Target Detection as the Rate-Limiting and Specificity-Limiting Step in CRISPR Genome Editing

The dominant metaphor for CRISPR is surgical. "Molecular scissors." The metaphor is not merely imprecise; it is inverted. It places the difficulty at the blade, where there is none, and hides the difficulty in the approach, where all of it lives.

Consider the actual sequence of events. A Cas9 ribonucleoprotein (RNP) enters a nucleus. It has one job: find a specific 20-bp sequence. It does not know where the sequence is. It has no map, no index, no directory, no ATP-driven motor to walk the genome systematically. It has thermal motion and a 3-bp recognition motif. The genome it must search is approximately 2 metres of DNA compacted into a sphere roughly 10 μm across, of which ~75–80% is wrapped around nucleosomes and physically inaccessible [E].

The RNP will collide with DNA on the order of 10⁵–10⁷ times before it finds its target. Each collision is a decision: probe or leave. Most of those decisions must be made in under 10 milliseconds, or the search never completes. And each decision is made on radically incomplete information — at the moment of collision, the RNP has read at most 3 bases.

Then, having found something that looks right, it must decide whether it is right, using a discrimination mechanism that runs on no fuel and must reach a verdict before performing an irreversible act.

This is the target detection problem. The cut is an afterthought.

1.1 The permission-structure reading

There is a natural way to state the apparent impossibility:

You cannot reliably address one site in 6.4 × 10⁹ using a 20-nucleotide guide and a 3-nucleotide motif, in a crowded nucleus, in minutes, with thermal energy alone.

Stated that way, it sounds like a prohibition. It is not. It is a specification with an unpriced term. The correct reading identifies the currency:

Apparent impossibility Actual specification Currency paid "Too many decoys to search" Search completes if PAM binding is weak enough that decoys do not sequester the enzyme Enzyme copy number × exposure time "Guide has enough bits but still misfires" The machine does not read all 40 bits at equal weight; only the seed is fully priced Effective address length; distal bases are nearly free "Binding energy cannot discriminate a 1-in-20 mismatch" Correct — equilibrium binding cannot. Kinetic gating can, at a cost On-target rate; discarded near-misses "Cannot exceed single-checkpoint discrimination" Correct for one checkpoint. Add an independent second gate External free energy, or a second molecular event (dimerization, nicking) "Cannot prove absence of off-targets" Correct at 10⁻³ floor. Not correct in principle Sequencing chemistry error rate; molecular consensus depth

Every row of that table is a place where the field has already paid, or where it has not yet noticed there is a bill. The rest of this paper prices each one.

1.2 What this paper does not claim

We do not claim that CRISPR is unsafe, or that clinical editing is unjustified. Approved therapies exist and work [E]. We claim something narrower and more structural: that the verification of specificity is bounded by physics and chemistry in ways that are not fully acknowledged in the standard reporting conventions, and that the boundedness is quantifiable rather than vague.

2.1 Information budget

Let $L=3.2\times {10}^9$ be the haploid human genome length. A target may lie on either strand, so the address space contains $2L=6.4\times {10}^9$ candidate positions.

Bits required to specify one position uniquely:

• —𝐼
• —req
• —=
• —log
• —2
• —(
• —2
• —𝐿
• —)
• —=
• —log
• —2
• —(
• —6.4
• —×
• —10
• —9
• —)
• —=
• —32.58
• —bits
• —I
• —req

=log 2

(2L)=log 2

(6.4×10 9 )=32.58 bits

Bits nominally carried by an SpCas9 guide:

• —𝐼
• —nom
• —=
• —2
• —𝑛
• —spacer
• —+
• —𝐼
• —PAM
• —=
• —2
• —(
• —20
• —)
• —+
• —2
• —=
• —42
• —bits
• —I
• —nom

=2n spacer

+I PAM

=2(20)+2=42 bits

where the NGG PAM contributes ${\log }^{2}16=4$ bits over a random trinucleotide... more carefully: NGG constrains 2 of 3 positions, so $I_{PAM}=2\times 2=4$ bits. (We use 4 bits below; the 42-bit total is unchanged only if we count the spacer as 40. Using $I_{PAM}=4$: $I_{\mathrm{n}\mathrm{o}\mathrm{m}}=44$ bits.)

Nominal margin: $44-32.6=11.4$ bits ≈ a factor of 2,700. On paper, an SpCas9 guide is massively over-specified. Uniqueness should be trivial.

[T-1] The nominal information content of a 20-nt spacer plus NGG PAM exceeds the human genome addressing requirement by ~11 bits. Sequence information is not the limiting resource.

This is the first and most important reframing. The off-target problem is not an information-shortage problem. The information is there. It is not read.

2.2 The effective address: positional weighting

Cas9 does not weight the 20 spacer positions equally. Mismatches in the PAM-proximal "seed" (roughly positions 1–10 counting from the PAM) are strongly rejected; PAM-distal mismatches are frequently tolerated [E]. Single-molecule FRET measurements quantify this directly: the dissociation rate of Cas9–RNA from a fully matched target is below 0.006 s⁻¹, but exceeds 2 s⁻¹ when mismatches are introduced near the PAM — a >300-fold shift in rejection kinetics driven by mismatch position, not mismatch count [E].

