One-Pass Bias-Corrected Isoform Quantification with a Certified Refinement Step

Counting RNA molecules by sequencing is distorted by the lab steps that prepare the sample. Amplification favors short fragments, enzymes prefer certain sequences, and read coverage piles up at particular positions along a transcript. The methods that correct for these distortions best do so by iterating. They alternate between guessing how much of each transcript (each RNA variant, or isoform, of a gene) is present and re-estimating the distortion, and they repeat until the two agree. That loop is affordable for a handful of bulk samples. It is not affordable for single-cell atlases with millions of cells, and each cell has too few reads to fit its own distortion reliably anyway. Single-cell practice has therefore dropped transcript-level correction and clusters cells on raw read-compatibility counts.

We propose to train a neural network that replaces the loop with one forward pass. The network reads a gene's compatibility counts and its exon structure, and it outputs bias-corrected isoform abundances together with the bias parameters. It is trained to minimize the same count likelihood the iterative solvers optimize, pooled across many cells, so it learns shared structure that a per-cell solver cannot see. A single cheap refinement step then acts as a certificate. If one step of the classical solver would improve the likelihood by more than a threshold, the system runs more steps for that gene and cell, and otherwise it stops. The goal is to recover most of the accuracy of the slow correcting solvers at close to the speed of quantification-free clustering. The certificate also says, per gene, when the fast answer can be trusted.

Core Claim

Amortized inference is our term for training one network to solve many instances of an optimization problem, instead of solving each instance from scratch. Our claim is that amortized inference over the iterative bias-and-abundance problem, trained on the likelihood rather than only by imitating a solver, matches converged iterative solvers on bias-heavy simulated and gold-standard data. We further claim it beats them on low-depth single cells, because pooling acts as a learned prior on bias.

The main contrast is with speed-oriented work that collapses iterations. DCFold shows that recycling and denoising loops in protein structure prediction can be distilled into a single pass. It warns, however, that this concentrates the output distribution, and it gives no signal for when the single pass is wrong. That warning matters here: when isoforms share most of their exons, the likelihood is flat, and a concentrated single-pass answer would hide real ambiguity. Our design addresses this with the likelihood-gap certificate and with uncertainty outputs that we check against the solver's own variability. Meta-Reflection likewise compresses an iterative critique loop into one pass, but only for language tasks, and it also has no stopping signal. The adaptive stopping rule in Deterministic Non-Autoregressive Neural Sequence Modeling by Iterative Refinement, which halts when consecutive drafts stop changing, is the nearest analogue to our certificate. We replace its heuristic draft distance with a quantity that has meaning in this problem: the likelihood improvement one solver step would achieve.

Research Program

The statistical model comes from the bias-modeling literature. For one gene, let $c_e$ be the number of reads in equivalence class $e$, meaning the set of reads compatible with exactly the same subset of transcripts. Let $\theta_t$ be the abundance of transcript $t$, and let $\beta$ be the bias parameters. The expected count is

$$\lambda_e(\theta,\beta) = \sum_{t \ni e} \theta_t \, \ell_{t,e}(\beta),$$

where $\ell_{t,e}(\beta)$ is the bias-adjusted effective length of class $e$ within transcript $t$. This quantity folds in positional, sequence, and fragment-length effects.

Two prior models supply the pieces. PGseq places Gamma-distributed exon-level bias on a Poisson rate, which yields a negative binomial likelihood. MultiSplice alternates between abundance and bias updates, with a lasso penalty that pushes unexpressed isoforms to zero. We adopt the negative binomial likelihood with a sparsity penalty as the shared objective $L(\theta,\beta; c)$. The alternating solver on this objective serves as both teacher and fallback.

The network is a set-encoder over a gene's splice graph. Its inputs are each equivalence class's count, its transcript membership, and its positional and sequence features. Its outputs are $\theta$, $\beta$, and a per-transcript log-variance. Bias parameters are split into two parts: a component shared across a library batch, which is estimated from all cells in that batch, and a per-cell deviation that shrinks toward it. This split is where pooling helps. The conservation assumption behind PGseq, that bias is stable across samples, becomes a learned and testable quantity rather than a fixed modeling choice.

Training minimizes the expected $L$ over many simulated and real genes and cells. We add an optional term that matches the converged solver on genes where the solver is well identified.

At inference, the network output initializes the solver, and we run one alternating step. If the relative likelihood improvement exceeds a threshold $\tau$, we continue iterating to convergence. The threshold $\tau$ sets where the system sits on the accuracy–latency curve, and sweeping it traces that curve.

