Many clinical decisions rest on a handful of gene-expression measurements from quantitative real-time PCR (qPCR). qPCR is a lab assay that tracks fluorescence over roughly 40 heating cycles as a DNA target is copied, producing an S-shaped curve for each gene. Standard practice keeps one number from that curve, the cycle threshold $C_t$: the cycle at which fluorescence rises above background. Each one-cycle shift means about a twofold change in starting transcript. That number is then normalized against "housekeeping" reference genes that are assumed to stay constant. When these numbers feed a prognostic model, two quiet failures stack up. Reference genes often do drift across conditions, as shown in Selection and Validation of Reference Genes For qRT-PCR Analysis of Rhopalosiphum padi and Selection of Stable Reference Genes for QRT-PCR in Tree Peony ‘Doulv’. The outcome labels also come from studies whose follow-up ended early, so most patients are "right-censored": still event-free when observation stopped. The 3–6-month window in the atrial fibrillation recurrence study missed every relapse after month 6.
This note proposes a system that borrows the "transcribe first" trick from speech language models. The model reads raw amplification curves and first writes a short, checkable assay transcript: per-gene $C_t$, amplification efficiency, curve-quality flags, a verdict on whether the reference genes are stable, and the normalized fold changes. Only then does it compare patients to rank their risk, using rewards that respect censoring. Because every line of the transcript can be checked against a deterministic curve fitter, we can test directly whether the reasoning drives the risk ranking or merely decorates it. Because training labels can be deliberately truncated, we can measure how much long-horizon ranking ability survives when follow-up is short.
Core Claim
The claim is narrow and falsifiable, and it has two parts. First, a model that generates an explicit, verifiable assay transcript before comparing patients will rank long-horizon risk more robustly than models fed only scalar $\Delta\Delta C_t$ values. The advantage should appear specifically when assays contain realistic artifacts, such as reference-gene drift, inhibited wells, or degraded efficiency. Second, the transcript makes faithfulness measurable. Perturbing a transcript line should move the ranking by an amount consistent with that line's biological meaning. A model that ignores its own transcript is caught by this test, which addresses the risk raised in Right Diagnoses, Decorative Reasoning.
We do not claim that raw curve encoding beats curve fitting on clean data. We expect it will not. A decisive arm of the study replaces the model-written transcript with a tool-written one. If that arm matches, the simpler pipeline becomes the recommended system, and the contribution reduces to the benchmark, the truncation protocol and the audit.
How This Differs From Existing Work
Two prior methods supply most of the machinery.
From speech, WHISMA pairs a frozen audio encoder with Llama-3-8B through small trainable adapters. It shows that making the model write a verbatim transcript before answering acts as an anchor: +26.6% relative F1 on SLURP slot filling and +33.0% relative accuracy over Qwen-Audio on unseen SLU-GLUE tasks. Can Speech LLMs Think while Listening? shows that text reasoning is roughly twice as token-efficient as reasoning over speech tokens and gives a 2.4x average accuracy gain. We reuse WHISMA's encoder-adapter-LoRA layout, with a small 1D curve encoder in place of Whisper. The key difference is that the qPCR "transcript" has a ground truth computable by a deterministic fitter, whereas speech transcripts require human labels.
From survival analysis, CACSurv trains a Qwen2.5-7B model to compare two or three patient text reports. It uses Group Relative Policy Optimization (GRPO), a reinforcement-learning method that scores groups of sampled answers against each other. Its rewards come only from patient pairs whose ordering is known despite censoring. It reaches an average concordance index (C-index) of 0.722 on TCGA-SurvReport, which has more than 60% censoring. We reuse its comparative training and its Monte Carlo reference aggregation at inference, which compares each test patient against random reference subsets. CACSurv only reads pre-written text reports and never sees raw assay signals. It also never tests what happens when follow-up is artificially shortened.
SurvAgent motivates the clinical value of step-by-step reasoning over pathology and genomics, reaching C-index 0.713. It treats genomics as static bulk categories and has no checkable intermediate.
Three methods constrain the design and serve as non-language baselines. Temporally-Consistent Survival Analysis shows that recursive targets beat landmarking on short, heavily censored sequences. Temporal Quilting shows that model reliability shifts across horizons as the at-risk pool thins. PrimerCast cautions that handcrafted kinetic features are surprisingly competitive: AUROC 0.930 for predicting amplification success. That is precisely why the tool-transcript arm exists.
