TransStop

Premature termination codons (PTCs) — nonsense mutations that insert a stop signal in the middle of a coding sequence — account for roughly 12% of rare human genetic diseases and are found in about 10% of tumor suppressor genes. One therapeutic strategy is translational readthrough: small molecules that coax the ribosome past the PTC so a full-length protein is still made. The catch is that readthrough efficacy depends heavily on which stop codon is present (UAA, UAG, or UGA), on the surrounding sequence, and on the specific compound used, so choosing the right drug for a given patient's mutation is a prediction problem.

TransStop, developed by the Complex Systems and Translational Bioinformatics group at the ICube laboratory of the University of Strasbourg and posted as a preprint in September 2025, addresses that problem with a pan-drug regression model. It is fine-tuned from the 500M-parameter multi-species checkpoint of Nucleotide Transformer v2, which supplies a pretrained representation of DNA sequence context, and adds a learnable drug embedding coupled to the sequence through cross-attention. A single model therefore covers eight readthrough compounds at once rather than requiring one model per drug.

The prior state of the art was an interpretable logistic regression from Toledano et al. (2024), fit on the same massively parallel reporter data. TransStop tests whether a transformer can capture the non-linear sequence-drug interactions a linear model cannot, and convert that gain into therapeutic recommendations at genome scale.

#Key Features

• Pan-drug in one model: A learnable embedding for each of eight compounds — G418, gentamicin, clitocine, FUr, DAP, SRI, SJ6986, and CC90009 — plus an untreated control, lets one checkpoint predict efficacy for any of them.
• Cross-attention drug conditioning: The drug embedding acts as the query and attends over the nucleotide token embeddings as keys and values, so the model reweights the sequence context differently for each compound rather than concatenating drug and sequence features.
• Compact sequence window: An ablation over context lengths from ±0 to ±72 nucleotides showed that ±6 nucleotides around the stop codon (a 15-nt input) captures nearly all the signal, making genome-scale inference cheap.
• Emergent biological structure: UMAP projections of the latent space separate cleanly into UAA, UAG, and UGA clusters without supervision, and correlating predicted profiles across drugs recovers the known eRF1/eRF3-inhibitor grouping (r > 0.89).
• Genome-wide prediction resource: 32.7 million predictions covering every possible human PTC for all eight drugs are deposited on Zenodo.

#Technical Details

The training data comes from Toledano et al.'s reporter assay of pathogenic ClinVar PTCs, filtered with the same criteria as the original study to roughly 5,400 PTCs per drug, with readthrough log(x+1)-transformed and the data split 80/10/10 stratified by drug. The regression head is a cross-attention layer followed by two linear layers with ReLU and dropout; an Optuna search over 30 trials selected a hidden size of 512, a drug embedding size of 32, and dropout of 0.17. On the held-out test set TransStop reaches a global R2 of 0.94, against 0.83 for the logistic regression baseline, and improves on it for every drug tested. Per-drug scores range from 0.92 for SRI and DAP down to 0.77 for gentamicin, with the largest gains on the eRF1/eRF3 inhibitors SJ6986 (0.91 vs. 0.71) and CC90009 (0.79 vs. 0.55). Weights ship on Hugging Face under Apache 2.0 with a model card; the GitHub repository holds the reproduction pipeline but carries no license file.

#Applications

The genome-wide predictions give clinicians and trial designers a lookup table from mutation to best-predicted compound. Across all human PTCs, clitocine is predicted best for 96.4% of UAA contexts, DAP for 87.4% of UGA, and SJ6986 for 84.1% of UAG, a specialization pattern that argues for stratifying nonsense-suppression trials by codon. The model also supports in silico saturation mutagenesis and epistasis analysis — applied to a DAP/UGA context it recovers the +4 cytosine as a dominant readthrough enhancer — and can generate a per-gene "therapeutic matrix", demonstrated for the CFTR nonsense mutations G542X, R553X, R1162X, and W1282X.

#Impact

TransStop is a worked example of adapting a genomic foundation model to a narrow, clinically actionable regression task: compared with the linear baseline it revises the recommended drug with a substantial predicted gain in 41,664 genomic positions, 92.6% of which involve UGA codons. The limitations are acknowledged by the authors. Training data comes from a single reporter system in one cell line (HEK293T), evaluation is on a held-out split of that same assay rather than an independent cohort, and the drug axis is a closed set of eight identities rather than a chemical representation, so the model cannot extrapolate to a new compound without retraining. The preprint has not yet been peer reviewed, and its therapeutic predictions are hypotheses awaiting experimental confirmation.