Model Card for RizFie/Eleuthia-v1

Eleuthia is a binary protein variant classifier that predicts whether a mutated protein sequence is likely Pathogenic or Benign. It is fine-tuned from ESM-2 and intended for research support in variant prioritization workflows.

Model Details

Model Description

Eleuthia takes an amino acid sequence containing a variant and returns class probabilities for:

• Benign (label 0)
• Pathogenic (label 1)

The model uses a maximum sequence length of 1022 tokens (ESM-2 limit). For longer sequences, inference applies a centered sliding window (optionally around the mutation position).

• Developed by: RizFie
• Funded by [optional]: Final Year Project (academic project)
• Shared by [optional]: RizFie
• Model type: Protein sequence classification (binary)
• Language(s) (NLP): Amino acid/protein sequence tokens
• License: Not yet formally specified (metadata currently set to other)
• Finetuned from model [optional]: facebook/esm2_t30_150M_UR50D

Model Sources [optional]

• Repository: https://github.com/RizFie/Eleuthia
• Paper [optional]: N/A
• Demo [optional]: FastAPI inference endpoint in project repository

Uses

Direct Use

Use Eleuthia to score mutated protein sequences and obtain:

• Predicted class (Pathogenic or Benign)
• Pathogenic probability
• Benign probability
• Confidence score

Intended users include bioinformatics students, researchers, and developers building exploratory tools for variant analysis.

Downstream Use [optional]

• Variant triage/prioritization in research pipelines
• Integration into web/API systems for rapid sequence scoring
• Feature input for higher-level decision-support models

Out-of-Scope Use

• Clinical diagnosis or treatment decision-making
• Standalone evidence for pathogenicity classification
• Use without domain-expert review and external validation
• Use on non-protein sequences or heavily out-of-domain inputs

Bias, Risks, and Limitations

• Training data is derived from curated cancer-gene variant records and may not generalize to all genes or populations.
• Label quality depends on source annotations and filtering logic.
• Oversampling can affect probability calibration.
• Sequence windowing for long proteins may omit distal context.
• Binary output simplifies a complex biological reality.

Recommendations

• Treat outputs as research signals, not clinical truth.
• Validate on independent datasets before deployment.
• Report uncertainty and inspect raw probabilities, not only hard labels.
• Combine with orthogonal evidence (functional assays, literature, clinical databases).

How to Get Started with the Model

Transformers inference

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "RizFie/Eleuthia-v1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

sequence = "MEEPQSDPSVEPPLSQETFSDLWKLLPENN..."  # mutated protein sequence
inputs = tokenizer(sequence, return_tensors="pt", truncation=True, max_length=1022)

with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=-1).squeeze().tolist()

print({"benign": probs[0], "pathogenic": probs[1]})

API-style decision rule used in deployment

threshold = 0.5
label = "Pathogenic" if probs[1] >= threshold else "Benign"

Training Details

Training Data

Training data was built from cancer-associated genes using EBI Proteins variation endpoints with ClinVar-style significance filtering. Variants were filtered to keep unambiguous single amino-acid substitutions with binary labels:

• 1: Pathogenic-only annotations
• 0: Benign-only annotations

Genes included in data collection:

• TP53, BRCA1, BRCA2, EGFR, APC, PTEN, RB1, MLH1, MSH2, KRAS, PIK3CA, CDH1

A dataset CSV is included in the project as Genetic_Variants_Dataset.csv.

Training Procedure

Preprocessing [optional]

• Construct mutated sequence from reference sequence + substitution.
• Enforce ESM-2 max length (1022) using mutation-centered sliding window when needed.
• Train/test split with stratification (test_size=0.25, random_state=42).
• Oversample the benign class in the training split to balance classes.
• Tokenize with AutoTokenizer from the base ESM-2 checkpoint.

Training Hyperparameters

• Training regime: fp16 mixed precision
• Base model: facebook/esm2_t30_150M_UR50D
• Optimizer/scheduler: default Trainer optimizer + cosine LR scheduler
• Learning rate: 1e-5
• Batch size: 8 (train), 8 (eval)
• Epochs: 15
• Weight decay: 0.01
• Early stopping: patience 3
• Loss: Focal Loss (alpha=0.25, gamma=2.0)
• Model selection metric: f1

Speeds, Sizes, Times [optional]

Please add your measured training wall-clock time, throughput, and final checkpoint size.

Evaluation

Testing Data, Factors & Metrics

Testing Data

Held-out stratified test split from the collected variant dataset (25%), with 2864 total samples.

Factors

Current evaluation is aggregate binary classification. Subgroup analyses (by gene, protein length, mutation type, domain) are recommended and currently not fully reported.

Metrics

• F1 score
• Accuracy
• Precision
• Recall
• ROC-AUC

Results

Final evaluation on the held-out test set:

• Overall Accuracy: 0.7105 (71.05%)
• ROC-AUC: 0.7731
• Macro F1: 0.68
• Weighted F1: 0.72

Per-class performance:

Class	Precision	Recall	F1-score	Support
Benign (0)	0.49	0.69	0.57	806
Pathogenic (1)	0.85	0.72	0.78	2058

Aggregate report:

• Macro avg: precision 0.67, recall 0.70, F1 0.68
• Weighted avg: precision 0.75, recall 0.71, F1 0.72

Summary

Eleuthia demonstrates feasible discrimination between benign and pathogenic variants on a cancer-focused held-out split, with stronger precision on pathogenic calls and moderate overall calibration. External validation on independent datasets is still required for broader use.

Model Examination [optional]

Evaluation visualizations from the held-out test split:

Confusion Matrix

ROC Curve

Additional recommended analyses for future versions:

• Per-gene performance breakdown
• Calibration plots
• Precision-recall curve (especially for class imbalance)

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: NVIDIA GPU (exact model to be added)
• Hours used: [More Information Needed]
• Cloud Provider: Google Colab (if applicable)
• Compute Region: [More Information Needed]
• Carbon Emitted: [More Information Needed]

Technical Specifications [optional]

Model Architecture and Objective

• Transformer encoder model from ESM-2 family
• Sequence classification head with 2 output logits
• Objective: binary classification of mutated protein sequences (Benign vs Pathogenic)

Compute Infrastructure

Training/inference were run in a Python environment using PyTorch + Transformers. GPU acceleration was enabled when available.

Hardware

• GPU: CUDA-capable GPU (exact model not recorded in this card)
• CPU/RAM: [More Information Needed]

Software

Core stack:

• torch
• transformers
• datasets
• evaluate
• scikit-learn
• pandas
• fastapi (deployment API)

Citation [optional]

If you publish a report/thesis/paper, add it here.

BibTeX:

@misc{eleuthia2026,
  title={Eleuthia: Cancer-Associated Protein Variant Classification Using ESM-2},
  author={RizFie},
  year={2026},
  howpublished={Hugging Face model repository},
  note={\url{https://huggingface.co/RizFie/Eleuthia-v1}}
}

APA:

RizFie. (2026). Eleuthia: Cancer-Associated Protein Variant Classification Using ESM-2 [Model]. Hugging Face. https://huggingface.co/RizFie/Eleuthia-v1

Glossary [optional]

• Pathogenic: Variant likely associated with disease phenotype.
• Benign: Variant likely not associated with disease phenotype.
• Focal Loss: Loss function that focuses training on harder examples.

More Information [optional]

For practical use, pair model output with domain constraints, curated databases, and expert interpretation.

Model Card Authors [optional]

RizFie

Model Card Contact

• Name: Mohamad Hariz Rafie
• Email: harizrafie86@gmail.com
• GitHub: https://github.com/RizFie