Improving screening efficiency by combining propensity-based sampling with interaction score prediction via matrix completion.
A: We extend the simpler protocol (propensity-based sampling only, Figure 1C), adding a projection-based predictor to choose likely synergistic pairs (steps 3 and 4). If the prediction-driven screening discovery rate is higher than the preceding propensity-based screening, a new prediction-driven screening cycle is started (step 5). We switch between propensity-based sampling and prediction to increase the fractional discovery rate. B: Fractional discovery rate across 9 data sets show marked improvement over brute-force screening. C: Estimates of the screening efficiency demonstrate that the full protocol (steps 1–5) gives better performance than propensity-based sampling only (steps 1–2). Yellow block: additional contribution by projecting onto in the largest yeast SGA screen.