Mucin 4 (MUC4) p.D4327Tfs*100.

Dose Effect, Gene-Environment Interaction, and Multiple Candidates in HRFLC Families

We assessed the dose effect of the 28 top candidate variants and observed a strong dose-dependent association between the number of variant alleles and LC risk, although sample sizes were limited. Compared with non-carriers (zero risk alleles), individuals carrying one or more than or equal to two risk alleles had ORs of 20.26 and 113.9, respectively (Supplementary Table 3). A similar trend was observed when combining FLC and SLC cases, with ORs of 3.06 and 15.20 for individuals carrying one or more than or equal to two risk alleles.

We next evaluated the distribution of these variants within GELCC families. In total, 101 individuals carried at least one candidate variant, including 61 FLC cases (one with both lung and breast cancers) and 40 relatives (seven with other cancers). Among these, 20 individuals were “multi-hit” carriers, harboring two or more candidate variants (Table 3). This subgroup comprised 15 FLC cases, three relatives with other cancers, and two unaffected relatives.

Table 3.

Summary of 20 Multi-Hit Carriers in the GELCC Study

Family		Individual Characteristics			Candidates
Site ID	n. FLC per Family	Subject	Affection Cancer Status	Agea ∣Sex∣PPY	n. Harbored: Candidate Genesb
European American
CIN04012	8	1	FLC	63∣M∣92	3: LAMA2, SYNE1, MLH1
		21	Head and neck cancer	61∣M∣135	2: LAMA2, MLH1
CIN04116	8	18	Unaffected relative	43∣F∣29	2: LRP1B, MUC4
		32	Breast cancer	65∣F∣41	2: LRP1B, MUC4
MAYO3821	3	12	FLC	78∣F∣36	2: LAMA2, GALNT6
		13	FLC	69∣F∣23	2: LAMA2, GALNT6
MAYO1645	5	14	FLC	58∣F∣54	2: PIM1, ATM
CIN11180	10	6	FLC with breast cancer	71∣F∣66	2: SYNE1, ASXL1
MCO432	9	180	FLC	n.a.∣M∣64	2: PRKN, ATM
MCO266	3	101	FLC	n.a.∣M∣9	2: ROS1, CHD
African American
MCO721	7	102	FLC	35∣F∣0	6: ROS1, SYNE1, COL6A3, BRCA2, FLG (R3009X, R3404Gfs*51)
		101	FLC	48∣F∣2	4: ROS1, PRKN (Q34R, R275W), BRCA2
KCI419	5	1	FLC	58∣M∣n.a.	3: PRKN, FLG, CFTR
KCI110	9	19	FLC	77∣M∣42	3: FLG (R3009X, Y3105X, R3404Gfs*51)
		46	Unaffected relative	63∣M∣0	3: FLG (R3009X, Y3105X, R3404Gfs*51)
KCI464	3	1	FLC	51∣F∣0	3: FLG (R3009X, Y3105X, R3404Gfs*51)
MCO643	4	101	FLC	60∣F∣2	2: FLG, CFTR
MCO273	5	101	FLC	71∣M∣19	2: ROS1, CFTR
		103	Other unspecified cancer	48∣M∣55	3: ROS1, FLG, CFTR
LSU2011	4	65	FLC	46∣F∣0	2: ROS1, PRKN

^aFour FLC cases and one relative with another cancer were diagnosed at an early age (<55 y).

^bGenes in bold contain two or more candidate variants.

F, female; FLC, familial lung cancer; GELCC, Genetic Epidemiology of Lung Cancer; M, male; n.a., not available; PPY, smoking pack-years.

