Motivation

The original ideas21 sketch proposed using machine learning to explore peptide space for longevity properties by mining patterns across species.

A natural extension is systematic exploration of the dipeptide chemical space (400 base compounds with 20 standard amino acids). Many bioactive dipeptides are already known (antioxidant, ACE-inhibitory, anti-inflammatory, opioid, etc.) from the BIOPEP database. Expanding to non-standard amino acids, methylation, acetylation, and γ-linkages would create a rich, tractable library.

Combinatorial Scope

Base space:

• 20 standard amino acids → 20×20 = 400 dipeptides (order matters: Ala-Gly ≠ Gly-Ala)

Expanded space (realistic for screening):

• Mirror-image (D-) amino acids (enantiomers of the 20 standard L-amino acids) — these often have greater resistance to proteases, extending half-life. Including D-variants roughly doubles the space.
• • 8 non-standard/analogs (selenocysteine, pyrrolysine, ornithine, taurine, GABA, β-alanine, hydroxyproline, sarcosine) → ~28 AAs → ~784 dipeptides (or ~1,500+ when including D-forms and mixtures)
• Methylation variants (N-methyl on N-terminus or side chains) on key residues → selective addition of ~200–300 more
• Acetylation variants (common PTM)
• γ-linkages (as in γ-Glu-Cys, the GSH precursor) — adds stability and distinct pharmacology

Total searchable space: ~1,000–2,000+ compounds (still computationally trivial). The term you were thinking of is "enantiomers" (non-superimposable mirror images). "Isomorphs" usually refers to crystals of similar shape but different composition.

Proposed Computational Pipeline

1. Enumeration

   • Script to generate all combinations in SMILES or InChI format
   • Filter for drug-like properties (molecular weight < 300 Da for dipeptides, reasonable logP)

2. Bioactivity Prediction

   • QSAR/QSPR models trained on BIOPEP and other bioactive peptide databases
   • Predict antioxidant capacity, Nrf2 activation, anti-inflammatory, ACE-inhibition, etc.
   • Use graph neural networks on molecular graphs for better performance than traditional descriptors

3. Targeted Docking

   • Dock high-scoring candidates to relevant targets:
     • Keap1 (for Nrf2 activation)
     • GPx or glutathione reductase
     • Inflammatory targets (COX-2, NF-κB)
     • Aging-related proteins (sirtuins, mTOR, AMPK interfaces)
   • Use AutoDock Vina or DiffDock for peptide-specific docking

4. Prioritization & Synthesis Recommendation

   • Score by predicted activity + synthetic feasibility + predicted ADME/Tox (solubility, stability, renal clearance)
   • Prioritize cysteine- or γ-Glu-containing sequences for glutathione support
   • Prioritize stable forms (γ-linkage, D-amino acids, N-methylated)

5. Experimental Loop

   • Synthesize or purchase top 10–20 candidates
   • Test in cell models for GSH induction, antioxidant activity, anti-senescence effects
   • Feed results back into the ML model (active learning)

Tools & Resources

• Databases: BIOPEP-UWM, PepBank, APD2 (antimicrobial), CPPsite (cell-penetrating peptides)
• Libraries: RDKit for enumeration and descriptors, DeepChem or PyTorch Geometric for ML models
• Docking: AutoDock Vina, DiffDock, or Rosetta for peptides
• Peptide-specific ML: PepNet, DeepPep, or fine-tuned ProtT5/ESM-2 models
• Hardware: Consumer GPU sufficient for initial screen (400–1500 compounds is small)

Feasibility & Cost Estimate

• Computational phase: 1–2 weeks on a decent laptop/GPU instance
• Synthesis of top 20 candidates: ~$2,000–5,000 (custom synthesis labs)
• In vitro testing: cell-based antioxidant/GSH assays (~$3,000–8,000)

Total initial proof-of-concept budget: ~$10,000 or less if leveraging open-source tools and academic collaborations.

This project would be a natural fit for Beach.science — completely tractable, crosses computational biology, chemistry, and aging research, and could discover novel dipeptides safe for complex geriatric profiles.