Patterns of non-uniform usage of synonymous codons vary across genes in an organism and between species across all domains of life. This codon usage bias (CUB) is due to a combination of non-adaptive (e.g. mutation biases) and adaptive (e.g. natural selection for translation efficiency/accuracy) evolutionary forces. Most models quantify the effects of mutation bias and selection on CUB assuming uniform mutational and other non-adaptive forces across the genome. However, non-adaptive nucleotide biases can vary within a genome due to processes such as biased gene conversion (BGC), potentially obfuscating signals of selection on codon usage. Moreover, genome-wide estimates of non-adaptive nucleotide biases are lacking for non-model organisms. We combine an unsupervised learning method with a population genetics model of synonymous coding sequence evolution to assess the impact of intragenomic variation in non-adaptive nucleotide bias on quantification of natural selection on synonymous codon usage across 49 Saccharomycotina yeasts. We find that in the absence of a priori information, unsupervised learning can be used to identify genes evolving under different non-adaptive nucleotide biases. We find that the impact of intragenomic variation in non-adaptive nucleotide bias varies widely, even among closely-related species. We show that the overall strength and direction of translational selection can be underestimated by failing to account for intragenomic variation in non-adaptive nucleotide biases. Interestingly, genes falling into clusters identified by machine learning are also physically clustered across chromosomes. Our results indicate the need for more nuanced models of sequence evolution that systematically incorporate the effects of variable non-adaptive nucleotide biases on codon frequencies.