Define a positional weight $w_i\in [0,1]$ for each spacer position, representing the fraction of the 2 bits at that position that actually contributes to discrimination:

• —𝐼
• —eff
• —=
• —2
• —∑
• —𝑖
• —=
• —1
• —20
• —𝑤
• —𝑖
• —+
• —𝐼
• —PAM
• —I
• —eff

=2 i=1 ∑ 20

w i

+I PAM

Empirical mismatch-tolerance profiles imply $w_i\approx 1$ for $i\lesssim 10$ and decaying to $w_i\approx 0.1$– 0.4 0.4 for $i\gtrsim 14$ [E, approximate]. Integrating:

• —𝐼
• —eff
• —≈
• —2
• —(
• —10
• —+
• —4
• —×
• —0.3
• —)
• —+
• —4
• —≈
• —26
• —bits
• —I
• —eff

≈2(10+4×0.3)+4≈26 bits

with a plausible range of 22–30 bits depending on guide sequence, chromatin context, and nuclease variant.

[T-2] The effective address length of SpCas9 is 22–30 bits against a 32.6-bit requirement. The system is under-specified by 3–10 bits, corresponding to an expected decoy multiplicity of $2^3$ to $2^{10}$, i.e. 8 to 1,000 genome-wide sites indistinguishable to the enzyme's discrimination machinery.

This single inequality — $I_{\mathrm{e}\mathrm{f}\mathrm{f}}<I_{\mathrm{r}\mathrm{e}\mathrm{q}}$

— is the origin of off-target editing. Everything else is detail.

2.3 Decoy census

We can compute the decoy population directly. The number of distinct 20-mers at exactly $k$ mismatches from a reference is ( 20 𝑘 ) 3 𝑘 ( k 20

)3 k , out of $4^{20}=1.10\times {10}^{12}$. Multiplying by the number of NGG-flanked positions in the genome ( $2L\mathrm{/}16=4.0\times {10}^8$):

Mismatches $k$ Distinct variants Expected genome-wide sites Cumulative ≤ k 0 1 0.00036 0.0004 1 60 0.022 0.022 2 1,710 0.62 0.64 3 30,780 11.2 11.8 4 392,445 143 155 5 3,767,472 1,370 1,525 6 28,256,040 10,300 11,800

Reading the table. At $k=0$ the expected count is 0.0004 — a perfect-match site is essentially unique, confirming T-1. But Cas9 routinely cleaves at 3–5 mismatches when those mismatches are PAM-distal [E]. At that tolerance, the expected decoy population is 10²–10³ sites per guide.

Observed off-target counts from unbiased cellular assays are typically 0–150 per guide [E] — one to two orders of magnitude below the naive census. The gap is real and has three components, which we quantify in later sections: chromatin inaccessibility (§7), sub-threshold kinetics that bind but never cleave (§5), and assay detection floors (§6). The third component is the dangerous one, because it is a property of the measurement, not of the biology.

[C-1] The discrepancy between the ~155-site census at $k\le 4$ and the typical 0–150 observed sites is not primarily explained by biology. We conjecture that the dominant term is the detection floor, and that duplex-consensus resequencing of the full $k\le 4$ census would recover a substantial fraction of currently "undetected" sites at frequencies of 10⁻⁴–10⁻⁶.

2.4 Bulges: the term everyone drops

The census above assumes ungapped alignment. RNA and DNA bulges — insertions or deletions in the heteroduplex — are tolerated by Cas9 and produce validated off-target cleavage [E]. Allowing a single bulge expands the candidate space by roughly a factor of $2\times 20\times 3\approx 120$ in the worst case, though thermodynamic penalties suppress most of them.

Most in-silico off-target predictors handle bulges poorly or not at all, and most experimental validation panels are constructed from those predictors. This creates a closed loop: predictors define the panel, the panel confirms the predictor, and the bulge class never enters either.

[O-1] The genome-wide contribution of bulged off-target sites to total unintended editing is unquantified. No unbiased assay has been analyzed with a gapped-alignment pipeline at therapeutic-scale sensitivity.

2.5 The PAM-relaxation trade

Engineered PAM-relaxed variants (VQR, EQR, VRER, SpG, SpRY) expand the addressable fraction of the genome [E]. The cost is exactly computable. Relaxing NGG (frequency 1/16) to NRN (~1/8 in practice for the most permissive variants) multiplies the PAM-site population by ~8:

Every row of the decoy table scales by the same factor. A near-PAMless nuclease at $k\le 4$ faces ~1,200 expected decoys instead of ~155.

[T-3] PAM relaxation purchases addressability at a directly proportional cost in decoy census and in search-time tax (§4). The exchange rate is linear in PAM degeneracy. There is no free relaxation.

Conversely, Cas12a's TTTV PAM (frequency ~3/256 ≈ 1/85 per strand) yields a decoy census ~5× smaller than SpCas9's before any consideration of its longer 23-nt spacer or its distinct kinetics [T]. Part of Cas12a's reported specificity advantage is not mechanism at all — it is arithmetic.

Before analyzing search kinetics, it is worth being concrete about the epistemic state of the enzyme at each moment.

Stage Bases read Bits acquired Cumulative Decision Free diffusion 0 0 0 — Nonspecific dsDNA contact 0 0 0 slide or release PAM engagement 3 4 4 unwind or release Seed R-loop (1–10) 10 20 24 propagate or reject Full R-loop (11–20) 20 40 44 dock HNH or reject Cleavage — — — irreversible

The critical observation: at the moment the enzyme commits to unwinding the duplex — an energetically expensive act — it has acquired 4 bits against a 32.6-bit requirement. It is operating at 12% of the information it needs. The unwinding decision is therefore made almost blind, and must be cheap enough to be made 10⁸ times.

This is why PAM binding must be weak. Not because weak binding is elegant, but because a decision made on 4 bits must be reversible on a millisecond timescale or the search deadlocks. §4 makes this quantitative.

