An open-source diagnostic toolkit for determining when a translated representation (h(X)) provides downstream predictive advantage over the original deployable representation (X), as a function of label budget.
TRACE implements the Advantage Representation Curve (ARC) introduced in the accompanying manuscript.
TRACE is the official companion repository for the paper:
Molecular Translators as a Computational Primitive for Biomarker Discovery: Learnability Gains Under Conserved Information Ceilings
Payam Saisan, Sandip Pravin Patel
bioRxiv (link coming soon)
TRACE serves two related purposes:
- a simulation and analysis framework for studying the paper’s central computational question;
- a practical diagnostic toolkit for comparing direct learning on (X) against translated learning on (h(X)) across label budgets.
The central question is:
If a model trained on (h(X)) outperforms a model trained on (X), is that because translation truly helps downstream learning, or because it only provides a temporary advantage in the low-label regime?
TRACE addresses this by analyzing paired learning curves across label budgets, rather than relying on a single benchmark AUC.
Timing of TRACE's development is tied to an emerging innovation in molecular data analytics. Translated molecular intermediates like MISO and GigaTime are emerging as a potentially game changing computational primitive for biomarker modeling. TRACE is built for studying this setting: when a downstream target (Y) may be better predicted indirectly through a translated representation (h(X)) derived from (X) vs. directly from it's original deployable representation (X).
In this view, a biomarker pipeline is organized as
Many biomarker pipelines now follow the pattern
deployable representation X -> translator -> translated representation h(X)
Examples include:
| Deployable representation (X) | Translated representation (h(X)) |
|---|---|
| H&E slide embeddings | predicted gene expression |
| H&E embeddings | predicted proteomics |
| H&E embeddings | predicted immune signatures |
| morphology features | predicted pathway activity |
| pathology foundation-model tokens | cross-modal biological embeddings |
Researchers often observe
but that single comparison is not enough.
A higher AUC from (h(X)) does not by itself tell you whether:
- translation provides a genuine practical advantage;
- the gain is mainly a finite-sample learnability effect;
- direct prediction from (X) will catch up once more labels are available;
- or the apparent gain reflects confounding, shortcut structure, or evaluation artifacts.
TRACE is built to distinguish those cases.
TRACE treats predictive performance as a function of training size.
where (X) is the original deployable representation, (h(X)) is the translated representation, and (n) is the number of labeled training samples.
TRACE computes the Advantage Representation Curve (ARC)
ARC is the expected generalization performance gap, indexed by label budget, between models trained on (h(X)) and models trained on (X).
The key point is that the shape of ARC across (n) is often more informative than any single endpoint metric.
TRACE uses four canonical ARC geometries as a compact diagnostic vocabulary:
| ARC pattern | Interpretation | Suggested action |
|---|---|---|
| positive at small (n), decaying toward zero | translation mainly improves sample efficiency | collect more labels if feasible |
| positive across the studied range | translated representation retains practical value | improve or exploit the translated representation |
| near zero throughout | translation adds little downstream value | prefer direct learning on (X) |
| negative throughout, or sign-reversing unfavorably | translation is lossy or distorting | avoid translated representation for this task |
These are empirical reference regimes, not rigid bins. Real studies may lie between them.
TRACE turns learning curves into decision-oriented summaries.
The accompanying paper studies a specific distinction:
- deployment-time information ceiling: the Bayes-optimal predictive limit available from the deployable representation (X);
- finite-sample learnability: how easily a downstream learner can extract useful predictive structure from limited labeled data.
If the translator (\hat Z = h(X)) is deterministic after paired-data training, then it cannot create new deployment-time information. Accordingly,
and, under the paper’s framing,
So whenever prediction from (\hat Z) exceeds prediction from (X) in practice, that gain must be interpreted as a learnability effect, not as creation of new deployment-time signal.
TRACE is the executable framework for studying exactly that distinction.
TRACE contains the code used to generate the paper’s synthetic experiments, figures, and summary analyses.
These experiments isolate the mechanism proposed in the manuscript:
translator-derived representations can improve downstream prediction in label-limited settings by reorganizing existing information into a more learnable form, even when deterministic translation does not increase the underlying deployment-time ceiling.
