The role of balancing selection (BS) is a long-standing evolutionary puzzle. The emergence of advanced genomic techniques has furnished researchers with enormous datasets, revitalising the exploration of this topic. Unlike directional selection that propels advantageous alleles to dominance, BS encompasses a spectrum of phenomena that sustains genetic variation in populations across extended periods. BS is of particular interest within the realm of immune function genes and host-parasite interactions in multiple organisms. In insect research, there is a noteworthy surge in interest regarding sexual antagonism levels and the genes associated with sex determination. Detecting BS is known to be a challenging problem that is often entangled with the effects of linkage disequilibrium (LD). LD refers to the inheritance of genome sections as units, leading to interconnections between genetic sites. It is possible to simulate both phenomena using ancestral recombination graphs (ARGs) generated in the forward simulation framework SLiM, however, inference frameworks often struggle to disentangle both effects. The integration of simulation data with Deep Learning (DL) algorithms in evolutionary biology has recently provided a means to account for these intertwined effects, while also enabling efficient parameter estimation. Notably, algorithms powered by Artificial Intelligence (AI) possess the computational power required to navigate the intricate genomic dependencies arising from the linkage. Our objective is to tackle this challenge by incorporating the interplay between genomic positions and BS using a mechanistic framework. We will investigate the accuracy of BS inference in the presence of LD. Our strategy consists of simulating genomic segments combining both effects and utilising a method for detecting BS, augmented by Machine Learning (ML) algorithms capable of simultaneously handling multiple genomic regions, to perform the inference. This investigation will shed light on the robustness and reliability of BS detection when LD is considered. To achieve this, we will construct a deep learning approach capable of (i) classifying the presence of BS, (ii) determining the best-fitting models among those with and without selection, and (iii) estimating selection parameters while accounting for linkage. We will use an approach developed in our group, Polymorphism-aware phylogenetic models (PoMos), which integrate the timescales of population genetic processes and phylogenetic divergence data. In particular, we will build on the recent development by research fellow and co-investigator Svitlana Braichenko to detect signatures of balancing selection (PoMoBalance). We will combine it with the model selection framework for AI proposed by Olivier Gascuel group based at the Paris Artificial Intelligence Research Institute (PRAIRIE), France. Furthermore, we will collaborate with the Adam Siepel group from Cold Spring Harbor Laboratories (CSHL), US with experience of generating ARGs data with forward simulations (SLiM) to train our new deep-learning approach. Our framework will be tested on simulated data and real sequences extracted from the great ape and fruit fly whole genome datasets.