Compared with non-carriers, carriers were more likely to be of African ancestry (42% versus 10%), female (53% versus 45%), and diagnosed with early onset LC (22% versus 13%). Multi-hit carriers had even more distinct patterns, with a disproportionate representation of African ancestry (n = 10) and early onset cases (n = 5, all female). Ancestry-specific trends were evident in CFTR, COL6A3, ERBB3, and FLG, where variants occurred predominantly among individuals of African ancestry (Table 2). Notably, several variants, including ROS1, PRKN, FLG, COL6A3, and BRCA2, were detected in individuals without tobacco exposure. In contrast, seven variants (e.g., LAMA2, PIM1, WNK1, MUC4) were found exclusively in smokers.

Although some variants were family specific (i.e., LRP1B, BRCA2), most recurred across multiple families, particularly in the 6q23-25 region (PRKN, SYNE1, LAMA2, and ROS1). The multi-hit carriers were particularly evident in three families: (1) MCO721 (African ancestry; seven affected): two early onset female FLCs carried multiple candidates, including BRCA2, COL6A3, FLG, ROS1, and PRKN; (2) CIN04012 (European ancestry; eight affected): two male heavy smokers (one FLC case, one relative with prostate cancer) carried MLH1 and LAMA2, along with SYNE1; (3) MAYO3821 (European ancestry; three affected), two female smoker FLCs carried GALNT6 and LAMA2 p.G2472V variants (Table 3).

Gene-Based Association and Network Analysis

Gene-based collapsing tests of 53,306 rare variants that passed the three-tier filtering pipeline identified three additional established cancer-related genes, including Retinoic Acid Receptor Beta (RARB), O6-methylguanine-methyl transferase (MGMT), and Early B-Cell Factor 1 (EBF1; Table 4 and Supplementary Table 4). Among the 20 candidate genes harboring the 28 top variants from the discovery and validation analyses discussed previously, the most significant associations were observed for KIF26B, ROS1, PRKN, and CHD2.

Table 4.

Gene-Based Association Collapsing Tests

Genea	n. Rare, Deleterious Variants, per Gene	n. Carriers in FLC∣ILCCO Controls (177 vs. 885)	CMC Test, FDR-Adjusted p Value
RARB	32	44∣0	2.7 × 10−31
MGMT	14	26∣0	4.4 × 10−25
EBF1	32	52∣1	2.7 × 10−20
KIF26B	14	20∣1	4.7 × 10−13
ROS1	15	24∣12	4.7 × 10−11
PRKN	13	25∣6	4.9 × 10−11
CHD2	9	16∣2	1.2 × 10−10
SYNE1	34	56∣74	1.1 × 10−9
LAMA2	19	26∣22	2.2 × 10−9
PIM1	5	14∣2	2.2 × 10−8
COL6A3	21	30∣38	2.9 × 10−8
FLG	9	47∣35	2.3 × 10−7
MUC4	37	66∣116	3.3 × 10−6
WNK1	4	8∣2	1.7 × 10−5
GALNT6	3	6∣0	2.0 × 10−5
CFTR	17	25∣40	0.0005
LRP1B	22	32∣74	0.0006
MLH1	6	14∣19	0.0056
BRCA2	5	8∣11	0.0073
ATM	7	12∣27	0.0259
ERBB3	5	10∣21	0.0262
JAK1	3	4∣6	0.0688
ASXL1	5	8∣22	0.1385

^aOnly rare, potential functional variants that passed the three-tier filtering pipeline in the discovery were included in the burden test (53,306 rare variants across 6049 genes). The full results are presented in Supplementary Table 2.

CMC, combined multivariate and collapsing; FDR, false discovery rate; FLC, familial lung cancer; GELCC, Genetic Epidemiology of Lung Cancer; ILCCO, International Lung Cancer Consortium.

Protein–protein interaction analysis of these 23 genes revealed a highly interconnected network (p value, 4.9 × 10⁻¹⁴; Fig. 1 and Supplementary Table 5). The core module comprises four DNA damage response (ATM, BRCA2, MLH1, and MGMT). The strongest and experimentally determined interactions (score > 0.90) include MUC4–ERBB3–JAK1, ATM–BRCA2–MLH1, and MGMT–MLH1. Other moderate interactions (score > 0.5) were observed among MUC4-GALNT6, BRCA2-MGMT, BRCA2-ROS1, LRP1B-SYNE1, and COL6A3–LAMA2. GO/KEGG analyses highlighted DNA repair and cancer pathways, especially adenocarcinoma (Supplementary Table 6).

