BGPT: Paper Review: CRISPR-Cas9 off-targeting assessment with nucleic acid duplex energy parameters

Explore by Goal

Quick Explanation Copied

Paper reviewed

Alkan et al. (2018) propose an energy-based Cas9-gRNA-DNA duplex model and two scores—CRISPRoff (off-target confidence per candidate) and CRISPRspec (gRNA specificity via a partition-function over target candidates). They report improved agreement with genome-wide experimental off-target datasets and suggest an inverse/on-target efficiency–specificity interplay mediated by binding energies. Evidence is strongest for the classification benchmarks and CIRCLE-seq correlation claims, but the model includes multiple approximations and heuristic parameters that could limit mechanistic universality.

Source:

Long Explanation

CRISPR-Cas9 off-targeting assessment with nucleic acid duplex energy parameters

Energy-parameterized off-target confidence (CRISPRoff) + ensemble specificity (CRISPRspec)

Genome Biology (2018) • DOI: 10.1186/s13059-018-1534-x

Reviewed paper:

What they built Duplex-energy Cas9 binding approximation → CRISPRoff + CRISPRspec

Model components: gRNA–DNA hybridization energy (position-weighted for Cas9 influence), DNA–DNA duplex opening penalty, gRNA folding penalty, and PAM-dependent correction factor.
CRISPRoff: for a specific predicted off-target, score = -G_B (binding free energy contribution).
CRISPRspec: partition-function specificity computed as a Boltzmann-weighted ensemble over all candidate on- and off-targets (excluding the on-target in the numerator) to yield an ensemble specificity score.

1) Reported benchmark performance (from the paper text)

The following plots use numeric values explicitly visible in the provided paper text (ROC/PR summary table). No additional data were inferred beyond what is stated.

2) Precision–Recall summary (text-reported)

3) Example operating point claim (text-reported)

The paper text claims: at fixed TPR = 0.9, CRISPRoff reaches FPR = 0.06, compared to CFD/Elevation higher FPRs (closest competitors mentioned: CFD and Elevation).

4) Mechanistic interpretation: what’s novel vs what’s approximated

Most plausible mechanistic core: the model assumes that Cas9 discrimination and off-target binding/cleavage can be approximated from relative duplex free energies plus a PAM correction and a gRNA folding penalty. This is implemented via CRISPRoff (per-site energy contribution) and CRISPRspec (ensemble weighting).

Key approximations / potential failure modes (skeptical audit):

Nearest-neighbor thermodynamics ≠ Cas9 physics. The approach relies on nearest-neighbor duplex parameters for RNA–DNA and DNA–DNA and then “projects” those into Cas9 binding stability via heuristic position weights and a PAM correction. This can be directionally useful but may not capture protein-mediated effects (e.g., specific contacts, conformational switching energy landscape) that go beyond duplex hybridization.
Missing energetic contributions. The paper states it ignores bulge energies because it scores mismatched off-target predictions and argues bulged off-targets are rarely reported at very low cleavage rates; this is a scope restriction.
Parameter transferability risk. δPAM values are set “arbitrarily for Cas9” based on “observations in the literature” about experimentally validated off-targets. If PAM effects differ across variants, conditions, or cell context, δPAM could become a systematic bias.
Ensemble specificity framing may hide heterogeneity. CRISPRspec is computed over all possible candidate targets in a predicted set. Agreement with experimental screens can be improved by better candidate enumeration—but any systematic mismatch between predicted candidate space and experimentally accessible targets (chromatin accessibility, replication timing, delivery mode effects) could reduce generalizability. The authors themselves focus on human genome datasets largely due to availability.

5) Evidence strength and what would most challenge the paper’s conclusions

Strongest reported evidence: improved ROC/PR benchmark summaries for CRISPRoff versus multiple competing tools (including CFD, MIT, CCTop, Cropit, Elevation, and VfoldCAS) using the Haeussler off-target benchmark protocol described in the paper.
Experimental alignment evidence: the authors report CRISPRoff correlates best with CIRCLE-seq cleavage activity across analyzed off-targets (while SITE-seq correlation is described as weaker/blurrier).
Challenging test: mechanistic falsification: the model assumes duplex energy dominance in binding stability; it would be most falsified if new experimental datasets show that off-target activity is largely decoupled from predicted duplex stability under conditions where duplex stability is systematically varied. The paper itself frames bulges, PAM effects, and other protein-driven kinetics as limitations/scope restrictions rather than fully solved physics.

6) Reproducibility & transparency signals (based only on what is in the provided text)

Code availability: the paper states source code is available via a GitHub repository (crisproff) and mentions a Zenodo DOI for the version used in submission, and GPLv3 licensing.
Webserver: computations can be performed through a webserver at rth.dk/resources/crispr/.
Data used: training/benchmarking relies on multiple published experimental datasets (Haeussler benchmark for ROC/PR; CIRCLE-seq and SITE-seq for activity/correlation; on-target datasets Doench2015 and Wang2015 for efficiency interplays). The paper also states which candidate enumeration tool it used for off-target site prediction in those datasets.

7) Counterpoints & blind spots to keep in mind

Assay/cell-context mismatch: The paper compares computational scores to multiple experimental screens that differ in lab protocols, concentrations, and sensitivity. Agreement in one context does not guarantee uniform performance elsewhere. The authors partially address this by noting weaker correlation in SITE-seq relative to CIRCLE-seq.
Candidate space limitation: CRISPRspec depends on predicted off-target sets (genome-wide enumeration up to 6 mismatches, filtered by PAM). If experimentally relevant off-targets lie outside the enumerated candidate space, partition-function specificity may misestimate risk.
Mechanism vs correlation: The inverse relationship between on-target efficiency and specificity is reported as evidence supporting an interplay mediated by binding energies, but the model is still an approximation—so it’s best treated as a hypothesis-generating pattern unless directly mechanistically validated in controlled perturbation experiments.

8) Notes on comparison to VfoldCAS (context for one baseline)

The ROC/PR table includes VfoldCAS as one of the compared baselines. VfoldCAS is described in the broader literature as a cleavage/active-structure folding ensemble approach; here, it is presented as a baseline method in the same evaluation.

Author reviews on BGPT

Feedback:

Updated: April 29, 2026