Computational Modeling and Machine Learning: The Cutting-edge Innovative Tools for Birth Defects Research, Prevention, and Treatment
By Sidra Shafique, PhD
“A baby learns to crawl, walk, and then run. We are in the crawling stage when it comes to applying machine learning.” -Dave Waters
Computational modeling and machine learning have applications in understanding the embryogenesis and the morphological basis of birth defects. Algorithms are created in machine learning to build a model based on a set of sample data. The model is then validated against experimental data for its accuracy and output. The disciplines most closely related to machine learning are bioinformatics, computational anatomy, computational modeling, and systems biology. All these disciplines act at the interface of mathematical, statistical, and data-analytical methods. Computational models can be used to run simulations to understand the outcomes of perturbations in genetic pathways to generate a hypothesis for testing in new experiments, to identify potential therapeutic targets, and so on.
I envision using available data in computational modeling for developmental sciences. As a researcher interested in the mechanisms of neural tube defects (NTDs), I understand that the NTDs result from failed neural tube closure during the first four weeks of embryonic development (24 days post-fertilization)(1). Underpinning these defects is a series of events involving signaling pathways, cytoskeletal components, and cell- and tissue-level mechanical interactions. Computational modeling has been effectively used to create a neural tube closure model. To understand how this model works, we need to know how it was created. The neural tube closure computational model of Brodland et al. (2010) effectively investigated the mechanics of neural plate morphogenesis, the formation of neural ridges, and prospective closure. This multiscale computational model of the whole embryo was created by dividing the surface epithelium into triangular regions consisting of several tens of cells. Each of these regions is represented along the apical and basal surfaces of the monolayer tissue, thus replicating the active forces produced by its cells. Thousands of cell-level simulations were incorporated to investigate the mechanics of embryonic epithelium closure. The cell-level constitutive equations relating stress, strain, cellular fabric, lamellipodium action, and other relevant factors were constructed. Then the model was created by first modeling a single cell, then a sheet of cells, followed by modeling a whole embryo using data from images taken from 63 different angles, consisting of 1282 surface points, and 2559 triangles of a live axolotl embryo. The simulation then was successfully run in the model to understand the impacts of absent neural ridge forces, unilateral absent neural ridge forces, increased thickness of neural ridge epithelium, and so on.(2)
Transforming the imaging data with calculation power and algorithms into sophisticated computation models and simulation is an emerging integrative discipline. Recently Dokmegang et al. (2021) created a computational model of the epiblast and trophectoderm using MG# (MechanoGenetic Sharp), an original computational model of biomechanics. The parent MG# project enables anyone with minimal or no programming experience to run biological simulations and visualise the results. MG#Core is the computational engine of MG#. It runs simulations and logs results into (custom) MG files. Using this MG# model, it was possible to reproduce key cell shape changes and tissue level behaviors in silico, while accounting for internal, cytoskeletal, and external forces. The simulations were successfully run to understand the epiblast remodeling and position relative to the trophectoderm.3
Computational models offer a holistic approach to combining, integrating, and visualizing experimental data. We see new scientific information coming along every day. However, this scattered and patchy information needs to be processed as ‘Big Data’. The discipline of bioinformatics applies network analysis to analyze and understand -omics data including genome, transcriptome, proteome, and metabolome. In most cases, the network analysis is formally focused on a defined set and source of data. Machine learning is the next step where algorithms would be used to predict the morphological outcome based on existing data. Machine learning, as a valuable visual tool, can help us understand the links between physical processes, gene expression, and cell fate. Although, it seems challenging to model the very dynamic multi-scale processes underlying embryogenesis by incorporating physics, genetics, epigenetics, and mechanics at the same time. The scientific community still needs to understand the importance of making available and maintaining open-source modeling codes. Despite this and other challenges, I believe that machine learning models created from embryonic morphogenesis and imaging data with multiscale biophysical algorithms can be used as a future clinical tool to predict birth defects. The obstetrical ultrasound and doppler scans provide us with high-resolution imaging data. For human pregnancies with a history of birth defects, patient data can be used to run simulations in already created models to evaluate the risk of prospective birth defects in the developing embryo. This information is of value from both the preventive and therapeutic perspectives. As mentioned before, we are just ‘crawling’ in this discipline. However, machine learning could be an effective and powerful diagnostic and therapeutic tool for human pregnancy evaluation in decades to come.
About the author
Sidra Shafique is a Postdoctoral Fellow studying developmental toxicology at Queen’s University, Canada. Her Master’s and Doctorate research focused on developmental sciences. She is a medical doctor (MBBS), Fellow (FCPS), and Member (MCPS) of the College of Physicians and Surgeons Pakistan in Obstetrics and Gynecology.
About the Society for Birth Defects Research and Prevention
Healthy pregnancies. Healthy babies. Better lives.
The mission of the Society for Birth Defects and Prevention (BDRP) is to understand the cause and pathogenesis of disorders of developmental and reproductive origin to prevent their occurrence and improve outcomes through research, collaboration, communication, and education.
Scientists interested in or already involved in research related to topics mentioned in this blog are encouraged to join BDRP and attend the 63rd Annual Meeting taking place June 24–28, 2023, in Charleston, South Carolina. BDRP is a multidisciplinary society of scientists from a variety of disciplines including researchers, clinicians, epidemiologists, and public health professionals from academia, government, and industry who study birth defects, reproduction, and disorders of developmental origin. Our members include those specializing in cell and molecular biology, developmental biology and toxicology, reproduction and endocrinology, epidemiology, nutritional biochemistry, and genetics, as well as the clinical disciplines of prenatal medicine, pediatrics, obstetrics, neonatology, medical genetics, and teratogen risk counselling. BDRP publishes the peer-reviewed scientific journal, Birth Defects Research. Learn more at http://www.birthdefectsresearch.org. Find BDRP on LinkedIn, Facebook, Twitter and YouTube.
References
1. Greene, N. D. E. & Copp, A. J. Neural tube defects. Annual Review of Neuroscience 37, 221–242 (2014).
2. Wayne Brodland, G., Chen, X., Lee, P. & Marsden, M. From genes to neural tube defects (NTDs): insights from multiscale computational modeling. HFSP Journal 4, 142 (2010).
3. Dokmegang, J., Yap, M. H., Han, L., Cavaliere, M. & Doursat, R. Computational modelling unveils how epiblast remodelling and positioning rely on trophectoderm morphogenesis during mouse implantation. PLoS ONE 16, 1–20 (2021).