Figure 1.

Protein-protein interaction network of top candidate genes. The predicted protein-protein interaction network for the 23 candidate genes based on STRING (Search Tool for the Retrieval of Interacting Genes/Proteins). The strongest and experimentally determined interactions (score > 0.90) include MUC4-ERBB3-JAK1, ATM-BRCA2-MLH1, and MGMT-MLH1. Other moderate interactions (score > 0.45) include MUC4-GALNT6, BRCA2-MGMT, BRCA2-ROS1, LRP1B-SYNE1, COL6A3-LAMA2, and PRKN-CFTR-WNK1. Edge colors indicate the source of interaction evidence: experimental data (purple), automated text mining (green), and curated databases (blue).

Discussion

GELCC’s longstanding efforts have enabled major advances in understanding the genetic basis of FLC. Using this resource, we identified 28 rare, high-penetrance variants across 20 genes, including novel candidates (COL6A3, GALNT6, MUC4) and established cancer genes (ATM, BRCA2, MLH1) and known FLC linkage loci (PRKN, LRP1B). Dose-effect and co-occurrence analyses across multiple families suggest a polygenic model of inherited risk. Several candidate variants were preferentially enriched among smokers, suggesting gene–environment interactions that may heighten cancer risk in the presence of tobacco exposure. These findings emphasize the importance of smoking cessation, particularly for individuals carrying high-penetrance germline variants.

Among the most compelling biological signals identified in this study is the disruption of the GALNT6–MUC4–ERBB3 axis, a pathway central to mucin-type O-glycosylation, epithelial integrity, and oncogenic signaling. GALNT6 (OMIM 605148) encodes a GalNAc-T6 that catalyzes the first step of mucin-type O-glycosylation and is frequently upregulated in LC.^66-69 Aberrant glycosylation, a hallmark of cancer, promotes progression.^69-78 Clinically, altered glycosylation of cancer glycoconjugates is widely used in biomarker development (e.g., CA-125, CA-15–3, CA-19–9).^71,74,79-81 GALNT6 also modulates immune checkpoints (e.g., PD-L1, TIM-3), contributing to immune evasion and therapeutic resistance,^82-88 supporting the emerging concept of “glyco-immune checkpoints” in cancer immunotherapy.^89-96

MUC4 (OMIM 158372), a heavily glycosylated transmembrane mucin, is essential for maintaining airway epithelial barrier function. When dysregulated, however, it forms oncogenic complexes with ERBB2, stabilizes EGFR/ErbB3 heterodimers, and activates MAPK and PI3K/AKT signaling.^97-108 MUC4 also interacts with EGFR and FGFR1,¹⁰⁹ further promoting immune evasion and tumor progression.^110,111 Clinically, MUC4 overexpression distinguishes lung adenocarcinoma from squamous cell carcinoma and correlates with poor prognosis.^97,112-115 ERBB3 (HER3, OMIM 190151) functions as a pivotal co-receptor that amplifies oncogenic signaling through dimerization with other EGFR family members. Although germline ERBB3 variants remain understudied, notable examples include p.I649R identified in a FLC family,¹¹⁶ and p.R1127H in never-smoker patients with LC.¹¹⁷ Together, disruption of GALNT6, MUC4, and ERBB3 may lead to defective mucin glycosylation, drive lung carcinogenesis, and increase inherited LC risk. Future studies should confirm whether mucin glycosylation is altered in lung tumors with these mutations and characterize the tumor immune microenvironment to clarify the functional consequences of this axis. This pathway represents a promising source of diagnostic biomarkers and glyco-immune checkpoints therapeutic targets, with potential to complement existing EGFR/ALK-targeted therapies and immunotherapies.

