What Can a Replication Movie Actually Distinguish?

1. Research question and scope

The starting question is an inference problem: when a polymerase pauses, do the measured signals identify what happened, or merely show that synthesis slowed? Published single-molecule studies motivate separating movement from molecular presence. Graham and colleagues measured stochastic synthesis pauses and slower helicase unwinding; Lewis and colleagues directly observed polymerase exchange. Neither observation makes every candidate mechanism distinguishable in every movie. [2] [3]

This study evaluates the information supplied by three idealized channels: helicase position H, the synthesis-front position P, and occupancy O. P denotes the front of synthesis, not the position of a detached protein. The scope is local DNA replication. It does not establish a transcription mechanism or explain genome-wide completion.

2. Models and observation protocol

All three models pause P. In A, H slows during the pause and immediately returns to full speed afterward, leaving a separation. In B, H remains slow until P catches up. In C, H and P follow B exactly, but occupancy falls from one to zero during the pause. A and B retain occupancy one. Thus B and C cannot be separated using H/P movement alone: that indistinguishability is built into the hypotheses, not discovered by the calculation. [1]

Normal and reduced H speeds were fixed at 510 and 84 nucleotides per second. Each independent molecule contained one candidate pause, beginning uniformly between 8 and 12 seconds. Pause duration followed an exponential distribution with known mean two seconds; a twelve-second sensitivity condition tested longer pauses. Each sequence covered 60 seconds with 0.4-second exposures. Position and occupancy errors were independent Gaussian noise with assumed standard deviations 100 nucleotides and 0.35. Constant position offsets were removed through contrasts.

The seven conditions comprised four cadences (0.7, 2.2, 4, and 8 seconds), two longer-pause conditions, and one noise-mismatch condition. Each had three true hypotheses, 100 groups per hypothesis, and 50 molecules per group. The comparison holds exposure duration and per-frame noise fixed; total frame count and total light dose vary with cadence.

3. Inference and decision rules

Likelihoods integrate over uncertain pause onset and duration instead of choosing whichever nuisance values fit best. Duration integration accounts for finite exposure and retains the probability of pauses extending beyond the recording window. Onset integration uses numerical quadrature. Equal model priors define pairwise comparisons: A/B uses H/P; B/C uses H/P/O.

For a homogeneous group, independent per-molecule log Bayes factors are summed. A correct-and-confident decision requires posterior probability at least 0.95 for the true model. The tested group sizes were 1, 2, 5, 10, 20, and 50; their nested prefixes make these comparisons correlated. Mixture analysis instead integrates the likelihood (1−θ)L_B + θL_C for each molecule with a uniform Beta(1,1) prior. Here θ is a fraction of molecules assigned to C, not time occupancy or a cell-wide detachment rate. [1]

4. Results: discrimination and precision

At four-second sampling, the first tested group sizes reaching the observed 95% correct-and-confident target were 10 for A/B and 20 for B/C. At eight seconds, B/C reached only 174/200 correct-and-confident groups at N=50, with a Wilson 95% interval of 81.6–91.0%. Its ordinary classification accuracy was higher, 196/200: accuracy and justified confidence are different endpoints. A result of 200/200 has a Wilson lower bound of 98.1%, not certainty. [1]

Homogeneous discrimination did not imply a precise mixture estimate. At four-second sampling and N=50, a true θ=0.5 produced mean estimate 0.505 and RMSE 0.108, but average 95% interval width 0.455. At eight seconds that width increased to 0.768. The finite-sample uncertainty remains material even when a homogeneous group is readily classified.

In the mismatch condition, actual noise rose to 150 nucleotides and 0.70 while inference retained the original assumptions. At four seconds, single-molecule B/C expected calibration error increased from 0.0060 to 0.1192. Among 200 groups of 50, 198 were correctly confident, one incorrectly confident, and one undecided. Pooling helped in this particular condition, but it did not validate the observation model.

5. Numerical verification and limitations

A saved sensitivity audit compared standard and finer integration on the same 1,050 sequences before the main calculation. Its predefined tolerance checks passed without changes to the confidence decisions. The saved artifact verifier recorded 493 passing checks. These are computational consistency checks; this publication did not rerun simulations or collect new laboratory data. [1]

The favorable assumptions include known speeds and pause distribution, calibrated binary occupancy, independent errors, and a preselected pause window. Bleaching, labeling efficiency, intensity drift, shared errors, unknown speeds, and systematic event-selection bias are omitted. The eight-second cadence is coarser than the four-second onset window, and sensitivity to sampling phase was not tested. A cadence threshold or molecule count should therefore not be transferred directly to an instrument specification.

6. Implications and evidence availability

The practical contribution is a way to compare acquisition choices while reporting confidence, mixture precision, and model mismatch separately. Observing polymerase presence or exchange is established prior work. Lewis et al. combined fluorescent polymerase with DNA-extension measurements; that assay does not automatically supply independent H/P positions and a binary occupancy signal for one synthesizing polymerase. A usable empirical test requires a channel mapping and an experimentally calibrated noise model. [3]

This article is a non-peer-reviewed research record. A newly curated public evidence file contains selected results, assumptions, and hashes identifying retained source artifacts. The full raw sequences and analysis environment are not released here, so the summary alone is not a complete reproducibility package. No real-data mechanism-classification result is claimed. [1]