4.1 The diffusion-limited ceiling

For a small target of radius $b\approx 1$ nm and an RNP diffusion constant $D\approx 3\mu m^2{\mathrm{s}}^{-1}$, the Smoluchowski association rate is

M − 1 s − 1 k 3D

=4πDbN A

=2.3×10 7 M −1 s −1

A single target locus in a nuclear volume of ~500 μm³ has an effective concentration of

= VN A

1

=3.3 pM

giving an idealized search time for a single RNP molecule of

= k 3D

c target

1

=1.3×10 4 s≈3.7 hours

[T-4] Under pure 3D diffusion with no decoys and no chromatin, one Cas9 RNP requires ~4 hours to locate one genomic target. This is the floor, not the estimate.

4.2 The decoy tax

The nucleus is not empty. Total DNA concentration is ~21 mM base pairs. The NGG PAM population is

= VN A

2L/16

=1.33 mM

If PAM engagement has dissociation constant $K_{D_{PAM}}$

, the fraction of RNP molecules free at any instant is

=(1+ K D PAM

c PAM

+ K D ns

c ns

) −1

and the search time is inflated by $1\mathrm{/}{f}_{\mathrm{f}\mathrm{r}\mathrm{e}\mathrm{e}}$

:

Search-time multiplier 10 mM 0.13 1.1× 1 mM 1.3 2.3× 100 μM 13 14× 10 μM 133 134× 1 μM 1,330 >1,300×

[T-5] (PAM affinity ceiling). The genome-wide PAM concentration of ~1.3 mM imposes an upper bound on useful PAM affinity. A nuclease with micromolar PAM affinity would spend >99.9% of its residence time on decoys and would not complete a genome search within a cell cycle. Weak PAM binding is not a compromise; it is a hard requirement derived from decoy concentration.

This is the cleanest permission-structure result in the paper. An engineer looking to improve Cas9 might reasonably propose tightening PAM recognition to reduce spurious engagement. The calculation says the opposite: tighten PAM affinity and the enzyme never arrives. The constraint does not prohibit high-affinity recognition — it specifies exactly where the affinity must be spent, which is on the heteroduplex, after commitment, not on the motif, before it.

4.3 Facilitated diffusion: how much does sliding buy?

Whether Cas9 uses one-dimensional sliding was contested. Early DNA-curtain work found predominantly 3D collision with no long-range sliding [E]. Subsequent single-molecule FRET and ensemble kinetics established short-range 1D diffusion with a sliding length of ~20–27 bp and a mean nonspecific dwell of ~11 ms on PAM-free duplex, accelerating target location by roughly 10-fold [E]. The sliding is biased 3′→5′ along the non-target strand under physiological salt, mediated by nonspecific contacts ~8 bp downstream of the PAM [E].

Applying the Halford–Marko combined-mode expression, the search time scales as

⇒

𝜏 ∝ 1 ℓ 𝑠 𝑙 τ search

V

⋅ something L DNA

1

with the antenna enhancement of order ℓ 𝑠 𝑙 / 𝑏 ≈ 27 bp / 3 bp ≈ 9 ℓ sl

/b≈27 bp/3 bp≈9, consistent with the measured ~10× [E, T].

[T-6] Facilitated diffusion contributes an order-of-magnitude search acceleration for Cas9 — not the 10²–10³ enhancement seen for classical sliders like lac repressor. Cas9 is a poor facilitated diffuser. It compensates by copy number.

4.4 The full budget

Assembling: with a 10× sliding benefit, a ~2–14× decoy tax, and a ~4× accessibility penalty from chromatin (§7):

𝜏 search ≈ 𝜏 ideal 𝑁 RNP × decoy tax × chromatin penalty sliding gain τ search

≈ N RNP

τ ideal

× sliding gain decoy tax×chromatin penalty

For $N_{RNP}={10}^4$ (a typical RNP electroporation) and mid-range parameters:

≈ 10 4 1.3×10 4 s

× 10 5×4

≈2.6 s

For $N_{RNP}={10}^2$ (a poorly delivered viral vector, or an RNP that has been mostly degraded):

≈260 s

[T-7] (The dose–specificity coupling). On-target search time scales as $1\mathrm{/}{N}_{RNP}$

. But off-target cleavage probability scales as $N_{RNP}\times t_{\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}}$

, because off-target sites are reached by the same search process and are cleaved by the same enzyme with a lower per-encounter success probability. Therefore:

on-target completeness ↑

⇔

𝑁 RNP 𝑡 exp ↑

⇔

off-target burden ↑ on-target completeness↑⇔N RNP

t exp

↑⇔off-target burden↑

Dose and specificity are not independent knobs. They are the same knob with a sign flip.

This is the mechanistic basis of the empirically well-known observation that RNP delivery (short exposure, rapid degradation) outperforms plasmid delivery (sustained expression) on specificity at matched on-target efficiency [E]. The paper's contribution here is to show that this is a theorem about the search process, not a delivery-modality quirk.

5.1 Why equilibrium binding cannot work

Naively, discrimination should follow from binding energy: the correct target binds more tightly, so it is preferentially cleaved. The ratio would be

For one RNA:DNA mismatch, $\mathrm{\Delta }\mathrm{\Delta }G\approx 1$–3 kcal/mol ≈ 1.7–5 $k_{B}T$, giving ${\eta }_{\mathrm{e}\mathrm{q}}\approx 5$–150.

But this argument fails for a reason that is easy to miss. The full 20-bp R-loop has $\mathrm{\Delta }G\approx -20$ to $-30$ kcal/mol. Both the target and a 4-mismatch decoy are bound far past saturation. In the saturated regime, the occupancy ratio collapses toward 1 even when the affinity ratio is large. Equilibrium binding discriminates affinity, but cleavage responds to occupancy, and occupancy is a saturating function of affinity.

[T-8] Equilibrium binding discrimination is structurally incapable of delivering the required specificity, because the on-target complex is saturated. Discrimination must be kinetic: it must occur during R-loop propagation, before saturation is reached.