Correlated with this axis, another key finding was oncogenic kinase domain variants, including JAK1 p.V651M and PIM1 p.E124Q. Both variants were located within the tyrosine kinase domain, likely promoting constitutive kinase activity and dysregulating JAK/STAT and NF-κB pathways. JAK1 (OMIM 147795) plays a central role in antitumor immunity, particularly in mucin-rich LC, where hyperactivation may intensify tumor growth and immune evasion.¹¹⁸ In line with our results, the JAK1 V651M mutation has been previously detected in children with Down syndrome acute lymphoblastic leukemia and in prostate carcinomas.^119-121 PIM1 (OMIM 164960) overexpression is also known to drive LC progression and therapy resistance.^122-125 A recent clinical study revealed that JAK inhibition plus PD-1 blockade improved outcomes in LC.¹²⁶

We also identified several novel candidate genes with potential roles in mucosal defense, epithelial barrier integrity, chromatin remodeling, and kinase regulation. FLG (OMIM 135940) and COL6A3 (OMIM 120250) are critical for epithelial structure and mucosal protection.^127-129 COL6A3 also regulates apoptosis, oxidative stress, metabolism, and proliferation. A Chinese case–control study linked the intronic variant rs115510139 in COL6A3 to increased LC risk, particularly in younger males.¹³⁰ CHD2 (OMIM 602119) and ASXL1 (OMIM 612990) regulate chromatin remodeling. ASXL1, frequently mutated in clonal hematopoiesis, interacts with BAP1 to regulate histone modification and engages the AKT–WNK1 signaling.^131-133 CHD2 is essential for chromatin accessibility, and its disruption may contribute to genomic instability.¹³⁴ In kinase regulation, WNK1 (OMIM 605232) plays a role in ion transport, cell signaling, and tumor progression,¹³⁵ and interacts with CFTR (OMIM 602421), consistent with our interaction analysis. Functional studies show that WNK1 suppresses CFTR activity in a kinase-dependent manner.¹³⁶ Reduced CFTR function has been implicated in lung epithelial barrier dysfunction and may promote lung tumorigenesis.^137,138

We replicated PRKN findings within our previously reported FLC linkage region at 6q23-25. Our analysis identified both a previously reported (p.R275W)^15,16 and a novel (p.Q34R) missense variant. We also uncovered three novel risk genes in this region: LAMA2 (OMIM 156225), which is involved in extracellular matrix structure^139-142; SYNE1 (OMIM 608441), a nuclear envelope protein essential for genome integrity; and the kinase oncogene ROS1 (OMIM 165020). Although somatic ROS1 fusions (e.g., CD74–ROS1) are established clinically actionable drivers in 1% to 2% of patients with LC, germline variants have rarely been reported. In our study, recurrent germline ROS1 variants (p.F1433S, p.N785S) in multiple early-onset, nonsmoking FLC families suggest that ROS1 may act as an LC susceptibility gene. ROS1-mutant tumors also exhibit enhanced DNA damage response and increased sensitivity to immune checkpoint inhibitors.¹⁴³ We also confirmed a FLC linkage locus at 2p12.2 LRP1B (OMIM 608766), frequently inactivated in LC,^144-146 and among the most frequently altered genes across cancers.¹⁴⁷ Its loss is associated with genomic instability and has been linked to improved responses to immune checkpoint inhibitors, reinforcing clinical and therapeutic relevance.¹⁴⁸

Remarkably, we identified a pathogenic BRCA2 frameshift, p.N986Kfs*2, that disrupts RAD51 recruitment by eliminating BRC repeats essential for homologous recombination. This variant has been reported in familial breast-ovarian cancer and other hereditary cancer syndromes.^149-158 We also detected missense variants in other DNA repair genes, including ATM (p.D1853V, p.S2146T) and MLH1 (p.V180G), the latter typically associated with Lynch syndrome. These findings reinforce the contribution of germline defects in DNA repair pathways to LC susceptibility, aligning with previous reports.^48-51,54 Such defects may confer therapeutic vulnerabilities: patients harboring these variants often show increased sensitivity to poly (ADP-ribose) polymerase (PARP) inhibitors, as observed in BRCA-deficient cancers.^159-161

