Real use case

Data

We ran the pipeline using about 200K samples from UKBiobank. We filtered for British to remove outliers (150K after filtering)

Gene scoring

We calculated the scored for the 150K samples.

MAF threshold:

1% MAF threshold.

Weighting function:

beta 1-25 weighting parameter.

Functional annotation:

CADD raw scores.

Association analysis

For each phenotype, we performed association analysis including all the samples (150K)

Method:

linear regression

Covarites:

age, sex, BMI and PC1-4

Prediction models

We also generated 3 models for each phenotype, a PRS prediction model, a gene-based prediction model and a combined model.

Feature Selection:

we used portion of the samples (50K) for feature selection with linear regression. The genes with p-values <0.05 were selected as features.

Covariates:

age, sex, BMI and PC1-4 were also included in the features.

model trainng:

Of the remaining 100K samples, 75% were used in training with 10-fold cross-validation.

model testing:

the 25% remaining of the dataset was used as external testing set for final model evaluation.

PRS calculation:

To avoid colinearty, we excluded variants that were included in the gene scores from the PRS calculation.

LDL Phenotype

We performed the analysis on the samples with LDL direct measurements as phenotype(quantitative).

Note

For this phenotype we adjusted the values for individuals who take statin.

Association analysis

The association analysis highlighted PCSK9 and LDLR as significant genes, both are known to be associated with LDL. QQ-plot and Manhattan plot are presented below.

The QQ-plot:

The Manhattan plot:

Regression model

For the prediction model, we used LDL direct measurements (adjusted for statin) as target. For features, we used the scores of 3 selected genes + BMI + age + sex + PC1-4. For the PRS and combined models we used the following PRS (PGS000688). The final prediction models was generated using gradiant boosting regression, evaluation metric are shown in the table below.

	Gene-based model	PRS model	Combined model
Rˆ2	0.092	0.322	0.321
RMSE	0.849	0.729	0.725

The images below are the output of the final combined model. Feature importance plot:

Actual vs Predicted:

Model residuals:

Alkaline phosphatase

We performed the analysis on the samples with ALP measurements as phenotype(quantitative).

Association analysis

We used linear regression for the analysis and age, sex, BMI and PC1-4 were used as covaraites. The association analysis highlighted ALPL, GPLD1 and ASGR1 as significant genes, all of which are known to be associated with alkaline phosphatase. QQ-plot and Manhattan plot are presented below.

The QQ-plot:

The Manhattan plot:

Regression model

For the prediction model, we used alkaline phosphatase measurements as target. For feature selection, For features we used 45 selected genes as features (45 genes) + BMI + age + sex + PC1-4. For the PRS and combined models we used the following PRS (PGS000670). The final prediction models was generated using gradiant boosting regression, evaluation metric are shown in the table below.

	Gene-based model	PRS model	Combined model
Rˆ2	0.084	0.255	0.281
RMSE	24.7	22.3	21.9

Feature importance plot for combined model:

Other phenotypes

Association analysis

biomarker	Top 3 significantly associated genes
Alanine aminotransferase	GPT, THRA, ACVR2B, More
Albumin	FCGRT, ALB, IQGAP2, More
Alkaline phosphatase	ALPL, GPLD1, ASGR1, More
Apolipoprotein A	ABCA1, LIPG, LCAT, More
Apolipoprotein B*	PCSK9, LDLR, NKPD, More
Aspartate aminotransferase	GOT1, GABRA5, THRA, More
Cholesterol*	PCSK9, LDLR, ABCA1, More
C reactive protein	CRP, PTGES3L, SLN , More
Creatinine (in serum)	NAA20, PRAMEF19, CLIC4, More
Gamma glutamyltranferase	GGT1, CCL1, RORC
Glucose	G6PC2, GCK, DYNLL1
Glycated Haemoglobin (HbA1C)	HBB, PIEZO1, GCK
Lipoprotein A	LPA, PLG, MRPL18
Triglycerides	APOA5I, APOC3, PLA2G12A

• values adjusted for statin

Note

summary statistics for biomarkers association analysis will be added soon.

Prediction models

Here we show a table of other phenotypes that we analyzed. For each phenotype we include the number of genes considered in the models as well as the Rˆ2 of the gene-based model, PRS model and combined model.

	Number of genes	Gene-based model	PRS model	Combined model
apolipoprotein a	6	0.227	0.413	0.403
apolipoprotein b*	5	0.059	0.267	0.269
aspartate aminotransferase	57	0.039	0.124	0.128
Cholesterol*	6	0.088	0.229	0.236
Creatinine	128	0.228	0.454	0.448
Hba1c	13	0.100	0.242	0.247
lipoprotein a	3	0.004	0.582	0.603
Triglyceride	5	0.143	0.316	0.315
urea	2	0.074	0.173	0.179
Urate	4	0.396	0.521	0.534

• values adjusted for statin