Implementation & Execution of DEEPSYNTH-AMP Pipeline

[soykothosen]

📝 Issue Description

This issue tracks the implementation, configuration, execution, and validation of the DEEPSYNTH-AMP (De novo Evolutionary Ensemble Policy Synthesis-Yaw Network Transfer Heuristic for Antimicrobial Molecular Peptide Design) pipeline.

The primary objective is to implement a lightweight, diffusion-based reinforcement learning pipeline that simultaneously optimizes antimicrobial peptides (AMPs) for three core objectives: Antibacterial Activity (MIC), Safety (Low Human Toxicity), and Chemical Synthesizability using a Pareto-based multi-objective optimization framework.

---

🎯 Project Scope & Core Innovations

• Weight-Free Multi-Objective Optimization: Generates a Pareto front of non-dominated candidates, eliminating arbitrary reward weights.
• Synthesizability-Aware Diffusion: Addresses the laboratory manufacturing gap by explicitly integrating synthesizability constraints.
• Democratized Design: Completely open-source pipeline reproducible on a single GPU within one week with < $10 compute cost.

---

🗺️ Implementation Milestones & Task Checklist

Module A: Data & Model Preparation

• Step 1: Environment Configuration
  • Install Python 3.10+, PyTorch (with CUDA support), HuggingFace transformers, RDKit, and AiZynthFinder.
  • Verification: All packages import successfully; GPU is correctly detected by PyTorch.
• Step 2: Dataset Acquisition and Processing
  • Download public datasets: DRAMP 3.0 (20,234), ToxinPred (2,000), PeptideAtlas (10,000), and AiZynth Database (500K rules).
  • Apply processing rules: Filter sequence lengths between 5-50 Amino Acids (AA), remove duplicates, precompute retrosynthetic synthesizability scores, and split data into 70% Train / 15% Val / 15% Test.
  • Verification: Training set contains 15,000+ sequences; synthesizability scores range from 0 to 1.
• Step 3: Model Loading and Initialization
  • Load the pretrained EvoDiff model (38M parameters) from HuggingFace hub.
  • Freeze base layers and add LoRA adapters (Rank=8, Alpha=16) for parameter-efficient fine-tuning.
  • Verification: 100 generated test peptides pass regular expression (regex) validation for standard amino acids.

Module B: Training & Optimization

• Step 4: Pareto Multi-Objective Framework Implementation
  • Code the three scoring functions: Activity ($f_1$), Safety ($f_2$), and Synthesizability ($f_3$).
  • Implement the Adaptive Reward Function based on conservative lower-bound mapping:
    $$R(s) = \min(f_1(s), f_2(s), f_3(s))$$.
  • Verification: All unit tests pass via pytest; rewards range dynamically between -0.5 and 1.0.
• Step 5: Reinforcement Learning Fine-Tuning
  • Setup Proximal Policy Optimization (PPO) fine-tuning loop for 50 epochs (Batch size = 64, 100 batches per epoch, generating 6,400 sequences/epoch).
  • Use a learning rate of 3e-4 and clip range of 0.2. Save model checkpoints every 5 epochs.
  • Verification: Mean reward increases over training; checkpoints successfully export.

Module C: Evaluation & Selection

• Step 6: Pareto Front Generation and Candidate Selection
  • Generate 10,000 candidate sequences using the final checkpoint.
  • Run non-dominated sorting to extract the Pareto-optimal set and calculate Crowding Distance to preserve structural diversity.
  • Filter duplicates and pick the Top 20 diverse Pareto-optimal candidate peptides.
  • Verification: Pareto set identified accurately; Top 20 sequences are well-distributed across the objective space.
• Step 7: Publication-Ready Packaging
  • Document the GitHub repository with a complete README.md.
  • Prepare a 1-page research summary report and a 5-slide presentation deck.

---

📊 Expected Performance Benchmarks

The generated output metrics should ideally align with or exceed the following target values based on proposal estimations:

• Hit Rate (MIC < 4 μg/mL): Competitive therapeutic threshold.
• Synthesizable Rate (Score > 0.7): ~78%.
• Toxicity Rate (Human cells): Restricted to ~20%.
• Diversity (Pairwise Levenshtein Distance): ~0.64.

---

⚡ Risk Mitigation Matrix

If execution encounters any bottlenecks, refer to these approved mitigation strategies:

Identified Risk	Probability	Impact	Mitigation Strategy
GPU Out of Memory (OOM)	Low	Medium	Reduce batch size and enable gradient accumulation.
Poor Pareto Front Quality	Medium	High	Adjust adaptive reward strategy or increase exploration parameters.
Low Sequence Novelty	Medium	High	Introduce a structural diversity bonus or increase sampling temperature.
Training Non-Convergence	Low	High	Reduce the PPO learning rate and expand the clip range.

---

🕒 Estimated Effort Summary

• Total Workload: ~28 hours.
• Suggested Timeline: 7 consecutive days at an average of 4 hours per day.

[soykothosen]

DEEPSYNTH_AMP_Research_Proposal.pdf