This is a genuine no-go, and it reads as a specification: the decision must be made in the transient, not at the endpoint.

5.2 R-loop propagation as a biased random walk

Model R-loop formation as a one-dimensional walk on states $n=0,1,\mathrm{.}\mathrm{.}\mathrm{.},N$ where $n$ is the number of base pairs of heteroduplex formed, $n=0$ is PAM-bound-but-unwound, and $n=N=20$ triggers the irreversible conformational commitment (HNH docking and cleavage). Let $G(n)$ be the free energy of state $n$ relative to $n=0$.

The probability that a walk starting at $n=0$ reaches $N$ before dissociating is the classical gambler's-ruin splitting probability:

=[1+ j=1 ∑ N

e G(j)/k B

T ] −1 ≡ 1+S 1

where $S=\sum jeG(j)\mathrm{/}{k}_{B}T$.

Now introduce a mismatch at position $m$. It raises the free energy of every state $j\ge m$ by $ϵ$. Writing $S<=\sum j<meG(j)\mathrm{/}{k}_{B}T$ and $S\ge =\sum j\ge meG(j)\mathrm{/}{k}_{B}T$:

P on

= 1+S <

+S ≥

1+S <

+e ε/k B

T S ≥

=1+(e ε/k B

T −1) 1+S <

+S ≥

S ≥

This equation derives the seed rule. For a downhill (favorable) propagation landscape, $G(j)$ decreases with $j$, so $eG(j)\mathrm{/}{k}_{B}T$ is dominated by small $j$. Therefore:

• —PAM-proximal mismatch (small $m$):
• —$S\ge$

contains the dominant terms → $S\ge \mathrm{/}(1+S)\to 1$ → $\eta \to e_{ϵ\mathrm{/}{k}_B}T$ (maximal discrimination) PAM-distal mismatch (large $m$): $S\ge$

contains only subdominant terms → $S\ge \mathrm{/}(1+S)\to 0$ → $\eta \to 1$ (no discrimination)

[T-9] The seed effect is not a property of protein contacts or structural recognition. It is a topological consequence of a downhill free-energy landscape in a splitting-probability problem. Any nuclease with a monotonically favorable R-loop propagation landscape will exhibit a seed, regardless of its structure.

This result is consistent with, and here derived independently of, published hybridization-kinetics models of CRISPR off-targeting [E].

5.3 The discrimination ceiling theorem

Read the expression for $\eta (m)$ again. The prefactor $(e_{ϵ\mathrm{/}{k}_B}T-1)$ is a hard bound:

T for all m

Equality requires $S\ge \mathrm{/}(1+S)\to 1$, i.e. $S<\to 0$ and $S\gg 1$. But $P_{\mathrm{o}\mathrm{n}}=1\mathrm{/}(1+S)$, so:

𝑆 ≫ 1

⇒

→ S 1

→0

[T-10] (Discrimination ceiling). For a single-checkpoint proofreader with no external free-energy input, the per-mismatch discrimination factor is bounded by ${\eta }_{\max }=e_{ϵ\mathrm{/}{k}_B}T$, and

⟺

⟺P on

→0

There is no landscape shape that achieves maximal discrimination at finite on-target efficiency. Specificity and activity are conjugate variables in a single-checkpoint nuclease.

With $ϵ\approx 2$–4 $k_{B}T$, ${\eta }_{\max }\approx 7$–55 per effectively read mismatch. A 4-mismatch decoy, if all four fall in the dominant region, is suppressed by $\sim {10}^3$ – ${10}^6$. If they fall PAM-distal, suppression approaches 1.

This theorem explains, in one line, the entire empirical record of high-fidelity variants. eSpCas9, SpCas9-HF1, HypaCas9, evoCas9, Sniper-Cas9 all work by the same mechanism: removing stabilizing nonspecific contacts, which raises the R-loop propagation landscape (increases $S$), which increases $\eta$ and decreases $P_{\mathrm{o}\mathrm{n}}$

[E, T]. The observed pattern — improved specificity accompanied by reduced or context-dependent on-target activity, particularly for suboptimal guides — is not a series of independent engineering compromises. It is T-10, measured five times.

[C-2] All existing high-fidelity SpCas9 variants lie on a single one-parameter curve in $(\eta ,P_{\mathrm{o}\mathrm{n}})$ space, parameterized by the mean shift in R-loop propagation free energy. Variants that appear to beat the curve are either (a) exploiting a second checkpoint, or (b) measured with insufficient off-target sensitivity.

C-2 is directly falsifiable and is the central testable claim of this paper.

5.4 Escaping the ceiling: the price of a second gate

Hopfield–Ninio kinetic proofreading achieves discrimination ${\eta }_n$ for $n$ sequential checkpoints, each with an irreversible discard branch, at a cost of one free-energy quantum per discard.

Cas9 has essentially one checkpoint: the conformational transition in which the HNH nuclease domain docks onto the scissile phosphate, allosterically gated by REC3 sensing of the PAM-distal heteroduplex [E]. And Cas9 hydrolyzes no ATP during interrogation. Its only irreversible step is cleavage itself — which is after the decision, not before it. Cas9 is an energy-unsupplied, single-checkpoint proofreader. T-10 applies with full force.