TRACE can also be used for studies that already have:
X : deployable features
H : translated features = h(X)
Y : downstream labels
and want to answer:
Should I train on (X) or on (h(X))?
Clone the repository:
git clone https://github.com/psaisan/TRACE
cd TRACEIf the repository includes packaging metadata for editable installation, install locally with:
pip install -e .If not, launch notebooks or scripts from the repository root so that local imports resolve correctly.
Python already includes a standard-library module named trace. Because this repository also uses trace/ as a package name, import behavior can depend on how Python is launched and what is on the path.
The most reliable ways to work with this repository are:
- run notebooks from the repository root;
- run scripts from the repository root;
- treat the notebooks in
Notebooks/as the primary runnable examples.
If you encounter an import collision with the Python standard-library trace module, first check that you are launching from the project root and that the local package has been installed correctly if editable installation is available.
The recommended entry point is the notebook sequence in Notebooks/.
-
Notebooks/00_overview_and_reference_scenarios.ipynb
Overview of TRACE outputs and reference synthetic scenarios. -
Notebooks/01_sample_efficiency.ipynb
Positive low-label ARC that decays toward zero. -
Notebooks/02_persistent_advantage.ipynb
Sustained positive ARC across label budgets. -
Notebooks/03_no_advantage.ipynb
Near-null ARC. -
Notebooks/04_lossy_translation.ipynb
Negative ARC / failure regime. -
Notebooks/05_custom_scenario_playground.ipynb
User-defined custom scenarios.
For paper-style synthetic outputs and the clearest overview, start with notebook 00.
The Examples/ directory contains additional example notebooks and scripts, including lightweight smoke tests and paper-style sweep runs.
A common applied setting is:
X = H&E embeddings
H = translated features, such as predicted gene expression from an H&E->RNA translator
Y = downstream task label, such as mutation status or response class
TRACE then estimates:
- paired learning curves from (X) and (H);
- the Advantage Representation Curve (ARC);
- compact summaries of low-label, late-label, and crossover behavior;
- regime-oriented interpretation.
At a high level, the analysis path is:
- provide
X,H, andy; - specify a label-budget grid
n_grid; - fit and evaluate paired downstream models across repeated subsamples;
- inspect:
- learning curves,
- ARC,
- scalar summaries,
- regime-like behavior.
Because the exposed programmatic interface may evolve, the notebooks in Notebooks/ should be treated as the primary runnable examples for this repository.
A typical sample-efficiency pattern looks like this:
Learning curves
AUC
0.9 | H
| /
0.8 | /
| /
0.7 | /
| /
0.6 | / X
+----------------
label budget
ARC
ARC
0.15 |\
| \
0.10 | \
| \https://github.com/psaisan/TRACE/tree/main
0.05 | \
+-----\-----------
label budget
Interpretation:
translation helps when labels are scarce
direct learning on X catches up with more data
Decision:
if labels are scarce -> use translation
if labels are plentiful -> direct modeling on X may suffice
TRACE is organized around three coordinated outputs:
-
paired learning curves
(A_X(n)) and (A_H(n)), typically with uncertainty bands; -
ARC curve
(\mathrm{ARC}(n)=A_H(n)-A_X(n)); -
compact summaries / regime scores
low-label advantage, late-regime behavior, and approximate crossover scale.
These outputs support interpretations such as:
- translation helps mainly when labels are scarce;
- direct learning on (X) catches up with increasing label budget;
- translated learning remains advantageous across scale;
- translation is effectively neutral;
- translation is lossy or misleading.
If you use TRACE in your work, please cite the associated manuscript:
Saisan, P., & Patel, S. P. (2026).
H&E-to-Molecular Translators as a Computational Primitive for Biomarker Discovery: Learnability Gains Under Conserved Information Ceilings.
bioRxiv.
@article{saisan2026trace,
author = {Saisan, Payam and Patel, Sandip Pravin},
title = {H\&E-to-Molecular Translators as a Computational Primitive for Biomarker Discovery: Learnability Gains Under Conserved Information Ceilings},
journal = {bioRxiv},
year = {2026},
note = {Preprint},
url = {https://github.com/psaisan/TRACE}
}MIT License