A functional smoothed machine learning approach for high-dimensional orofacial clefts genetic data analysis project summary orofacial clefts (OC) are one of the most challenging newborn defects that…

The National Institute of Dental and Craniofacial Research awarded the University of New Hampshire $299,950 on August 19, 2026, under the Oral Diseases and Disorders Research program (CFDA 93.121) to develop a functional smoothed deep learning approach combining deep neural networks and functional data analysis techniques for analyzing high-dimensional genetic data associated with orofacial clefts and predicting disease risk. The research will build and validate a functional smoothed deep learning (FSDL) method using functional data analysis techniques to improve disease risk prediction for orofacial clefts based on high-dimensional genetic data, with capability to capture dependent structures among genetic variants and handle both gene sequencing data and conventional covariates such as demographic data. The project will employ larger-scale orofacial clefts databases for training and validation. The award supports work at the University of New Hampshire in New Hampshire with a performance period from August 19, 2026, through August 18, 2028. The research is led by an early-stage new investigator with a multidisciplinary team experienced in statistics, computer science, biostatistics, and orofacial clefts. This is a Project Grant, the standard NIH research project grant assistance type.

A functional smoothed machine learning approach for high-dimensional orofacial clefts genetic data analysis project summary orofacial clefts (OC) are one of the most challenging newborn defects that demand innovative quantitative analysis methods and tools based on genetic data. Machine learning methods, including deep neural networks (DNN), have been increasingly used in modeling the complex relationships between human diseases including OC and genetic variants due to their flexibility for complex pattern recognition. Meanwhile, the statistical techniques in functional data analysis (FDA) have also gained popularity in genetic studies benefiting from its capability of handling high dimensionality, missing data and dependency structure. The goals of this project are to develop a deep learning approach based on both DNN and FDA techniques to effectively predict the disease risk and identify predictive genes based on high dimensional genetic data, and to customize the proposed method for the analysis of OC. Several larger scale OC databases will be used to train and validate the proposed approach. The main hypothesis is that the proposed approach will take full advantage of both DNN and FDA and overcome the pitfalls of each method alone. Particularly, the approach will be capable to capture the dependent structure among the genetic variants for accurate prediction and flexible to handle both gene sequencing data and conventional covariates such as demographic data. The research will be led by an early-stage new investigator, who has assembled a multidisciplinary team experienced in statistics, computer science, biostatistics, and orofacial clefts. The specific aims of the project are to: 1) build a functional smoothed deep learning (FSDL) approach using the FDA techniques for disease risk prediction, particularly for OC risks, based on high dimensional genetic variants, investigate its computational efficiency and compare its performance with existing approaches through extensive simulation studies; and 2) develop a permutation- based association test procedure and feature importance scores for orofacial clefts by considering a large number of genetic predictors and covariates such as environmental factors. Based on the selected predictive genes, an aggregated FSDL prediction model will be trained using the pre-identified OC databases followed by meta-analysis. If successful, this new approach will provide a unified machine learning framework that can provide both risk prediction and an association test for orofacial clefts. Moreover, IT will promote the machine learning methods for the high-dimensional data in general (e.g., high-dimensional genetic and imaging data) and could be extended to other omics data, such as transcriptomic, epigenomic and proteomic data.