Figure 4: Performance of human scorers (H1,H2) and OWL software.

A) Relationship (left) between the total number of worms detected by H1 and H2 (solid blue line, slope = .846 $R^2=0.828$) and by the average human and OWL software (dashed black line, slope = 0.519; $R^2=0.812$). Shaded areas show the 95% confidence intervals of the fit. The fit residuals (right) indicate no systematic effect of the number of worms.

B) Relationship (left) between the mean worm position detected by H1 and H2 (solid blue line, slope = 0.993; $R^2=0.994$) and by the average human and OWL software (dashed black line, slope = 0.774; $R^2=0.957$). Shaded areas show the 95% confidence intervals of the fit. The fit residuals (right) indicate no systematic effect of the mean position.

C) Density as function of distance along the chemical gradient for three conditions (top to bottom): null condition, DMSO:DMSO, known attractant, isoamyl alcohol, known repellent, 1-octanol. Distributions scored by humans (light blue and aqua) and determined by OWL software (dark blue) are similar. Each image in the test dataset was scored by two human experimenters and by the OWL software, as described in Methods. The test dataset included 19 assay images (4 of diacetyl assays and 3 each for all other conditions).