The escape routes are therefore predictable from the theory, and all of them have been found empirically:

• —Strategy Mechanism in T-10 terms Measured gain Price
• —Paired nickases (Cas9n D10A ×2) Two independent searches must both succeed:
• —${\eta }^1{\eta }^2$

50–1,500× [E] 2 guides; reduced efficiency; risk of unpaired nicks dCas9–FokI dimers Obligate dimerization = independent AND gate ~10³–10⁴× [E] Strict spacing constraints; narrow targeting range Truncated guides (17–18 nt) Reduces $S<$

, steepens effective landscape 10–10³× [E] Lower on-target activity (T-10 again) High-fidelity variants Raises landscape; increases $S$ 10–10³× [E] On-target loss, guide-dependent RNP + short exposure Reduces $N{t}_{\exp }$

(T-7), not $\eta$ 2–20× [E] Delivery constraints Anti-CRISPR timed shutoff Truncates $t_{\exp }$

variable [E] Timing control required

[T-11] Every specificity gain exceeding the single-checkpoint ceiling $e_{ϵ\mathrm{/}{k}_B}T$ per mismatch has been achieved by introducing a second independent molecular requirement — a second guide, an obligate dimerization, or a temporal gate. There is no counterexample in the literature.

T-11 is the paper's strongest structural claim and the one most worth attacking.

5.5 The unexploited currency

T-10's escape clause names a currency the field has not spent: external free energy during interrogation.

Type I CRISPR systems do this. Cascade surveils, and Cas3 — an ATP-dependent helicase-nuclease — is recruited only upon a validated R-loop. The ATP hydrolysis provides exactly the irreversible discard energy that Hopfield's scheme requires. Type I systems are reported to exhibit high specificity, and the theory says they should [E, C].

[C-3] An engineered Cas9 coupled to an ATP-consuming conformational cycle — a fused helicase, a chaperone-like reset, or a designed futile cycle that discards non-committed complexes — could exceed the single-checkpoint ceiling without the on-target loss mandated by T-10, because the discard energy comes from ATP rather than from the binding landscape. This is the one route to breaking the specificity–activity conjugacy that does not require a second guide.

[O-2] No engineered ATP-driven proofreading module has been demonstrated for a Class 2 nuclease. Whether the required conformational coupling is achievable by fusion or requires ground-up design is open.

Sections 2–5 concern what the enzyme does. This section concerns what we can know about what it did. It is the harder problem, and it is where the field's reporting conventions are weakest.

6.1 A taxonomy with failure modes Class Methods Sensitivity Blind to In silico Cas-OFFinder, CFD, Elevation, CRISPR-Net, DeepCRISPR, GuideScan2, language-model predictors N/A (ranking only) Bulges (mostly); chromatin; novel motifs; anything outside training distribution Cell-free biochemical CIRCLE-seq, SITE-seq, Digenome-seq, ONE-seq, CHANGE-seq Very high (~10⁻⁵ of input DNA) Chromatin entirely; over-predicts by 10–100× Cell-based DSB capture GUIDE-seq, DISCOVER-seq, BLISS, TAG-seq ~0.1% [E] Rare events; non-DSB editors; structural variants Targeted deep sequencing rhAmpSeq, amplicon NGS ~0.1–1% (error-limited) Anything not on the panel; large events (dropout) Consensus sequencing Duplex-seq, UMI consensus, CRISPR-amplification enrichment 10⁻⁶–10⁻⁷ Anything not on the panel; cost-limited to small panels Structural CAST-seq, LAM-HTGTS, optical mapping, long-read Moderate Small indels Single-cell scDNA-seq of edited populations Cell-limited Genome-wide depth per cell

The critical structural observation: no method in this table is simultaneously genome-wide, chromatin-aware, and sensitive below 10⁻⁴. The cell-free assays give you genome-wide and sensitive but not chromatin-aware. The cell-based assays give you genome-wide and chromatin-aware but not sensitive. The consensus methods give you chromatin-aware and sensitive but not genome-wide — they require a candidate list, which comes from the methods that lack sensitivity.

[T-12] (Detection trilemma). Genome-wide coverage, chromatin fidelity, and sub-10⁻⁴ sensitivity cannot currently be obtained simultaneously. Every published specificity claim occupies one vertex of this triangle and is silent about the other two.

6.2 The detection floor is set by chemistry, not depth

A common misconception is that off-target sensitivity is a sequencing-depth problem: sequence deeper, see rarer events. This is false past a certain point.

Standard Illumina per-base error rates are ~10⁻³ per base. To call a variant at allele frequency $f$, the signal must exceed the error background. With $N$ reads, the observed variant count is Binomial ( 𝑁 , 𝑓 + 𝜖 ) Binomial(N,f+ϵ) where $ϵ$ is the error rate. As $N\to \mathrm{\infty }$, the estimate converges to $f+ϵ$, not to $f$. Depth reduces variance; it does not reduce bias.

𝑓 detectable ≳ 𝜖 regardless of 𝑁 f detectable

≳ϵregardless of N

The floor is broken only by changing $ϵ$: unique molecular identifiers and duplex consensus require independent observation of both strands of an original molecule, driving $ϵ$ to ~10⁻⁷–10⁻⁹ [E]. Prediction-guided enrichment approaches — using the nuclease itself to destroy wild-type molecules and enrich mutant ones before amplification — report 1.6–984-fold sensitivity gains over standard targeted amplicon sequencing [E].

[T-13] Off-target detection sensitivity is bounded below by the sequencing chemistry's error rate, not by read depth. Claims of "no detectable off-targets" from standard NGS are statements about a ~10⁻³ floor and carry no information about the 10⁻⁴–10⁻⁷ regime.

6.3 The therapeutic gap

This matters because of a simple arithmetic mismatch between assay scale and dose scale.

