Fig. 1.

Illustration of the self-training algorithm. Step 1: train a base predictor, $f$, using ${\text{G}}_1$ and ${\text{P}}_1$ from animals with measured phenotypes. Step 2: predict self-trained phenotypes, ${\widehat{\text{P}}}_2$, based on ${\text{G}}_2$ for animals without measured phenotypes. Step 3: combine ${\text{G}}_1$, ${\text{G}}_2$, ${\text{P}}_1$, and ${\widehat{\text{P}}}_2$ to train a new predictor, $f^{\ast }$. In the testing phase, compare accuracies of $f$ and $f^{\ast }$ on the testing set (${\text{R}}_{\text{SL}}$ and ${\text{R}}_{\text{SSL}}$). First, predict phenotypes ${\widehat{\text{P}}}_{\text{T}}$ and ${\widehat{\text{P}}}_{\text{T}}^{\ast }$ based on ${\text{G}}_{\text{T}}$ using $f$ and $f^{\ast }$, respectively and second, calculate ${\text{R}}_{\text{SL}}$ (${\text{R}}_{\text{SSL}}$) as the correlation between ${\widehat{\text{P}}}_{\text{T}}$ (${\widehat{\text{P}}}_{\text{T}}^{\ast }$) and ${\text{P}}_{\text{T}}$