Background: Large-scale phenotyping for detailed milk fatty acid (FA) composition is difficult due to expensive and time-consuming analytical techniques. Reliability of genomic prediction is often low for traits that are expensive/difficult to measure and for breeds with a small reference population size. An effective method to increase reference population size could be to combine datasets from different populations. Prediction models might also benefit from incorporation of information on the biological underpinnings of quantitative traits. Genome-wide association studies (GWAS) show that genomic regions on Bos taurus chromosomes (BTA) 14, 19 and 26 underlie substantial proportions of the genetic variation in milk FA traits. Genomic prediction models that incorporate such results could enable improved prediction accuracy in spite of limited reference population sizes. In this study, we combine gas chromatography quantified FA samples from the Chinese, Danish and Dutch Holstein populations and implement a genomic feature best linear unbiased prediction (GFBLUP) model that incorporates variants on BTA14, 19 and 26 as genomic features for which random genetic effects are estimated separately. Prediction reliabilities were compared to those estimated with traditional GBLUP models. Results: Predictions using a multi-population reference and a traditional GBLUP model resulted in average gains in prediction reliability of 10% points in the Dutch, 8% points in the Danish and 1% point in the Chinese populations compared to predictions based on population-specific references. Compared to the traditional GBLUP, implementation of the GFBLUP model with a multi-population reference led to further increases in prediction reliability of up to 38% points in the Dutch, 23% points in the Danish and 13% points in the Chinese populations. Prediction reliabilities from the GFBLUP model were moderate to high across the FA traits analyzed. Conclusions: Our study shows that it is possible to predict genetic merits for milk FA traits with reasonable accuracy by combining related populations of a breed and using models that incorporate GWAS results. Our findings indicate that international collaborations that facilitate access to multi-population datasets could be highly beneficial to the implementation of genomic selection for detailed milk composition traits.