An autologous ex vivo cell therapy infuses on the order of $3\times {10}^6$ CD34⁺ cells/kg, i.e. $\sim 2\times {10}^8$ cells for an adult. Consider an off-target edit occurring at frequency $f$:

Off-target frequency $f$ Detectable by Edited cells in a 2×10⁸ dose 10⁻¹ Anything 2 × 10⁷ 10⁻³ GUIDE-seq (marginal) 2 × 10⁵ 10⁻⁴ Duplex-seq only, if on panel 2 × 10⁴ 10⁻⁶ Duplex-seq only, if on panel 2 × 10² 10⁻⁸ Nothing currently 2

[T-14] A therapeutic dose delivers 10⁵ cells carrying an off-target edit that sits exactly at the detection floor of the most widely used unbiased assay. Absence of evidence at 10⁻³ is not evidence of absence at therapeutic scale.

[C-4] The clinically relevant risk regime is $f\in [{10}^{-6},{10}^{-3}]$ — above the level at which a single edited clone is improbable, below the level at which standard assays report anything. This regime is systematically under-measured. We conjecture that the majority of clinically deployed guides have $\ge 1$ off-target site in this window.

C-4 is uncomfortable and should be tested rather than assumed. The test is specified in §9.

We state clearly what C-4 does not claim: it does not claim harm. Most off-target edits are in non-coding, non-regulatory DNA and are biologically inert. The claim is epistemic — that we currently do not know, and that our reporting conventions obscure the not-knowing.

6.4 Structural variants: the blind class

Short-amplicon sequencing has a failure mode that is worse than insensitivity: it is anti-correlated with the severity of the event.

CRISPR-induced DSBs produce, at appreciable frequency, kilobase-scale deletions, inversions, translocations, chromothripsis, and whole-chromosome-arm loss [E]. When such an event removes a primer binding site, the amplicon simply fails to form. The affected allele is not sequenced. The measured indel spectrum is computed only over surviving amplicons.

[T-15] Amplicon-based editing quantification is systematically biased against detecting the most consequential outcomes, because large deletions cause allelic dropout. The bias direction is: the larger the event, the less likely it is to appear in the data.

This is not a subtle statistical issue. It means the standard readout for editing outcome — indel frequency by amplicon NGS — is a measurement conditioned on the event being small.

6.5 Base and prime editors: the search problem becomes delocalized

Base editors and prime editors do not create DSBs, which removes the substrate that GUIDE-seq, DISCOVER-seq, and BLISS detect. Off-target detection for these tools requires entirely different assays.

Worse, base editors have an off-target class that is not sequence-addressed at all. Cytosine base editors exhibit Cas9-independent deamination: the deaminase acts on transiently single-stranded genomic DNA arising from replication and transcription, anywhere in the genome, uncorrelated with the guide [E]. Whole-genome sequencing of cells expressing evolved CBEs shows C→T and C→G mutations bearing an APOBEC mutational signature outside all predicted off-target sites [E]. Engineering has substantially mitigated this — narrowed-window variants such as YE1 show background-level Cas9-independent editing [E] — but the class exists and is invisible to any guide-directed assay.

[T-16] For base editors, the off-target set is not defined by sequence similarity to the guide. Guide-directed off-target search is therefore structurally incapable of bounding total off-target burden, regardless of its sensitivity. The problem changes from a localized search to a delocalized one, and requires whole-genome or signature-based assays.

Recent assay development responds directly to this. R-loop-detection methods — labeling transient single-stranded DNA genome-wide rather than detecting breaks — target the actual mechanistic intermediate that base editors exploit, and are being developed specifically because standard guide-directed workflows fail to capture the newly characterized editing outcomes or to satisfy emerging regulatory requirements [E]. Single-cell DNA sequencing approaches that resolve on- and off-target editing in the same cell add a dimension no bulk assay provides: whether off-target edits co-occur with on-target edits in the same genome [E].

[O-3] No assay currently bounds total genome-wide mutational burden from a base or prime editor at therapeutic sensitivity. Whole-genome duplex sequencing of edited primary cells at sufficient depth is technically possible and economically prohibitive at present.

Cell-free assays over-predict off-targets by 10–100×. Cell-based assays under-detect them. The reconciling variable is chromatin.

Nucleosomes strongly impede Cas9 target engagement; naked DNA and nucleosomal DNA differ by roughly an order of magnitude in accessibility, and nucleosome breathing at entry/exit sites — not steady occupancy — governs much of the residual access [E]. Roughly 75–80% of genomic DNA is nucleosome-associated.

This produces a specific, testable structure:

• —Cell-free assay finds a site, cell assay does not → likely closed chromatin. Prediction: the site becomes editable if chromatin is opened (e.g., in a different cell type, during replication, or under a remodeler perturbation).
• —Cell assay finds a site, cell-free assay does not → likely a bulged or non-canonical site missed by the biochemical pipeline's alignment.

[C-5] The observed 10–100× over-prediction by cell-free assays is quantitatively accounted for by nucleosome occupancy alone, with no additional free parameters, if occupancy is measured in the same cell type by ATAC-seq or MNase-seq. This is directly testable and, to our knowledge, has not been done as a closed quantitative comparison.

The clinical implication is uncomfortable and follows immediately: off-target profiles are cell-type-specific. A guide validated in HEK293T carries no guarantee in CD34⁺ HSPCs, and a guide validated in quiescent cells carries no guarantee in cycling cells where replication transiently exposes the entire genome.

[T-17] Because chromatin gates accessibility and chromatin differs by cell type and cell-cycle phase, off-target specificity is not a property of a guide. It is a property of a (guide, cell type, cell state) triple. Reporting specificity as a per-guide scalar is a category error.

The results above are not independent. They form a coupled system with a small number of controls and no free lunches.