Research Program
Building a benchmark. No dataset in our material links raw qPCR curves to survival outcomes, so the first deliverable is a semi-synthetic benchmark built on the six TCGA-SurvReport cohorts (BLCA, BRCA, COADREAD, HNSC, LUAD, STAD; over 3,300 patients). For each cohort we choose a low-plex panel of 8–12 prognostic genes plus 3 candidate reference genes from TCGA bulk expression. We then render each patient's expression as amplification curves using a standard kinetic model with four noise sources: baseline drift, plateau saturation, well-level inhibition, and efficiency degradation. Efficiency distributions are matched to the spread observed in the PrimerCast dataset (50,760 primer–target combinations).
We inject reference-gene drift that is correlated with tumor biology. This produces exactly the case where naive $\Delta\Delta C_t$ normalization misleads. The simulator, its seeds and the artifact rates are released, so the artifact level is a controllable experimental axis rather than a hidden confound. Transferring to real clinical qPCR cohorts is a separate, later validation that this note does not claim to have.
The model. A small 1D convolutional encoder maps each well's fluorescence curve to a few tokens. A trainable adapter projects these tokens into a Qwen2.5-7B backbone fine-tuned with LoRA (low-rank adapters, a cheap way to fine-tune a large model). The output has two parts. The first is the assay transcript in a fixed schema. The second is a comparative rationale and ranking over a mini-cohort of two or three patients, as in CACSurv.
Training. We first train the transcript in a supervised way against the deterministic fitter on clean wells. We then run GRPO with the reward
$$R = C_{\text{pairs}} + \lambda \, A_{\text{transcript}},$$
The two terms are:
- $C_{\text{pairs}}$ is the fraction of censoring-comparable patient pairs in the mini-cohort that the model orders correctly.
- $A_{\text{transcript}}$ is the fraction of transcript fields matching the fitter within tolerance, computed only on wells where the fitter is reliable.
The weight $\lambda$ is tuned on validation data. Wells with injected artifacts receive no transcript reward. On these wells the model must learn from the ranking reward alone whether to trust or discount them. This is where we expect it might beat a fixed pipeline.
Truncated-follow-up protocol. This is how limited follow-up enters the study. Training labels are administratively censored at a horizon $\tau$ of 6, 12 or 24 months: any event after $\tau$ becomes a censoring at $\tau$. Test patients keep their full follow-up. We report:
- C-index restricted to test pairs whose ordering is decided by events after $\tau$, which isolates "stamina" beyond the training window;
- horizon-wise Brier score, a calibration measure;
- the gap between truncated and untruncated training.
Baselines
The naive reference is a Cox proportional-hazards model on scalar $\Delta\Delta C_t$ computed with the nominal reference genes, the way a lab would do it today. It sets the floor and should be hurt most by reference drift.
The strongest classical competitor is a tool-feature tabular model. A deterministic fitter extracts the full curve parameters ($C_t$, efficiency, plateau height, takeoff slope, QC flags) and a geNorm-style reference-stability score. These features feed:
- Random Survival Forests;
- Survival Quilts (Temporal Quilting), which keeps performance stable across horizons;
- SurvTRACE (SurvTRACE), a transformer survival model with reported relative gains of 1.7–2.1% on METABRIC.
This family receives the same features the transcript contains. It therefore tests whether language reasoning adds anything beyond better feature extraction.
The strongest language-model competitor is CACSurv run on text reports generated from the same fitted features. Its published average C-index on TCGA-SurvReport is 0.722, against 0.657 for the best non-LLM baseline (both from CACSurv). Those numbers come from clinical text reports, not qPCR. On our benchmark there is no reported result found on Lacuna for qPCR-curve survival ranking, so we rerun CACSurv ourselves.
The key ablation is the tool-transcript arm. It is our exact model, but the transcript is written by the fitter and inserted rather than generated from raw curves. A second ablation removes the transcript entirely: the model goes straight from curve tokens to the comparative rationale. This isolates the anchoring effect that WHISMA reported for speech.
Fair comparison. All arms share the same patient splits, the same truncation horizons and the same hyperparameter search budget (equal numbers of configurations on the validation fold). All LLM arms use the same 7B backbone and the same GRPO step budget.
What counts as beating a baseline. We beat the tabular family only if the post-$\tau$ C-index is higher with a bootstrap confidence interval excluding zero, at the two highest artifact rates. We beat the tool-transcript arm only if the gain appears on artifact-bearing wells and not merely on clean ones.
Why not just…?
