#36 · ML / AI methods

Multi-objective genetic algorithms constrain sequence design to trusted regions

The common thread is multi-objective search over biological sequence space in which constraints are part of the objective function. Plant DNA Designer optimises codon usage and cassette design with a 19-objective genetic algorithm across 18 crop species. A distribution-constrained 5'UTR design method restricts a genetic algorithm to a KNN-defined trust region in predictor embedding space, keeping candidates in-distribution while preserving optimisation. The genetic-code study frames the standard code as a multi-objective optimum balancing mistranslation error load against amino acid composition, solved by simulated annealing. What is new is treating distribution membership and error load as explicit terms in the search rather than as post hoc filters on unconstrained scores. The first two items are applied design tools, while the third is a theoretical optimality analysis, so the shared thread is methodological rather than a common benchmark.