Our gene-based collapsing analysis highlighted three genes, MGMT, RARB, and EBF1, with established roles in cancer biology. MGMT, a DNA repair enzyme often silenced by promoter methylation, contributes to LC development and therapy resistance.^162-164 RARB, a tumor suppressor, regulates cellular differentiation through retinoid signaling,^165-167 while EBF1 promotes tumor progression via cell cycle and p53 pathway regulation in solid tumors.^168-170 While biologically plausible, these gene-level rare variant burden signals should be interpreted with caution, particularly in highly polymorphic genes or when driven by singletons (variants observed only once).

A major strength of this study is its focus on individuals with extreme phenotypes, specifically large multigenerational families with multiple affected members. These pedigrees are enriched for rare and ultra-rare, high-penetrance variants, offering over fivefold greater statistical power than conventional population-based studies.²⁸ The identification of significant candidate variants highlights the value of this approach in uncovering inherited risk factors. However, several challenges remain due to differences between family-based and population-based designs. In the discovery, within-family segregation analysis was limited by the absence of key affected individuals, many of whom were deceased prior to biospecimen collection due to the aggressive nature of LC. To overcome this, we leveraged external control data sets and validated findings in large biobanks. While this strategy enhanced statistical power, demographic differences, such as the higher African American representation in AoU, lower in UKB, and markedly higher smoking prevalence among FLC cases, along with missing smoking data in validation sets, introduce heterogeneity and limit adjustment for smoking as a confounder, potentially attenuating genetic signals from high-risk families. Moreover, filtering strategies that prioritize rare variants based on allele frequencies in gnomAD may enrich for variants specific to FLC cases, limiting generalizability to population-level screening.

Going forward, we will expand sequencing in ancestry-diverse, multigenerational pedigrees to improve representation and segregation power. Given that several pathogenic germline variants (e.g., BRCA2, MLH1, ROS1, LAMA2, SYNE1) predispose to multiple cancers, we will assess their co-segregation with other cancers to inform prevention strategies beyond LC in high-risk families. In parallel, functional validation using CRISPR-based gene editing, immune profiling, and somatic variant analysis from tumor samples is planned to elucidate the biological impact of prioritized variants.

In summary, our findings provide compelling evidence that rare, inherited variants substantially contribute to FLC susceptibility, candidate genes involved in DNA repair, cancer driver genes, and an MUC4–GALNT6–ERBB3–JAK1 signaling axis. Although current genetic testing in LC focuses on somatic alterations, these germline (inherited) variants, if validated, could profoundly advance the translation of rare variant polygenic risk models, inform precision screening strategies (e.g., low-dose CT), guide genetic counseling and testing for patients and families, and enable targeted smoking cessation interventions, addressing a critical gap in routine clinical care.

Supplementary Material

Note: To access the supplementary material accompanying this article, visit the online version of the Journal of Thoracic Oncology at www.jto.org and at https://doi.org/10.1016/j.jtho.2025.12.004.

Data Availability

Data from this study are available through NCBI dbGaP: Genetic Epidemiology of Lung Cancer Consortium (GELCC, phs000629.v1.p1), International Lung Cancer Consortium (ILCCO, phs000878.v2.p1), and The Cancer Genome Atlas (TCGA, phs000178.v9.p8). The U.K. Biobank whole-exome sequencing data are accessible via the U.K. Biobank Research Analysis Platform to researchers with an approved project. All of Us whole-genome sequencing data are available through the All of Us Researcher Workbench, a secure cloud-based platform, to researchers from eligible organizations on completion of the required training and authorization process.

Code Availability

All analyses were performed using standard publicly available software. Any specific analysis code details are available from the authors on request.