Why not just fit curves with standard software and run a survival forest on the numbers? This is the most serious threat. PrimerCast's success with handcrafted kinetic features suggests extracted curve parameters carry most of the signal. CACSurv's 6.5-point lead over non-LLM models comes from text reports under heavy censoring, not from assay data, so it does not settle the question here. The tool-feature tabular baselines answer it directly. If they match our model at every artifact level, we conclude that the language model adds nothing for low-plex qPCR, and we report exactly that.
Why not just write the fitted values into a prompt and use CACSurv as-is? This is the tool-transcript arm, and on clean curves we expect it to tie. The open question is whether a fixed fitter's errors on inhibited wells or drifting reference genes pass silently into the ranking. A model that sees the raw curves and the other patients might discount those errors. The speech analogy only partly supports this. Can Speech LLMs Think while Listening? found that cascaded pipelines lose fine acoustic detail, but qPCR curves are far simpler than audio. The comparison stratified by artifact rate settles it. If the gap stays near zero even under heavy artifacts, the simpler cascade wins.
Why not just use temporal-consistency regression or landmarking to handle short follow-up? Temporally-Consistent Survival Analysis matches landmarking with four times fewer samples, but it needs repeated visits that form a Markov sequence. Our benchmark is single-snapshot, so TCSR's recursive targets do not apply directly. Landmarking reduces to Cox at one landmark, which is already a baseline. The cheaper way to handle short follow-up is not a new model but a check that any model's ranking survives truncation. That check is our truncation protocol, applied to every arm.
Why trust the rationale at all? Because it is checkable, unlike the free-text clinical reasoning audited in Right Diagnoses, Decorative Reasoning. In the audit below, a model whose ranking does not move when its transcript is corrupted fails, whatever its C-index.
Proposed Experiments and Decisive Checks
Experiment 1: factorial comparison. We cross four factors and report post-$\tau$ C-index, full C-index and integrated Brier score:
- arm: naive Cox, tabular family, CACSurv-on-text, tool-transcript, no-transcript, full model;
- artifact rate: 0%, 10%, 25%;
- truncation horizon $\tau$: 6, 12, 24 months;
- cohort: the six cohorts.
The decisive pattern for the core claim is specific. The full model must beat the tool-transcript arm and the tabular family at high artifact rates and short $\tau$, while tying them on clean data. Any other pattern falsifies the claim.
Experiment 2: faithfulness audit. On held-out patients we apply three edits to the generated transcript, then let the model continue reasoning from the edited version:
- shift a prognostic gene's $C_t$ by a controlled amount;
- flip the reference-stability verdict;
- mark a well as failed.
A faithful model's risk score should move in the biologically expected direction, with magnitude roughly increasing with the size of the $C_t$ shift. It should be insensitive to edits on genes it calls irrelevant. We report the rank correlation between edit size and score change, and the fraction of edits producing no change. The same audit on CACSurv-on-text gives a reference point.
Experiment 3: transcript accuracy as a predictor of ranking errors. We test whether errors in the model's transcript explain its ranking mistakes. If they do, transcript accuracy, which is cheap to verify, becomes a usable confidence signal at deployment. This mirrors WHISMA's observation that upstream transcription errors propagate downstream.
Cost. We report wall-clock inference cost against the tabular baselines, since comparative inference with 40 reference subsets is far slower than a forest. A gain under about one C-index point would not justify that cost.
References
- Selection and Validation of Reference Genes For qRT-PCR Analysis of Rhopalosiphum padi (Hemiptera: Aphididae)
- Selection of Stable Reference Genes for QRT-PCR in Tree Peony ‘Doulv’ and Functional Analysis of PsCUC3
- A predictive model using left atrial function and B‐type natriuretic peptide level in predicting the recurrence of early persistent atrial fibrillation after radiofrequency ablation
- Right Diagnoses, Decorative Reasoning:A Perturbation Audit of Medical Chain-of-Thought
- WHISMA: A Speech-LLM to Perform Zero-Shot Spoken Language Understanding
- Can Speech LLMs Think while Listening?
- CACSurv: Concordance-Aligned Comparative Learning with Large Language Models for Cancer Survival Prediction
- SurvAgent: Hierarchical CoT-Enhanced Case Banking and Dichotomy-Based Multi-Agent System for Multimodal Survival Prediction
- Temporally-Consistent Survival Analysis
- Temporal Quilting for Survival Analysis
- PrimerCast: Predictive Modeling of PCR Amplification with an AI-Ready Experimental Dataset
- SurvTRACE: transformers for survival analysis with competing events