bioRxivSource publication:
Loading content…
Perturb-seq often pools several guides per target as replicates, but a guide's off-target effects can bias the target-level estimate. PETRO assumes a strict majority of guides are valid for each outcome and uses robust weighting to downweight disagreeing guides, recovering the on-target perturbation effect without removing guides or using sequence annotations. In RPE-1 CRISPRa target-outcome pairs whose longest seed match was 13–18 bp, mean pooled bias was 0.087 and PETRO's residual bias was 0.005; in simulation, pooling stayed near 0.075 as cells increased while PETRO fell toward zero.
When several guides for one target are collapsed into a single target label, the target-level perturbation estimate still carries each guide's own off-target effects, even when those guides do perturb the intended gene. Conventional practice pools cells carrying guides for the same target or averages the estimated guide effects; this removes bias only if the weighted off-target effects cancel, so more cells sharpen a biased estimate rather than recovering the on-target component. In RPE-1 CRISPRa, using PAM-proximal promoter seed-match length as an off-target benchmark independent of expression response, mean guide-level bias rose from 0.024 at 9 bp to 0.194 at 11 bp and 0.460 at 13–18 bp; for target-outcome pairs with a 13–18 bp longest match, mean target-level bias under pooling was 0.087, about 19% of the guide-level bias and close to the longest-matching guide's mean pooling weight of 18%.
PETRO replaces the target-level estimate with a robust, outcome-specific weighted combination of guide effects, so a minority of guides that disagree with the majority is downweighted rather than removed from every outcome. PETRO applies a majority rule: for each target-outcome pair it assumes more than half the guides are valid, making the on-target perturbation effect the guide median without knowing which guides are valid; estimation uses robust M-estimation with Tukey bisquare weights computed separately per outcome. On unfiltered RPE-1 guide effects, PETRO kept mean target-level bias near zero, with residual bias of 0.005 at 13–18 bp; in K562 CRISPRi, for seven promiscuous guides independently flagged by Drepanos et al., PETRO's transmission of the on-target perturbation effect was 0.091 at day 7 and 0.111 at day 10 relative to pooling normalized to 1, without using the flags during fitting.
In simulation with one off-target guide among four, pooling's bias does not vanish as cells increase, whereas PETRO's bias approaches zero. With Γ=0 and δ=0.30, raising cells per guide from 150 to 750 left pooling's bias near 0.075 while PETRO's bias fell from 0.032 to about −1.78×10⁻⁴. The comparison comes from four-guide simulations with 10,000 replicates per condition, three valid guides and one with an unknown off-target effect; at Γ=0, δ=0.80 and 150 cells per guide, pooling had bias 0.20, the guide median 0.035 and PETRO 5.37×10⁻⁴.
The method targets standard multi-guide Perturb-seq analysis and takes guide-level effect estimates and their covariance from an upstream model, so linear, negative-binomial or pseudobulk DESeq2 models can supply the input; it was evaluated on RPE-1, Hs27 and A375 CRISPRa data, HEK293FT and K562 CRISPRi data, and four-guide simulations. The source states that tolerating up to k invalid guides requires at least 2k+1 guides per target, so odd guide counts use replication more efficiently under this criterion.
The majority rule is an identification assumption that generally cannot be verified from observed guide effects alone; if at least half the guides are invalid, especially when invalid guides agree with one another, PETRO may fail to recover the on-target perturbation effect. In finite samples small off-target effects are hard to distinguish from sampling noise, and the common-effect model or proportional extension may not fit every target-outcome pair. The cause of the ZNF219 guide disagreement remains unresolved, and the data do not identify which guide phenotype represents the biological effect of ZNF219.