Original 2018 concept: Use SVM to discriminate peptide sequences between short vs long-lived species, train LSTM/GAN to generate new peptides along the "multi-species aging vector", then manufacture and test oligopeptides.

The core biological insight remains excellent. Here is how recent advances (2020-2026) dramatically upgrade it.

Modernized Pipeline

Phase 1: Feature Extraction (Biggest upgrade)

• Use ESM-2 (or ESM-3 when available) embeddings instead of hand-crafted features.
• Compute "longevity direction vectors" in embedding space between orthologs from short-lived (C. elegans, mouse) vs long-lived (elephant, human, bowhead whale, naked mole rat) species.
• Augment with AlphaFold3 structural features and known longevity transcriptomic signatures (Tyshkovskiy et al., 2023).

Phase 2: Classification/Discrimination

• Keep SVM on ESM-2 embeddings as interpretable baseline.
• Optional upgrades: XGBoost, attention-based classifiers, or contrastive learning on the longevity vector.

Phase 3: Generation (LSTM → modern generators)

• Best drop-in replacement: Fine-tuned ProGen2 or ProtGPT2 — these understand protein "grammar" at a level plain LSTMs cannot match.
• Strong alternative: Latent-space VAE with longevity direction arithmetic (encode → add vector → decode). This is the cleanest mathematical upgrade to your original concept.
• State of the art: Diffusion models (RFdiffusion, EvoDiff, Chroma). Condition on the longevity vector or specific aging targets (AMPK, sirtuins, mTOR interfaces).

Phase 4: Validation & Filtering

• AlphaFold3 for structure quality (pLDDT filtering) and binding prediction.
• DiffDock or similar for affinity scoring.
• Toxicity, solubility, and aggregation predictors.
• Final scoring against biological age clocks (epigenetic, proteomic).

Phase 5: Wet Lab Loop

• Synthesize only top ranked oligopeptides (hexapeptides remain practical).
• Test on short-lived models.
• Feed results back into the model (active learning).

Recommended Stack (2026)

• Embeddings: ESM-2/3
• Generator: ProGen2 fine-tune or EvoDiff/RFdiffusion
• Structure filter: AlphaFold3
• Classifier: SVM on ESM embeddings (keep for interpretability) or modern GNN
• Experiment tracking: Weights & Biases + the Prometheus-style agentic network we looked at earlier

Why This is Better Than 2018

• Protein language models understand long-range dependencies and evolutionary context far beyond LSTM.
• Diffusion/VAE methods produce much more diverse and foldable sequences.
• AlphaFold3 removes most of the "will this even fold?" uncertainty before spending money on synthesis.
• The longevity vector concept maps beautifully onto latent space arithmetic and conditional generation.

This approach is now very feasible as a serious research project. The biological hypothesis was ahead of its time—the tools finally caught up.

This approach is feasible today as a serious research project. The biological hypothesis was ahead of its time — the tools finally caught up.

Open to collaboration or further development.

#longevity #ai #peptides #bioinformatics