longevity-esm2-v6

Longevity Protein Classifier

Fine-tuned ESM-2 150M for binary classification of protein sequences as longevity-associated or not. Trained on multi-species GenAge data with LoRA adapters. Built to connect protein language models to longevity biology.

Performance

Benchmark Results

Novel Predictions Not in GenAge

Proteins scoring above 0.50 that are not in the GenAge human database. These are model predictions only — not experimentally validated.

TFEB is the strongest novel prediction. It is mechanistically connected to mTOR (already in GenAge), independently predicted at 0.502 by a model trained with no pathway information.

Recommended Thresholds

Known Limitations

1. Protein length Sequences longer than 512 amino acids are truncated from the C-terminus. This causes failures on long proteins where the functional domain sits in the C-terminal half. MTOR (2,549 aa) and TERT (1,132 aa) both fail for this reason. Do not use this model on proteins above 800 amino acids without validating first.

2. Family-specific blind spots The model learned sirtuin and tumour suppressor sequence features well but has insufficient training examples to generalise to forkhead transcription factors. FOXO3 (402 aa, fits within 512 window) scores 0.000 despite being a canonical longevity gene. This is a training data coverage problem, not a truncation problem.

3. Direction of effect not captured The model cannot distinguish pro-longevity proteins (overexpression extends lifespan) from anti-aging-disease proteins (loss of function accelerates aging). A high score means associated with longevity biology, not activating this protein extends lifespan.

4. Not for clinical use Research screening tool only. Do not use for clinical, diagnostic, or therapeutic decisions.

How to Use

from transformers import AutoTokenizer, EsmForSequenceClassification
from peft import PeftModel
import torch

base = EsmForSequenceClassification.from_pretrained(
    "facebook/esm2_t30_150M_UR50D",
    num_labels=2,
    ignore_mismatched_sizes=True
)
model = PeftModel.from_pretrained(base, "mawe/longevity-esm2-v6")
tokenizer = AutoTokenizer.from_pretrained("mawe/longevity-esm2-v6")

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)
model.eval()

def score_sequence(sequence, threshold=0.06):
    inputs = tokenizer(
        sequence,
        max_length=512,
        padding="max_length",
        truncation=True,
        return_tensors="pt"
    )
    with torch.no_grad():
        outputs = model(
            input_ids=inputs["input_ids"].to(device),
            attention_mask=inputs["attention_mask"].to(device)
        )
        prob = torch.softmax(outputs.logits, dim=1)[:, 1].item()
    return {
        "probability": round(prob, 4),
        "prediction": "Longevity" if prob >= threshold else "Non-longevity",
        "warning": "Sequence truncated to 512aa" if len(sequence) > 512 else None
    }

Training Details

Positive set: GenAge database

• Human GenAge: 306 genes (all entries)
• C. elegans Pro-Longevity: 283 genes
• D. melanogaster Pro-Longevity: 125 genes
• M. musculus Pro-Longevity: 85 genes
• Total positives: 574

Negative set: Swiss-Prot reviewed proteins

• NEG_RATIO: 10 negatives per positive
• Species weights: human 2.0x, mouse 1.5x, worm and fly 1.0x
• Necessary-for-fitness genes excluded from universe
• Anti-Longevity genes excluded from positives

Architecture: ESM-2 150M + LoRA r=16, alpha=32, dropout=0.15

Loss: Focal loss gamma=1.0, label smoothing=0.1, contrastive margin=0.30

Optimiser: AdamW lr=2e-4, cosine schedule, 10% warmup

Hardware: NVIDIA T4 16GB on Kaggle

Experiment History

Contact

Feedback and collaboration welcome.