Bio Genome Engineering Off Target Prediction

Version Compatibility

Reference examples tested with: Cas-OFFinder 3.0+, pandas 2.2+, Python 3.10+.

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures
• CLI: <tool> --version then <tool> --help to confirm flags

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Results depend on inputs far more than tool versions: the candidate list is bounded by the reference genome build, the mismatch/bulge tolerance, and the PAM pattern searched, not by the Cas-OFFinder version. The CFD matrix is SpCas9/NGG-specific and a relative ranker, not a calibrated cutting probability. Load the published CFD tables (Doench 2016 / CRISPOR distribution) rather than hand-typing values. Cas-OFFinder is the maintained snugel/cas-offinder repository (native DNA/RNA bulge support from v3.0.0).

Off-Target Prediction

"Check my guide for off-targets" -> Enumerate candidate sites genome-wide by mismatch/bulge tolerance, rank them by a per-site score, decide whether in-silico is sufficient or empirical discovery is required, and report each claim at the right rung: predicted, detected, or validated.

• CLI: cas-offinder input.txt G output.txt enumerates sites (no ranking)
• Python: CFD scoring from the published mismatch/PAM tables; aggregate specificity
• Web/CLI: CRISPRme for variant-aware (gnomAD + individual) nomination; CRISPOR to aggregate

The Single Most Important Modern Insight -- in-silico enumeration nominates candidates; it does not measure which sites are cut

The naive model -- "search the genome within N mismatches, score by CFD, the high scorers are my off-targets" -- is wrong in three structural ways no better scoring fixes:

1. Mismatch count is not cleavage. A 2-mismatch site in closed chromatin may never be cut; a 3-mismatch site in open chromatin near an active promoter is. Cellular cutting depends on chromatin, dose, and exposure time -- invisible to a sequence search.
2. Bulges and non-canonical PAMs are routinely missed. Real validated off-targets occur with 1-2 nt DNA/RNA bulges and at NAG/NGA PAMs; fixed-alignment mismatch-only search misses them. The failure is silent -- a clean report looks identical whether the guide is specific or the search just couldn't see the off-target.
3. CFD is a narrow, SpCas9-only relative ranker. A CFD of 0.08 is not "8% chance of cutting"; comparing two guides' aggregate scores is fine, reading an absolute CFD as a safety threshold is not.

The corollary, and the central professor-level point: in-silico lists overlap only partially with empirically validated off-targets, and the empirical genome-wide assays disagree with each other too. No single method is authoritative. Off-target evidence escalates: predicted -> detected by an unbiased assay -> validated by targeted amplicon deep-seq. Conflating these rungs is the field's most common error. Therapeutic-grade assessment is triangulation (variant-aware in-silico + >=2 orthogonal empirical assays + amplicon validation + a structural readout), never one tool's output.

In-Silico Taxonomy -- enumerate, then score, then aggregate

The unifying caveat: every score is bounded by the enumerator's coverage -- if the enumerator didn't propose a site (bulge, distal PAM, beyond the mismatch cutoff), no scorer will ever flag it.

Empirical Discovery Assays -- each has a characteristic bias; concordance is partial

The load-bearing reality: in-vitro assays over-call (high sensitivity, low cellular specificity); cell-based assays under-call rare sites and are cell-type-dependent (K562 yields far more hits than HEK293 for the same guide). Cross-method discordance is information, not noise -- sites found by both are high-confidence; in-vitro-only sites are likely chromatin-protected. The defensible workflow is the VIVO logic (Akcakaya 2018): sensitive in-vitro generator -> cell-based assay in the relevant cell type -> amplicon validation.

Decision Tree by Scenario

High-Fidelity Nucleases -- often a bigger lever than guide reselection

Two tacit points: (1) delivery format matters -- eSpCas9/HF1 can lose their fidelity advantage delivered as a high transient RNP bolus; HiFi Cas9 and Sniper-Cas9 stay specific and active as RNP. (2) Fidelity has a guide-dependent on-target tax -- a variant clean and active on guide A can be nearly dead on guide B. Pick the variant, then test it on the target guide in the intended delivery format; transferability is not assumable.

