A three-tower PyTorch fusion network that classifies human genes into inherited retinal dystrophy (IRD) phenotypic modules using evolutionary conservation, protein language model embeddings, and protein-protein interaction network topology - then applies the trained model genome-wide to surface novel candidate genes.
Part of a broader research portfolio: this repository is the flagship ML/DL extension of the Evolutionary Genomics & Multi-Omics Portfolio, building directly on the NPP and HPO-based phenotypic clustering methodology developed there.
A three-tower fusion model recovers known Inherited Retinal Diseases (IRD) gene-phenotype relationships at 0.40 macro-F1 (up to 0.77 F1 on individual phenotypic modules) - roughly 6.8× the random baseline on a 17-class problem trained on only 434 genes - by integrating evolutionary conservation, protein sequence, and protein-interaction (PPI) network topology.
A PPI-only single-tower variant scores higher on average (0.44) - expected, since PPI representation is biased toward well-studied genes, and the 434 training genes are themselves already well-characterized IRD genes.
For an interactive walkthrough of the research - click to launch:
Narrative walkthrough of the research - NPP methodology, project case studies, and contact.
pytorch-multimodal-gene-classifier/
├── scripts/
│ ├── config.py # hyperparameters and constants
│ ├── dataset.py # PyTorch Dataset for multi-modal NPZ loading
│ ├── model.py # ThreeTowerClassifier architecture
│ ├── train.py # 5-fold stratified cross-validation training
│ ├── evaluate.py # per-variant evaluation and confusion matrices
│ ├── inference.py # genome-wide inference and concordance analysis
│ ├── align_npp_esm2.py # data alignment: NPP × ESM2
│ ├── append_ppi_embeddings.py # appends PPI to aligned datasets
│ └── umap_visualization_demo.ipynb # UMAP of learned embedding space
├── HPC_run_public/ # sanitized SLURM/cluster pipeline, ready to adapt to any HPC environment
│ ├── esm2/ # ESM2 embedding generation
│ ├── ppi/ # STRING → node2vec embedding pipeline
│ └── README.md # variable reference, execution order, compute requirements
├── ppi/
│ ├── scripts/ # local PPI pipeline: STRING ID mapping, edge list construction
│ └── processed/ # intermediate mapping and edge list files
├── input/ # committed documentation/sample files; real data is gitignored
│ ├── gene_classification_SAMPLE.csv # synthetic demo file illustrating expected input format (real labeled dataset is confidential and not distributed)
│ └── DATA.md # data requirements and processing instructions
├── output/
│ ├── evaluation/ # confusion matrices and per-module metrics
│ ├── training/ # cross-validation metrics per variant
│ └── figures/ # UMAP visualization
├── docs/
│ ├── training_methodology.md # methodology and results: two-tower ceiling → three-tower breakthrough
│ ├── inference_methodology.md # genome-wide inference methodology and aggregate results
│ └── ppi_methodology.md # STRING v12.0 → node2vec pipeline and biological validation
├── requirements.txt
└── .gitignore
docs/ also contains per-run training logs and per-gene inference output locally - these stay private (gitignored), since they include candidate gene-level data pending biological validation and publication. The three files listed above are their public methodology and aggregate-results summaries.
The question: can a deep learning model, built from three complementary representations of a gene - evolutionary conservation across roughly 1,900 species, a pretrained protein language model's view of its sequence, and its position in the human protein-protein interaction network - recover known inherited retinal dystrophy (IRD) gene-phenotype relationships well enough to be trusted for genome-wide candidate discovery, in a clinical area where a substantial fraction of cases still lack a confirmed causal gene?
Built with: PyTorch · deep learning (DL) · multi-modal fusion architectures · ESM2 protein language model embeddings · node2vec graph embeddings · scikit-learn
The result: combining all three modalities, and testing every possible sub-combination as a controlled ablation, surfaced PPI network topology as the single strongest predictor of IRD module membership - a result that was not anticipated going in.
| Variant | Val Macro-F1 (5-fold CV, mean ± SD) |
|---|---|
ppi_only |
0.4379 ± 0.0502 — best overall |
fusion_3tower |
0.4012 ± 0.0200 |
npp_ppi |
0.3940 ± 0.0179 |
esm2_ppi |
0.3936 ± 0.0127 |
npp_esm2 (Phase 1 ceiling, no PPI) |
0.2921 ± 0.0361 |
esm2_only |
0.2488 ± 0.0241 |
npp_only |
0.2212 ± 0.0179 |
| Random baseline (1/17 classes) | ~0.059 |
The best model reaches roughly 7.5× the random baseline on a 17-class problem, trained on only 434 labeled genes. For context: a problem this small and this biologically noisy does not behave like image classification, where 95%+ accuracy is routine on clean, abundant data.
A macro-F1 of 0.44 here reflects genuine biological heterogeneity within disease modules and a training set two orders of magnitude smaller than a typical deep learning benchmark - it is an honest, usable signal for prioritizing candidates, not a solved classification problem.
