Résultats du processus de synthèse de l'article arXiv : 2101.1234v1

Résumé : A first analysis applying compressive sensing to a quantitative biological trait and its compressible "frequency domain" is presented. Consider an $n$-bit genetic sequence and suppose we want to discover a function that maps participating alleles (or even environmental influences) to a particular trait. Under plausible assumptions of how they evolved, certain traits can be viewed as "smooth" functions on the $n$-dimensional Boolean lattice of possible genomes. This allows approximation of their Fourier transforms, i.e., their gene networks, as sparse, dominated by "low-frequency" components. In turn, we can apply compressive sensing methods to collect relatively few samples, yet achieve accurate recovery. For an arbitrary quantitative trait affected by $n=26$ genes and with $812$ meaningful gene interactions, our simulations show noisy trait measurements ($\mathrm{SNR}=20\,\mathrm{dB}$) from just $M=44,\!336$ genomes in a population of size $N = 2^{26}$ (undersample ratio $M/N\approx0.00066$) permit discovering its gene network and predicting trait values, both with about $97.6\%$ accuracy. More dramatic undersample ratios are possible for traits affected by more genes. Work is currently underway to see if empirical data fit the proposed model. If so, it could offer a radical reduction in the number of measurements -- from exponential to polynomial in some cases -- necessary to quantify the relationship between genomes and certain traits.