The brain is a highly heritable organ, and the field of imaging genetics has a big potential. However, it has been difficult to determine how gene variants affect brain phenotypes. We have been involved in the development of novel brain imaging analytical approaches to improve genotype-phenotype mapping of the human brain. Further, we have developed methodology to improve analysis of cross-ethnic samples, with implications for both research as well as translation to clinical use.Although the molecular processes for neurodevelopment seldom occur in discretized sets, the imaging genetic analyses often rely on measures of morphologically defined region of interests (ROI). As the morphological variations are not confined by the anatomical landmarks, arbitrarily defined ROI can sometimes introduce bias, as we have shown using 3D modeling of human brain based on genetic ancestry (Fan et al., 2015).We have recently increased the size of our neuroimaging training data to ~5K individuals. This has enabled us to extract reliable imaging features pertaining to genetic variations other than the traditionally defined ROI. Here we demonstrate our recent studies on learning new endophenotypes and its utilities in discovering the genetic influence on neuroanatomy, and show evidence of improved signal to noise ratio as well as better replication rate.The learnt manifolds of anatomical structure can either be used for correcting the bias, e.g. population structure, or improving the power to detect genetic effects. This is important for extending imaging genetics research beyond Caucasian samples.