Predictive rubber tree farming via machine learning: basis for increased production of the main renewable source of latex

• Grant number: 20/08293-0
• Support Opportunities: Regular Research Grants
• Start date: February 01, 2022
• End date: July 31, 2024
• Field of knowledge: Agronomical Sciences - Agronomy - Crop Science
• Principal Investigator: Anete Pereira de Souza
• Grantee: Anete Pereira de Souza
• Host Institution: Centro de Biologia Molecular e Engenharia Genética (CBMEG). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Associated researchers: Antonio Lucio Mello Martins ; Elaine Cristine Piffer Gonçalves ; Livia Moura de Souza

Abstract

The high price of oil and the growing interest in consuming raw material from renewable sources has increased the demand for natural rubber, which that the only commercial source is the rubber tree (Hevea brasiliensis). Brazil came from a situation which was the original center of production of rubber and had the greatest diversity, became the largest producer of it and now is importing most of what consumes. The rubber plantations implanted in the State of São Paulo resumed the growth of the country's crop, despite still contributing only 1.44% of the natural rubber consumed worldwide. To make possible an increase of the heveculture in new potential areas investment in the predictive agriculture is necessary: a combination of technological advances in genomics, phenomics, artificial intelligence and quantitative models, which enable the selection of superior clones tolerant to planting in colder and drier regions in the Southeast and South. The multiplication of the clones is done using grafting, and the uniformity in the performance of the clones is of great importance for the production of latex and its industrial purposes. However, although the grafting method characterizes a genetic-cultural strategy to ensure that the grafting of a monoclonal block is isogenic, this uniformity expected by the vegetative propagation in the rubber groves has not been observed. Thus, high coefficients of variation for vigor and rubber production have been observed among different clones and even within the same clonal varieties. However, there are no studies that use estimates of genetic and phenotypic parameters to understand this variation found between the vigor and productivity variables in rootstocks. In this context, the transdisciplinary domain of combining the different methodological strategies of genomics, machine learning and quantitative genetics expands the potential to elucidate the molecular mechanisms of latex production. In the absence of studies on genomics and machine learning applied to rootstock improvement, the present work proposes: (i) Genomic prediction (PG) using characteristics of economic importance (such as latex production) for selection of potential grafting combinations; (ii) Evaluation of models based on machine learning for this predictive task, contrasting them with parametric and Bayesian statistical techniques; (iii) Genome-Wide Association (GWAS) of the graft / rootstock relationship to identify chromosomal regions, genes and metabolic pathways related to different phenotypic characteristics; (iv) Use of attribute selection techniques and interpretable machine learning models to complement traditional GWAS techniques. As a result of the application of PG and GWAS to rubber rootstocks, it is expected to accelerate the selection gains, reducing the time and cost of a breeding cycle through early selection. Furthermore, it is intended to assess the potential for increased associated selection gains using, in conjunction with genomic approaches, artificial intelligence models. In addition, the associations of markers found in regions of quantitative locus in rootstocks can be applied in breeding programs via marker-assisted selection, in addition to directing the identification of genes that are related to latex production. (AU)