Variant-Aware Screening (reference-only is a clinical liability)

A patient is not GRCh38. A common SNP can restore a PAM or remove the protective mismatch at a near-target site, creating an off-target that exists only in some individuals -- and because variant frequencies differ by ancestry, reference-only screening systematically misses off-targets common in under-represented populations. For a human therapeutic guide, an off-target check must expand from "checked off-targets" to "checked off-targets variant-aware, across ancestries" -- run CRISPRme against gnomAD (and the treated individual's genome).

Enumerate Candidate Sites with Cas-OFFinder

Goal: Generate the genome-wide candidate-site list for one or more guides, including bulges and relaxed PAMs.

Approach: Write the Cas-OFFinder input file -- genome path, an optional DNA/RNA bulge line (v3.0.0+), a pattern with N's at guide positions and the PAM (use NRG to also catch NAG/NGG), then one query line per guide (guide bases + N's for the PAM positions, same length as the pattern) with its mismatch tolerance. Run on GPU if available. The output is a flat site list with mismatch counts -- it is a hypothesis set to score downstream, not a verdict.

# input.txt
# /path/to/genome_dir            # directory of FASTA (Cas-OFFinder indexes it)
# 2 2                            # DNA bulge, RNA bulge (omit this line for no-bulge search)
# NNNNNNNNNNNNNNNNNNNNNRG        # 20 N (guide) + NRG PAM -> also catches NAG
# GGCCGACCTGTCGCTGACGCNNN 4      # query: 20 guide bases + NNN (PAM positions), <=4 mismatches
cas-offinder input.txt G output.txt   # G=GPU, C=CPU, A=auto

Score Candidates with the Published CFD Tables

Goal: Rank candidate sites by relative cleavage propensity and compute an aggregate guide-specificity score for comparing guides.

Approach: Do NOT hand-type the CFD matrix. Load the published Doench 2016 tables (mismatch_score.pkl, pam_scores.pkl -- they ship with CRISPOR and the Doench code), take the product of per-position mismatch penalties x the PAM penalty for each site, and aggregate as 100/(1 + sum(CFD)) with per-site CFDs on a 0-1 scale (the CRISPOR specificity formulation; equivalently 10000/(100 + 100*sum)). Compare aggregate scores among candidate guides; never read an absolute CFD as a safety guarantee. (See examples/off_target_analysis.py.)

import pickle

def load_cfd_tables(mismatch_pkl, pam_pkl):
    '''Load the published Doench 2016 CFD tables (distributed with CRISPOR) -- do not fabricate.'''
    with open(mismatch_pkl, 'rb') as f:
        mismatch = pickle.load(f)   # keys like 'rA:dG,3' -> penalty
    with open(pam_pkl, 'rb') as f:
        pam = pickle.load(f)        # keys like 'AG' -> penalty
    return mismatch, pam

Structural Consequences Amplicon Panels Miss

Validating only with a short amplicon at each predicted site systematically misses the large-scale outcomes that are often the real safety concern:

• Large deletions / complex rearrangements at the on-target (Kosicki 2018 Nat Biotechnol 36:765) -- kilobase deletions whose alleles often drop out of the PCR, so the amplicon reads back more wild-type than it is.
• Chromosomal translocations between on- and off-target (or multiplexed) cuts -- need junction-capture (PEM-seq, UDiTaS, HTGTS, CAST-seq), not amplicon panels. A "clean" amplicon panel does not certify the absence of these.

The Evidence Ladder & Limit of Detection

"Not detected" means "below the LoD," never "zero." State the LoD: 0.05% editing is irrelevant for a research knockout but is ~50,000 mis-edited cells in a 10^8-cell therapy.

Per-Method Failure Modes