After an acute kidney injury, some kidneys repair and others slide into fibrosis, the permanent scarring that drives chronic kidney disease. The decision to intervene has the same structure as a problem in live translation. Wait longer after the injury and the call on which way the kidney is heading becomes more reliable. Meanwhile the scar tissue sets and the window for treatment closes. Simultaneous machine translation has a mature toolkit for this "wait or commit" dilemma. The central tool is the wait-$k$ policy: a fixed rule that reads $k$ input words before emitting the first output, plus methods for measuring how quality falls as $k$ shrinks. We propose to use that toolkit to measure, for the first time, how reliability trades against earliness when fibrotic commitment is judged from RNA sequencing after kidney injury.
The part we think matters most is a second kind of delay that the translation analogy makes visible. Standard RNA sequencing measures the whole pool of transcripts in a cell, and that pool is dominated by stable, long-lived messages. So it reports changes in gene activity late. Nascent-RNA methods tag freshly made transcripts and see those changes almost immediately; Dyna-vivo-seq found 90 transcription factors changing activity within 5.3 hours of reperfusion that were hidden in total RNA. The question we pose is quantitative: at a fixed reliability target, how many hours or days earlier can a fibrosis call be made from nascent signal than from steady-state signal? We also ask whether early calls fail in the same way low-latency translators do, by falling back on learned priors instead of the evidence in the sample.
Core Claim
We claim that early fibrosis prediction should be reported as a frontier, not as a single accuracy at a chosen checkpoint. The frontier has two latency axes: time after injury at which tissue is sampled (the biological analogue of $k$) and the lag built into the measurement itself (nascent versus steady-state RNA). We expect nascent or nascent-proxy signal to shift the frontier toward earlier times for gene programs whose transcripts are long-lived. We expect little shift for programs built from short-lived transcripts. Concretely, let $Q_m(k)$ be the area under the ROC curve (AUROC) for calling late fibrosis from data gathered up to $k$ hours post-injury using measurement modality $m$. Let $q^*$ be a fixed reliability target. The quantity we estimate is the time bought by switching modality:
$$\Delta = \min\{k : Q_{\text{total}}(k) \ge q^*\} \;-\; \min\{k : Q_{\text{nascent}}(k) \ge q^*\}.$$
A positive $\Delta$ with a confidence interval excluding zero is the headline result. A $\Delta$ near zero is also informative, because it would say metabolic labeling does not buy intervention time for this disease.
What We Borrow From Simultaneous Translation, and What We Do Not
Three results from simultaneous translation supply method. STACL introduced wait-$k$ with "prefix-to-prefix" training, where one model is trained on every truncated input it will face at test time. It showed that restricting a full-context model at inference time collapses quality, while its wait-9 model came within 2.5 BLEU of full-sentence translation on Chinese–English. We reuse the idea of training and evaluating on truncated observation windows, and we sweep $k$ to trace the curve as STACL did. Dalvi et al. showed that fixed read/write schedules give a clean, reproducible reference curve, which is why we define $k$ by the clock rather than by a learned controller. Zhong et al. identified how low-latency decoders fail. When starved of source context, the model leans on its own prior outputs and produces fluent but unsupported text. They measured this with a ratio of target-side to source-side reliance, and found hallucination at 31.28% for wait-1 versus 11.50% offline on German–English.
We translate that last diagnostic rather than the architecture. A biological forecaster with little post-injury information can default to the population prior, for example "injured kidneys at this time usually progress." Its apparently confident calls would then be priors dressed as predictions.
We deliberately do not import autoregressive sequence-to-sequence decoders. The file makes clear why they would be overkill here. Kidney transcriptomes come from destructive sampling, so each animal yields one snapshot. The longest public time course, the 12-month ischemia/reperfusion series, has 10 time points. A "stream" of that length, measured on different animals at each point, does not support a learned read/write controller. The translation work also constrains us. Adaptive segmentation (Zhang et al., SASST) beats fixed wait-$k$ when information density varies. The 7 temporal gene modules in the ischemia series suggest density varies a great deal. So an adaptive stopping rule is a required comparison, not an afterthought.