Controls available to the engineer:

— dose and exposure Landscape steepness $S$ — nuclease variant, guide truncation PAM degeneracy — nuclease choice Checkpoint count $n$ — paired nickase, dimerization, timed shutoff Chromatin state — cell type, cycle phase (largely not controllable) Assay sensitivity $ϵ$ — measurement only, does not change biology

Coupling relations:

Push this Get this Pay this Source

↑ on-target completeness ↑ off-target burden ↑ (linear) T-7

$S$ ↑ (hi-fi variant) $\eta$ ↑ (bounded by $e_{ϵ}$) $P_{\mathrm{o}\mathrm{n}}$

↓ T-10 PAM degeneracy ↑ addressability ↑ decoy census ↑, search tax ↑ T-3, T-5

↓ (tighter) — nothing search time ↑↑ T-5 checkpoints $n$ ↑ ${\eta }_n$ a second guide, a dimerization, or ATP T-11, C-3

$ϵ$ ↓ knowledge ↑ cost; panel-limited, not genome-wide T-12, T-13

Row four is the notable one. Tightening PAM affinity buys nothing and costs a great deal — a rare case of a control with no favorable direction.

Row five is the only route that increases $\eta$ without paying in $P_{\mathrm{o}\mathrm{n}}$

, and it is the only row whose currency (external free energy) has never been spent in a Class 2 system.

The central thesis, stated plainly: the target detection problem admits no solution that improves specificity without payment. But every payment is identified, finite, and in a named currency. The problem is not that precise editing is forbidden. The problem is that the field has been paying in $P_{\mathrm{o}\mathrm{n}}$

(the expensive currency, per T-10) when a cheaper currency — ATP, per C-3 — has been sitting unspent because Class 2 systems inherited an architecture that never needed it.

Consistent with the discipline that a claim without a failure condition is not a claim, here are the specific experiments that would break the arguments above.

9.1 Test of C-2 (the single-curve conjecture)

Design. Measure $(\eta ,P_{\mathrm{o}\mathrm{n}})$ for a panel of ≥8 SpCas9 variants (WT, eSpCas9, HF1, Hypa, evo, Sniper, plus two designed intermediates) across ≥20 guides spanning a range of predicted R-loop stability, in a single cell type, using a matched off-target assay at ≥10⁻⁵ sensitivity.

Prediction. All variants collapse onto a one-parameter curve $\eta (P_{\mathrm{o}\mathrm{n}})$ with the functional form implied by $P_{\mathrm{o}\mathrm{n}}=1\mathrm{/}(1+S)$, $\eta =1+(e_{ϵ}-1)S\ge \mathrm{/}(1+S)$.

Falsified if. Any single-checkpoint variant lies significantly above the curve at matched $P_{\mathrm{o}\mathrm{n}}$

, and the excess is not attributable to a second molecular requirement.

9.2 Test of T-5 (the PAM affinity ceiling)

Design. Engineer a graded series of PAM-interacting-domain variants with $K_{D_{PAM}}$

spanning 10 μM to 10 mM (measurable by ITC or single-molecule dwell-time on PAM-only substrates). Measure genome-wide search completion time by live-cell single-particle tracking and time-to-half-maximal editing.

Prediction. Search time scales as $(1+c_{PAM}\mathrm{/}{K}_D)$, with the ~1.3 mM crossover clearly visible. Variants at $K_D<100\mu$ M show >10× slower target location despite unchanged catalytic parameters.

Falsified if. Tighter PAM binders search as fast or faster — which would imply that PAM decoy sequestration is not rate-limiting and that some unmodeled release mechanism operates.

9.3 Test of C-1 and C-4 (the detection-floor conjecture)

Design. For 5 clinically deployed guides, enumerate the full $k\le 4$ decoy census computationally with gapped alignment (addressing O-1). Assay every site by duplex-consensus sequencing at $\ge {10}^{-6}$ sensitivity in the relevant primary cell type. Compare against the published GUIDE-seq/CIRCLE-seq call sets for the same guides.

Prediction. ≥3× more true off-target sites are recovered than appear in the published call sets, concentrated in the 10⁻⁵–10⁻³ frequency band, with an enrichment of bulged sites among the newly detected.

Falsified if. The census is essentially exhausted by existing methods — which would mean the biology, not the assay floor, is doing the suppressing, and C-1 is wrong.

9.4 Test of C-5 (chromatin reconciliation)

Design. Same cell type, same guide, same day: CIRCLE-seq (chromatin-free), GUIDE-seq (chromatin-aware), ATAC-seq (accessibility). Fit a single accessibility-weighting parameter.

Prediction. A one-parameter accessibility model reconciles the two off-target lists to within a factor of 2 across ≥80% of sites.

Falsified if. Residual discrepancy requires site-specific parameters — implying an additional unmodeled variable.

9.5 Test of C-3 (the ATP escape route)

Design. Construct a Cas9 fused to an ATP-dependent conformational module positioned to reset non-committed R-loops (candidate architectures: helicase fusion at the REC lobe; designed futile cycle coupled to the HNH gate). Measure $(\eta ,P_{\mathrm{o}\mathrm{n}})$ against the C-2 curve.

Prediction. The construct lies above the curve — higher $\eta$ at matched $P_{\mathrm{o}\mathrm{n}}$

— with the offset scaling with ATP concentration and vanishing in a Walker-B-dead mutant.

Falsified if. No offset, or an offset that persists in the ATPase-dead control (implying the gain came from steric effects, not proofreading energy).

This is the experiment we consider most valuable and least likely to have been attempted.

