These prediction analyses of measured total and regional fat mass confirmed that BMI, based on weight and height, and waist size measurements together predict total body fat very well (R2
85%). However, we found that measures of abdominal and intra-abdominal (visceral and liver) fat were predicted less optimally by these anthropometric variables and that the addition of adiposity-associated biomarkers improved their predictions. About half of the variation in abdominal adiposity was predicted by anthropometry, with the prediction of this variability further improved by adding the top 5 predictors from the Random Forest biomarker model (R2
0.53 to 0.58). The prediction of visceral fat also improved slightly (R2
0.65 to 0.68) by adding the top 5 biomarker predictors. The largest contribution from the biomarker model was observed for the prediction of liver fat, for which R2
increased from 29% with the anthropometry model to 44% with the model that also included the top 5 biomarkers. Blood adipokines (leptin, leptin-adiponectin ratio, sLEPR, PAI1) contributed to the prediction of both total and regional fat. Other top predictors included markers of insulin resistance and the IGF pathway (insulin, IGFBP1, uric acid), sex hormones (free estradiol, SHBG), lipid-soluble micronutrients (vitamin D3
, lycopene, CoQ10, alpha-tocopherol) and markers of inflammation (CRP).
It is well established that adipose tissues are active endocrine organs, with each regional depot having intrinsic secretory profiles 
. Thus, blood concentrations of depot-specific adipocyte-derived biomarkers and their metabolites may reflect relative body fat distribution and also contribute to associated metabolic risks. Metabolic syndrome has been associated more with abdominal fat than total or gluteofemoral fat 
, and more with visceral fat compared to abdominal subcutaneous fat 
. Accordingly, in past studies, certain circulatory markers have shown a strong association with visceral fat specifically (Table S1
), including low blood levels of adiponectin 
and SHBG 
, and high levels of PAI1 
, visfatin 
, systemic inflammatory markers 
, insulin 
and free estradiol 
. Also, liver fat has been associated with blood levels of liver enzymes 
, insulin and sLEPR 
, adiponectin 
, PAI1 
, fetuin A 
, retinol binding protein-4 (RBP4) 
, and free fatty acids 
. Our study included most of these biomarkers associated with regional adiposity.
There have been few published studies that have attempted to optimally predict body composition with a comprehensive list of biomarkers. In a study of 56 middle-aged and 20 older adults who were healthy but overweight, 124 proteins in fasting blood analyzed with a Luminex multiplex assay were tested for their prediction of BMI using Random Forest modeling 
. Similar to our study, the candidate markers were selected a priori
, based on their association with chronic diseases, inflammation, endothelial function and metabolic signaling. BMI was best predicted, positively, by leptin, complement 3 (C3), CRP, amyloid P and vascular endothelial growth factor, and, negatively, by IL-3, IL-13 and apolipoprotein A1. In another study of 20 postmenopausal women, DXA-based percent lean body mass was predicted by fasting blood levels of 90 cytokines analyzed with a Luminex multiplex assay 
. Random Forest modeling identified 7 top predictors of percent lean mass (serum leptin, adiponectin, insulin, C3, amyloid P, growth hormone, eotaxin) and discriminated high vs. low lean mass groups with less error (mean error
5.0%) compared to an alternative Recursive Partitioning model (mean error
Our findings support the contention that adding key biomarkers to usual anthropometric variables may enhance the prediction of body fat distribution patterns when reference imaging-based methods are not practical, such as typically in large epidemiologic studies. Past studies that compared anthropometric measures to imaging of fat topography observed a good correlation between anthropometry and total fat mass 
but detected lower correlations for intra-abdominal fat distribution 
. Our study results are consistent with this literature.
Certain biomarkers performed far better than others in predicting specific adiposity, such as leptin for total fat, lycopene, leptin-adiponectin ratio and leptin for visceral fat, and insulin and SHBG for liver fat (). We did not observe one or two predominantly strong predictors for abdominal fat like we did for the other adiposity measures. Leptin, a well-established indicator of total adiposity, also predicted visceral fat, together with leptin-adiponectin ratio, which may independently reflect leptin resistance due to excess intra-abdominal adiposity 
. Insulin resistance markers (insulin, HOMA-IR, HOMA-beta) were consistently among the most important predictors of visceral fat and liver fat, although we included only insulin in the final model due to their high correlations. These results are consistent with the notion that visceral fat carries a greater metabolic risk than subcutaneous fat by inducing fatty acid drainage into the liver through the portal venous system, which then may impair insulin/glucose homeostasis 
. Endogenous synthesis of estrogen from androstenedione in adipocytes is known to be particularly active in the subcutaneous adipose tissue, whereas visceral fat and subsequent increase in liver fat may interfere with the production of SHBG 
. This is also consistent with our findings, where blood levels of bioactive free estradiol were shown to predict total adiposity (mostly subcutaneous fat) and SHBG predicted hepatic adiposity.
CRP ranked high for predicting visceral adiposity. However, in contrast to previous studies 
, other common markers of systemic inflammation were either mostly undetectable (TNFα) or showed only modest to low predictive ability for total adiposity (IL6). This may be because our study participants were mostly healthy adults who were non-diabetic and without overt low-grade inflammation. Lipid-soluble micronutrients, especially D vitamers, also showed prediction capacity for abdominal, visceral and hepatic adiposity, as noted before 
A key strength of the present study is the implementation of Random Forest modeling. The use of stepwise linear regression to screen biomarkers resulted in over-fitting of the training data (leading to many predictors in the final model and a R2
>95%), with a low predictive R2
in the testing data, in our analysis (data not shown), as well as in past studies 
. The tree-based Random Forest modeling also allowed the incorporation of potentially important interactions among predictors. This is the first time that this analytic approach was used to predict detailed, imaging-based regional body fat measurements. The study limitations include a relatively small sample size and the possibility that potential confounders were not accounted for. Also, there may be other (as yet unidentified) biomarkers that could substantially improve the predictions. Replications in larger datasets are warranted, especially to compare the prediction performance of biomarkers in men and across ethnic groups with varying body fat distribution. In this sample of Caucasian and Japanese American women, ethnicity was an important determinant of fat distribution 
. Interestingly, it did not remain an important predictor after accounting for anthropometry and the biomarker predictors.
In summary, we provide preliminary evidence that supports the utility of measuring key blood biomarkers to improve the performance of usual anthropometric variables in predicting abdominal, visceral and liver fat. Discovery of additional biomarker predictors and generalization of this research to other populations may allow for the development of accurate prediction models for specific body fat compartments. Such prediction equations may be very useful in predicting risk of obesity-associated diseases at the individual and population levels.