Baselines

The naive reference point is uniform-coverage, length-normalized quantification with no bias model, as used in the segment-wise model evaluated by rQuant. It runs in negligible time, which fixes the floor for both accuracy and cost. We will run it on the same equivalence-class input. Beating it is necessary but not interesting on its own.

The strongest accuracy competitor is the converged iterative solver on our shared negative binomial and lasso objective. It stands in for the MultiSplice- and PGseq-style methods. We will also run rQuant's position-wise quadratic program on bulk data. On Flux Simulator data, rQuant reports correlation with true abundance of 0.820 across genes and 0.635 within genes. On MAQC paired-end data, PGseq reports squared correlation with qRT-PCR of up to 0.860, against 0.800 for RSEM and 0.723 for Cufflinks. On its low-expression group PGseq reports 0.492–0.530, while competitors fall to 0.20–0.35. No reported result was found on Lacuna for MultiSplice's own benchmarks. Solvers and network see identical reads and annotations. Solvers get tuned penalty and convergence settings on the same validation genes that tune the network. We report wall-clock time and core-hours on the same hardware. Beating the solver means one of two things. On bulk data, the network plus certificate must stay within 0.01 correlation of the converged solver at a tenfold lower median cost. On single cells at low depth, it must exceed the solver's accuracy.

The strongest speed competitor is transcript-compatibility-count clustering from Ntranos et al., which reports about 2.86 core-hours on the 3,005-cell Zeisel dataset, against 88 for HISAT and 433 for Bowtie pipelines. We will also run its reverse pipeline: pool reads within each cluster, quantify with bias correction, and give each cell its cluster's isoform proportions. Our method must reach comparable core-hours while producing per-cell isoform estimates that beat the reverse pipeline.

Three ablations isolate the key ingredients. The first trains by pure imitation of the solver with no likelihood term. The second removes cross-cell pooling of bias parameters. The third removes the certificate and always stops after the single pass.

Why not just…?

Why not just warm-start the iterative solver and cap it at a few iterations? This is the most serious threat. Alternating solvers often do most of their work in the first steps, and a good initialization from pooled clusters might make the network unnecessary. Our accuracy–cost curves will include truncated solvers at one to twenty iterations, initialized from cluster-pooled estimates. If truncated solvers match the network plus certificate at equal cost, the network adds nothing, and the contribution reduces to the certificate as a stopping rule. We would report it that way. The case where we expect the network to win is low per-cell depth. There, the per-cell likelihood is nearly flat, so more iterations fit noise, whereas pooled training supplies a prior that the solver does not have.

Why not just use the cluster-pooled reverse pipeline? It corrects bias once per cluster and is cheap. However, it assigns every cell in a cluster the same isoform proportions, so by construction it cannot detect isoform switching within a cluster. Ntranos et al. themselves note that pooling brings back the unresolved bias and identifiability problems. Our simulations will plant within-cluster isoform gradients and test whether the network recovers them while the reverse pipeline does not.

Why not skip bias correction altogether? The file's evidence is strong on this point. In rQuant's comparison, correlation within genes falls from 0.635 to 0.367 without position-wise profiles. In PGseq's comparison, differential-expression AUC on colorectal cancer data falls from 0.912 to 0.422 for Cufflinks. We include the uniform model to confirm that this gap survives in our single-cell simulations.

Why not just sequence native RNA and avoid library bias? Nanopore direct RNA sequencing, as in Coral, removes amplification bias. It moves the cost into slow neural basecalling, and existing short-read single-cell atlases cannot be re-sequenced. This proposal targets the data that already exists.

Proposed Experiments

Bulk accuracy. On Flux Simulator data with injected positional, fragment-length, and sequence bias, we measure across-gene and within-gene correlation with truth, alongside core-hours, against rQuant's 0.820 and 0.635. On MAQC, we measure $R^2$ with qRT-PCR overall and on the low-expression group, against PGseq's 0.860 and 0.492–0.530.

Single-cell accuracy. Isoform ground truth does not exist for real single cells, so we simulate cells with Flux Simulator at depths down to 1% of bulk, matching the downsampling regime in Ntranos et al. We plant cell-type-specific and within-cluster isoform differences and score per-cell isoform error.

Scale. On the Zeisel data we report total core-hours and check that clustering on our isoform estimates preserves the published subtype structure.

Decisive checks. Three results decide the claim. The first is whether, at matched core-hours, the network plus certificate dominates truncated warm-started solvers at low depth. The second is whether the certificate's flagged genes are exactly those where single-pass error is large, measured by rank correlation between the one-step likelihood gain and the true error. The third is whether predicted uncertainty is calibrated on genes whose isoforms share most exons, which we check by bootstrap coverage against the solver. A failure on the second or third check would reproduce DCFold's collapse concern in this setting, and that failure would itself be a publishable negative result.