Research Program
The first stage uses existing data only and needs no outcome labels. It asks how long each fibrosis-relevant gene program takes to become detectable in steady-state versus nascent signal.
In bulk RNA-seq, reads falling in introns come mostly from unspliced pre-mRNA. They are therefore a rough, free proxy for ongoing transcription, while exonic reads reflect the steady-state pool. We will first check this proxy against real metabolic labeling in the Dyna-vivo-seq dataset (GSE253836). That dataset has 27,590 single cells across pre-ischemia, ischemia and reperfusion, with new and old RNA separated by 4-thiouridine labeling and T-to-C conversion, at about 20.7% new RNA in vivo. The check is gene by gene and cell type by cell type: pseudo-bulk intronic counts are compared with labeled-new counts. Genes where the proxy fails are flagged and excluded from later nascent-proxy analyses.
We then take the ischemia/reperfusion time course (GSE98622), which runs from 2 hours to 12 months after a 21-minute bilateral ischemia with sham controls. For each of its 7 temporal modules, and for markers the authors tie to maladaptive repair (de novo Krt20 tracking Havcr1/Kim-1), we estimate the earliest time point at which intronic and exonic signal separate injured from sham kidneys. The difference is a per-program measurement lag. We expect it to scale with transcript stability. We can test that directly by correlating per-program lag with turnover rates estimated from the Dyna-vivo-seq new/old ratios. Dyna-vivo-seq also reported that proximal tubule cells with high baseline turnover are the most injury-susceptible, which gives a mechanistic reason the lag should differ across cell types.
The second stage adds fate labels and builds the frontier. Public data give only group-level contrasts in which the same injury leads to different fates. Examples are Isg15 knockout versus wild type (Cui et al.) and Butyrolactone I versus vehicle in folic-acid and obstruction models (Ha et al.). Replicate numbers there are small. Whether early-time profiles are even deposited for those arms must be checked. We will use them as pilot contrasts only.
The decisive data come from a new paired cohort. Mice receive graded ischemia durations so that outcomes spread between repair and fibrosis. Each animal gets an early small-tissue sample at one of several post-injury times, profiled with in vivo nascent labeling as in Dyna-vivo-seq. Library cost there is under $0.10 per cell. Late fibrosis in the same animal is scored by histology and collagen content at a fixed chronic endpoint. Serum creatinine and blood urea nitrogen are drawn at every early time point. Pairing early sample and late outcome within one animal removes the destructive-sampling objection. It is also the only way to label individual trajectories.
On this cohort, for each $k$ and each modality (total, intronic proxy, labeled nascent, clinical chemistry), we fit calibrated fate classifiers with animal-level cross-validation. From these we read off $Q_m(k)$ and $\Delta$.
The third stage is the prior-reliance diagnostic, adapted from Zhong et al. For each early call we compare the full model's prediction with a model that sees only time since injury and ischemia duration. We also measure how much the prediction changes when the sample's expression features are permuted within its time stratum. We call a prediction prior-driven when it tracks the covariate-only model and barely responds to permutation. The analogue of Zhong et al.'s finding would be that false fibrosis calls, made in animals that went on to repair, concentrate among prior-driven predictions at small $k$.
Baselines
The naive reference point is the covariate-only model. It predicts fate from ischemia duration and sampling time alone, with no molecular data. Every molecular curve must clear it, and it also anchors the prior-reliance diagnostic.
The clinical reference is serum creatinine plus blood urea nitrogen at the same time points and the same cross-validation folds. Ha et al. use these to report kidney damage. For their value as early predictors of fibrosis: no reported result found on Lacuna for this.
The strongest existing molecular approach is a single-checkpoint classifier on total bulk RNA. We give it both genome-wide features and the known injury markers (Havcr1, Krt20, the late modules from the ischemia time course, and ISG15/TGF-β receptor signalling from Cui et al.). It is the strongest competitor because it is what a nephrology lab would actually do. For its early predictive accuracy on fibrosis outcome: no reported result found on Lacuna for this. For the translation analogy, the numbers that motivate the design are STACL's wait-9 gap of within 2.5 BLEU of offline, and Zhong et al.'s 31.28% versus 11.50% hallucination rate. Neither is a biological baseline.