Applied genome-wide to ~19,573 unlabeled genes, just over half receive matching predictions from at least two independently-trained model variants, and a smaller, more selective subset achieves unanimous agreement across all variants tested - full breakdown of the concordance methodology and reliability framework in docs/inference_methodology.md.
One module that failed completely throughout the first project phase (F1 = 0 in most runs without PPI) produced its first strong result the moment PPI features were introduced - a concrete illustration of what this modality contributes that the others structurally cannot.
434 labeled training genes projected into 2D via UMAP from the learned three-tower embedding space. High-reliability modules form visually distinct clusters, consistent with their validation performance.
NPP Tower: Linear(1905→64) → ReLU → Dropout(0.3) ────────────────┐
│
ESM2 Tower: Linear(1280→128) → ReLU → Dropout(0.3) ──────────────┼── Concatenation (256-dim)
│ │
PPI Tower: Linear(129→64) → ReLU → Dropout(0.3) ─────────────────┘ ▼
Fusion Head
Linear(256→64) → ReLU → Dropout(0.3) → Linear(64→17)
│
▼
Softmax over 17 IRD modules
The network integrates three independent modalities:
- NPP: 1905-dimensional evolutionary conservation profiles across ~1,900 species.
- ESM2-650M: 1280-dimensional protein language model embeddings (frozen).
- PPI: 128-dimensional node2vec embeddings generated from STRING v12.0 networks, plus a 1-bit coverage flag indicating genes that received mean-vector imputation rather than genuine network signal.
Total trainable parameters (fusion_3tower): 311,825
The architecture supports dynamic ablation studies using boolean flags (use_npp, use_esm2, use_ppi), enabling training and evaluation of all 7 possible tower combinations from a single codebase.
git clone https://github.com/<your-username>/pytorch-multimodal-gene-classifier
cd pytorch-multimodal-gene-classifier
pip install -r requirements.txtOne note: pecanpy (PPI pipeline only) downgrades numpy to 1.26.4 due to numba constraints - install it in a separate environment if needed.
A note on reproducibility: This repository includes the full training and inference pipeline, code, and all aggregate results (cross-validation metrics, confusion matrices, the UMAP visualization). It does not include the labeled training dataset (confidential, pending publication) or trained model checkpoints (~25 MB, derived directly from that data). As a result,
train.py,evaluate.py, andinference.pycannot be run end-to-end from a fresh clone. The code is provided for transparency, architecture review, and adaptation to your own labeled dataset of the same format (seeinput/DATA.md).
Refer to input/DATA.md for full instructions.
python scripts/align_npp_esm2.py && python scripts/append_ppi_embeddings.pyNote: edit VARIANTS in scripts/config.py to select which tower combinations to train.
python scripts/train.pyOutputs confusion matrices and per-module metrics to output/evaluation/.
python scripts/evaluate.pyOutputs per-variant predictions and concordance analysis to output/inference/.
python scripts/inference.py434 labeled genes, 17 IRD modules (HPO-based clustering, thesis Stage 1), 5-fold stratified cross-validation, macro-F1 as primary metric (chosen specifically for class imbalance, since module sizes range from 10 to 63 genes). AdamW optimizer, early stopping on validation macro-F1, class-frequency-weighted cross-entropy loss.
Ablation strategy: one variable changed per run throughout nine controlled Phase 1 experiments (architecture size, regularization, augmentation, learning rate, class weighting), followed by a full 7-variant ablation across all combinations of the three modalities once PPI was introduced. Full run-by-run record in docs/training_methodology.md.
- Fourth tower: tissue expression data (Human Protein Atlas, chosen over GTEx for its retinal tissue coverage), planned as an independent post-hoc validation layer rather than additional training input
- Investigating why the PPI-only model outperforms the full three-tower fusion - architecture (fusion mechanism, capacity) and data (study bias in curated interaction databases) are both live hypotheses, discussed in
docs/training_methodology.md - Biological sanity-checking of top genome-wide candidates against independent expression and locus evidence
The Tabach Lab, Hebrew University of Jerusalem. STRING database v12.0. ESM2 (Meta AI). The HUJI Moriah HPC cluster.
This model's training labels derive from the HPO-based phenotypic module clustering (Stage 1) developed in the Evolutionary Genomics & Multi-Omics Portfolio.
- Lab: Prof. Yuval Tabach Lab, Faculty of Medicine
- Institution: Hebrew University of Jerusalem
- Role: M.Sc. Candidate in Genomics & Bioinformatics
This is an active MSc research project. Results presented here are hypothesis-generating - statistically supported candidates for follow-up, not clinically validated findings. Gene-level outputs from genome-wide inference are withheld pending biological validation and publication.
Questions, feedback, and collaboration inquiries are welcome via the contact details on the visual portfolio site.
Built as part of ongoing MSc research at the Hebrew University of Jerusalem, Tabach Lab - ML/DL applied to inherited retinal disease gene discovery.