HybriTE

HybriTE is a graph neural network for mRNA translation efficiency (TE) prediction. It represents each transcript with region-aware graph nodes that combine sequence-derived features, RNA secondary-structure priors from RNAplfold, and biochemical annotations. This repository contains the HybriTE code used for data preparation, graph construction, HybriTE training, LightGBM baseline training, cross-species evaluation, interpretability, and figure generation.

Requirements

• Python 3.10
• ViennaRNA (RNAplfold on PATH)

Setup

conda env create -f environment.yaml
conda activate hybrite
pip install -e .

Dataset

Download the prepared dataset files from this Google Drive folder:

https://drive.google.com/drive/folders/1h4gT797xGT1nZgT0iuO-dTuFukhOvsNT?usp=sharing

Place the downloaded files at the paths expected by the configs:

• data/raw/human/translation_efficiency_with_biochemistry.csv
• data/raw/mouse/translation_efficiency_raw.csv
• data/derived/mouse/translation_efficiency_with_biochemistry.csv (if using the transferred mouse table directly)

If you want to rebuild the mouse transferred table yourself, also provide:

• data/raw/orthology/mart_export.txt

Main folders

• hybrite/ — core model and utilities
• scripts/ — runnable paper scripts
• configs/ — paper configs only

Scripts

Data and graph preparation

• scripts/transfer_orthology.py — transfer human biochemical features to mouse through one-to-one orthology
• scripts/precompute_structure.py — run RNAplfold and save structure caches
• scripts/build_graphs.py — build HybriTE transcript graphs from input tables
• scripts/prepare.py — create CV folds and fold-specific biochemical feature manifests

Training

• scripts/train.py — train the HybriTE graph model
• scripts/train_baseline.py — train the LightGBM biochemical-only baseline

Cross-species evaluation

• scripts/cross_species.py — evaluate a HybriTE checkpoint across species
• scripts/cross_species_baseline.py — evaluate a LightGBM baseline across species

Analysis and figures

• scripts/compare_runs.py — compare two matched run directories fold by fold
• scripts/generate_interpretability_artifacts.py — generate human interpretability outputs
• scripts/generate_all_figures.py — generate all main paper figures

Figure scripts

• scripts/plots/plot_figure_performance_analysis.py — main benchmark, ablation, and transfer figure
• scripts/plots/plot_figure_interpretability_analysis.py — interpretability figure
• scripts/plots/plot_figure_per_target_performance.py — per-source performance figure
• scripts/plots/plot_figure_target_correlation.py — prediction-similarity figure

Main configs

• configs/main/human.yaml — full human HybriTE model
• configs/main/human_nobio.yaml — human -Bio ablation
• configs/main/human_nostruct.yaml — human -Struct ablation
• configs/main/mouse.yaml — full mouse HybriTE model
• configs/main/mouse_nobio.yaml — mouse -Bio ablation
• configs/main/mouse_nostruct.yaml — mouse -Struct ablation
• configs/baselines/human_lightgbm.yaml — human LightGBM baseline
• configs/baselines/mouse_lightgbm.yaml — mouse LightGBM baseline

Sensitivity configs used in the paper

• configs/controls/human_bins_coarse.yaml
• configs/controls/human_bins_fine.yaml
• configs/controls/human_threshold_1e_2.yaml
• configs/controls/human_threshold_1e_1.yaml
• configs/controls/human_hp_layers_2_hidden_64.yaml
• configs/controls/human_hp_layers_2_hidden_128.yaml
• configs/controls/human_hp_layers_2_hidden_256.yaml
• configs/controls/human_hp_layers_3_hidden_64.yaml
• configs/controls/human_hp_layers_3_hidden_128.yaml
• configs/controls/human_hp_layers_3_hidden_256.yaml