yeast_sl_benchmark.md

Yeast Synthetic Lethality Benchmark Construction

---

Dataset construction

We construct a benchmark for S. cerevisiae synthetic lethality prediction using quantitative genetic interaction data from condition-specific SGA screens. We use two stress conditions from the Costanzo et al. 2021 dataset: bleomycin (BLEO), a DNA-damaging agent, and dimethyl sulfoxide (DMS), a solvent control that induces mild oxidative stress. Each condition profiles approximately 200,000 gene pairs, with four measurements per pair: the genetic interaction score $\pi$ (the difference between observed and expected double-mutant fitness, estimated as the ensemble mean of $10^3$ Monte Carlo simulations), the standard deviation of the $\pi$ distribution across simulations ($\sigma_\text{pair}$), a Z-score normalizing $\pi$ by the global standard deviation ($Z = \pi / \sigma_\text{experiment}$), and a Benjamini–Hochberg false discovery rate (FDR).

We label gene pairs as synthetic lethal (positive) or non-interacting (negative) using stringent criteria adapted from the validation thresholds in Shen et al. A gene pair is labeled synthetic lethal only if all three conditions are met: (1) FDR $<$ 0.05, (2) $Z < -3$, and (3) $\pi + 3\sigma_\text{pair} < 0$, ensuring that the interaction remains negative even at the upper bound of estimation uncertainty. A gene pair is labeled non-interacting if FDR $>$ 0.9 and $|Z| < 1.0$, requiring that the double-mutant effect is statistically indistinguishable from background. Gene pairs that do not meet either set of criteria (approximately 93% of the data) are excluded. Under these thresholds, the BLEO condition yields 139 SL pairs and 15,912 non-interacting pairs, and the DMS condition yields 204 SL pairs and 15,683 non-interacting pairs, for a combined pool of 343 SL pairs (approximately 1% of labeled pairs), consistent with the expected rate of synthetic lethal interactions in yeast.

For each condition, we sample 200 SL pairs and construct matched negative pairs at a 1:2 ratio. To ensure that negative pairs are not trivially distinguishable from positives, we adopt a shared-gene sampling strategy: for each sampled SL pair $(A, B)$, we draw one non-interacting pair of the form $(A, x)$ that shares the query gene and one non-interacting pair of the form $(y, B)$ that shares the array gene. This yields 600 gene pairs per condition (200 SL, 400 non-interacting) and 1,200 gene pairs in total. Prompts are paraphrased with o3-mini-0131 (temperature $= 1.0$) under a "biologist" role.