Tags: T = theorem / derived from stated premises · E = empirical, established in literature · C = conjecture, testable, not established · O = open problem

Claim Tag Load-bearing for Falsifier

T-1 Guide + PAM carry 44 bits vs 32.6 required; information is not limiting T §2 entire Arithmetic T-2 Effective address is 22–30 bits; deficit of 3–10 bits T (from E weights) Off-target origin Mismatch-tolerance profiles showing uniform positional weighting T-3 PAM relaxation costs decoy census linearly T Nuclease choice — T-4 Single-RNP idealized search time ≈ 3.7 h T §4 budget Diffusion constant off by >10× T-5 PAM affinity ceiling at ~1.3 mM decoy concentration T §4.2; design guidance §9.2 T-6 Cas9 sliding gain ≈ 10×, not 10²–10³ T + E Search budget Measured $l_{\mathrm{s}\mathrm{l}}\gg 27$ bp T-7 Dose and specificity are conjugate: both scale with $Nt$ T Delivery strategy Regime where off-target saturates but on-target does not T-8 Equilibrium binding cannot discriminate (saturation argument) T §5 entire — T-9 Seed effect derives from landscape topology, not structure T Generality across nucleases A nuclease with downhill landscape and no seed T-10 Discrimination ceiling: $\eta \le e_{ϵ\mathrm{/}{k}_B}T$; $\eta \to {\eta }_{\max }\Rightarrow P_{\mathrm{o}\mathrm{n}}\to 0$ T Whole paper §9.1 T-11 All super-ceiling gains came from a second independent requirement T (empirical survey) §8 synthesis A single-checkpoint counterexample T-12 Detection trilemma: genome-wide ∧ chromatin-aware ∧ <10⁻⁴ unattainable T §6 entire An assay achieving all three T-13 Detection floor set by $ϵ$, not depth T Interpretation of null results — T-14 Therapeutic dose delivers ~10⁵ cells at the assay floor T Clinical framing Dose arithmetic T-15 Amplicon assays biased against large events (dropout) T + E Outcome measurement — T-16 Base-editor off-target set is not sequence-addressed T + E §6.5 Demonstration that Cas9-independent class is negligible in all deaminases T-17 Specificity is a (guide, cell type, cell state) property, not a guide property T Reporting conventions Cell-type-invariant off-target profiles C-1 Detection floor, not biology, dominates the census–observation gap C §6.3 §9.3 C-2 All hi-fi variants lie on one $(\eta ,P_{\mathrm{o}\mathrm{n}})$ curve C §5.3 §9.1 C-3 ATP-driven proofreading escapes the specificity–activity conjugacy C §5.5, §8 §9.5 C-4 Clinically relevant risk regime ${10}^{-6}<f<{10}^{-3}$ is systematically under-measured C §6.3 §9.3 C-5 Nucleosome occupancy alone reconciles cell-free vs cellular assays C §7 §9.4 O-1 Genome-wide contribution of bulged off-targets is unquantified O §2.4 — O-2 No ATP-driven proofreading module demonstrated for Class 2 O §5.5 — O-3 No assay bounds total mutational burden for base/prime editors O §6.5 — O-4 The value of $ϵ$ (per-mismatch penalty) is not independently measured in the R-loop context — literature values are inferred from cleavage data, making T-10 partly circular O T-10 quantitation Direct thermodynamic measurement

Note on O-4. This is the weakest joint in the paper and should be stated plainly. T-10's form is derived and secure. Its numerical ceiling depends on $ϵ$, which is extracted from the same cleavage assays the theorem is meant to explain. An independent measurement — magnetic-tweezer R-loop stability on matched and mismatched substrates, or calorimetry on isolated heteroduplexes — would remove the circularity. Until then, treat ${\eta }_{\max }\approx 7$–55 as an estimate, not a derived constant.

[O-2 / C-3] The ATP question. Can a Class 2 nuclease be given an external free-energy source for proofreading? This is the only identified route past the specificity–activity conjugacy that does not cost a second guide. Highest leverage in the paper. [O-4] Independent measurement of $ϵ$. Removes the circularity in the discrimination ceiling and converts T-10 from a structural theorem to a quantitative one. [T-12] Breaking the detection trilemma. A genome-wide, in-cellulo, sub-10⁻⁵ assay. Whole-genome duplex sequencing is the brute-force answer; the interesting question is whether a smarter one exists — e.g., enrichment on the mechanistic intermediate (R-loops, single-stranded DNA) rather than on a predicted candidate list. [O-1] Gapped-alignment census. Cheap, computational, and nobody has closed it at therapeutic sensitivity. [T-17] Cell-state-resolved specificity. If specificity is a triple, the field needs a specificity surface, not a specificity number. This is a data-infrastructure problem as much as a biology problem. [Unlisted] The functional-consequence question. Most off-target edits are inert. Nothing in this paper addresses which ones are not. A specificity theory that cannot rank consequences is incomplete — a 10⁻³ edit in a gene desert and a 10⁻⁶ edit in a tumor suppressor are not comparable quantities, and we have been treating them as though they were.

The analysis above is internally consistent. The information budget balances. The search-time budget balances. The discrimination ceiling follows from the splitting-probability algebra, and the escape routes it predicts are exactly the escape routes the field found empirically without knowing they were predicted.

That internal consistency is worth something, and it is worth less than it feels like it is worth.

The model treats the R-loop as a one-dimensional walk on 20 states with additive mismatch penalties. Real R-loop propagation involves DNA supercoiling, protein conformational states that are not slaved to $n$, sequence-dependent stacking, and a genome that is not a naked polymer but a phase-separated, protein-dense, actively transcribed and replicated object. The chromatin term in §7 is a single scalar standing in for the entire field of nuclear organization. The $ϵ$ in T-10 is fitted, not measured (O-4).

What the analysis establishes is not that CRISPR specificity is understood. It is that the structure of the problem is a permission structure: a set of coupled constraints, each of which names a currency and a price, none of which forbids the outcome. The apparent impossibility — addressing one site in 6.4 billion with a 20-letter word and no fuel — dissolves into a bill. The bill is large, itemized, and in several places has not yet been read.

The most consequential line on it is the one nobody has paid: Cas9 proofreads for free, and gets what free proofreading buys.