For fairness, every modality gets the same animals, folds, feature-selection budget, hyperparameter search size and model family. Nascent signal counts as beating total RNA only if $\Delta > 0$ with a bootstrap interval, over animals, excluding zero, at two preset targets $q^*$.
Two ablations isolate the proposal's ingredients. The first swaps labeled nascent RNA for the intronic proxy. This asks whether the cheap proxy recovers the gain, which would matter for reanalysing every existing bulk dataset. The second compares prefix-shared training (one model trained on all windows up to each $k$) with independent per-$k$ classifiers. This tests whether the one idea we take from STACL's architecture helps at these sample sizes.
Why not just…?
Why not just train a classifier at one early checkpoint, say 24 hours, on total RNA? That classifier is one point on our curve, and we include it. The trouble is that a single checkpoint hides the quantity a clinician needs: how much reliability is lost by deciding earlier, and how much time a different assay buys. We do not claim sequence models beat it. Our prefix-sharing ablation may well show independent per-time classifiers do equally well with this few animals. If so, we report that, and the contribution rests on the frontier and on $\Delta$, not on an architecture.
Why not just use creatinine and urea nitrogen, which are cheap and validated? These markers rise only after filtration capacity is already lost. The ischemia time course shows molecular reprogramming such as Krt20 and Havcr1 induction and lymphoid-structure programs appearing before scarring is established. Still, "cheaper and nearly as early" is a live possibility. So clinical chemistry is its own modality on the frontier. If its curve reaches $q^*$ within a day of the nascent curve, that is a practically important negative result.
Why not just use an adaptive rule that stops as soon as the model is confident, instead of a fixed wait? This is the most serious threat. Adaptive policies beat fixed wait-$k$ in translation when information density varies (Zhang et al.), and fibrosis clearly passes through distinct phases. We will run a confidence-threshold stopping rule on each modality. Following Iranzo-Sánchez, who showed average lag hides harmful latency spikes, we will report both the mean and the spread of decision times. Adaptive stopping does not replace the fixed-$k$ frontier; it is evaluated on top of it. Measurement lag should shift both, so the modality comparison survives either way. If adaptive stopping on total RNA matches fixed-$k$ nascent RNA, that is itself the answer to whether labeling is worth it.
Decisive Checks
The first check is whether the intronic proxy tracks labeled nascent RNA in GSE253836 for the fibrosis-relevant programs. If it does not, stage one on GSE98622 is reported only for programs where it does.
The second check is whether per-program measurement lag in GSE98622 correlates with transcript turnover estimated from Dyna-vivo-seq. A positive correlation would support the mechanism behind any $\Delta$.
The third and main check, in the paired cohort, is whether $\Delta$ is positive with an interval excluding zero at both reliability targets, against the total-RNA classifier and clinical chemistry.
The fourth check is whether false fibrosis calls at small $k$ concentrate among prior-driven predictions, as hallucinations concentrate in source-starved intervals in translation. If they do, the diagnostic becomes a practical filter: an early call that barely depends on the sample should be withheld rather than acted on.
References
- Dyna-vivo-seq unveils cellular RNA dynamics during acute kidney injury via in vivo metabolic RNA labeling-based scRNA-seq
- STACL: Simultaneous Translation with Implicit Anticipation and Controllable Latency using Prefix-to-Prefix Framework
- Incremental Decoding and Training Methods for Simultaneous Translation in Neural Machine Translation
- On the Hallucination in Simultaneous Machine Translation
- Molecular characterization of the transition from acute to chronic kidney injury following ischemia/reperfusion
- Learning Adaptive Segmentation Policy for Simultaneous Translation
- SASST: Leveraging Syntax-Aware Chunking and LLMs for Simultaneous Speech Translation
- ISG15 accelerates acute kidney injury and the subsequent AKI-to-CKD transition by promoting TGFβR1 ISGylation
- Butyrolactone I blocks the transition of acute kidney injury to chronic kidney disease in mice by targeting JAK1
- Going Beyond Your Expectations in Latency Metrics for Simultaneous Speech Translation