MOZAIC: Compound Growth via In Silico Reactions and Global Optimization using Conformational Space Annealing

MOZAIC is a novel computational framework that combines reaction-based fragment growing with Conformational Space Annealing to generate chemically diverse, synthesizable, and high-affinity drug candidates while optimizing key properties like synthetic accessibility and solubility.

The Gist

Imagine you are a master chef trying to create the perfect new dish to cure a specific disease. You have a small, tasty ingredient (a "fragment") that you know works well with the disease, but it's not strong enough on its own. Your goal is to add other ingredients to make it a full, powerful meal, but you have two big rules:

1. The final dish must be delicious (it must bind tightly to the disease).
2. The dish must be easy to cook in a real kitchen (it must be possible to synthesize chemically).

This is the challenge of Fragment-Based Drug Discovery (FBDD). For a long time, computers could suggest new "recipes" (molecules), but often those recipes were impossible to cook in a real lab, or the computer got stuck making the same boring dish over and over again.

Enter MOZAIC, a new computer program developed by Jinhyeok Yoo and Woong-Hee Shin. Think of MOZAIC as a super-intelligent, global-searching sous-chef that never gets stuck in a rut.

Here is how MOZAIC works, broken down into simple concepts:

1. The "Recipe Book" of Real Chemistry

Many computer programs invent molecules that look cool on a screen but don't exist in the real world. MOZAIC is different. It only uses a "recipe book" based on real chemical reactions (like following a strict cookbook).

• The Analogy: Instead of just gluing random Lego bricks together (which might fall apart), MOZAIC only snaps bricks together using the specific, proven connectors that real chemists use. This ensures that every molecule it designs can actually be built in a real laboratory.

2. The "Conformational Space Annealing" (CSA) – The Smart Search

This is the secret sauce. Imagine you are looking for the highest peak in a massive, foggy mountain range (the "Chemical Space").

• The Problem: A normal computer might climb a small hill, think it's the top, and stop there. This is called getting stuck in a "local minimum."
• MOZAIC's Solution: MOZAIC uses an algorithm called Conformational Space Annealing (CSA). Think of this as sending out a whole team of hikers (a "population") to explore the mountains at the same time.
  • At first, the team spreads out wide to see the whole map (exploring diverse options).
  • As they get closer to the best peaks, they start narrowing their search, focusing only on the very highest spots.
  • They constantly swap ideas and combine their best routes (like mixing two good recipes to make a better one).
  • The Result: MOZAIC doesn't just find a good solution; it finds the best possible solution by avoiding the "fake peaks" that trap other programs.

3. The "Scorecard" (The Objective Function)

MOZAIC doesn't just look for strength; it balances three things, like a judge scoring a talent show:

1. Binding Affinity (The "Grip"): How tightly does the drug grab onto the disease? (Measured by a tool called AutoDock Vina).
2. Synthetic Accessibility (The "Cooking Difficulty"): How hard is it to make this drug? (A low score means it's easy to cook).
3. Drug-Likeness (The "Health Score"): Is it safe and likely to work in the human body? (Measured by QED).

MOZAIC tweaks the recipe until it finds the perfect balance where the drug is strong, easy to make, and safe.

4. What Did They Prove?

The authors tested MOZAIC on three different "challenges":

• Challenge 1 (PDE10A): They started with a weak drug and asked MOZAIC to make it stronger. MOZAIC created a molecule that was even better than the current "gold standard" drug used in clinics, and it did so by following real chemical rules.
• Challenge 2 (TrmD): They compared MOZAIC to another famous program (CReM-Dock). While the other program made many similar-looking drugs, MOZAIC created a much wider variety of unique shapes (like a diverse menu vs. a menu of just three variations of pasta).
• Challenge 3 (The "Blind" Test): They tested MOZAIC on a protein structure that wasn't perfectly known (like a blurry photo). Even with imperfect information, MOZAIC found strong drugs, proving it's robust and doesn't panic when the data isn't perfect.

Why Does This Matter?

In the past, computer-designed drugs were often "theoretical ghosts"—they looked great on a screen but couldn't be made. MOZAIC bridges the gap between the computer screen and the real-world lab.

It's like having a chef who not only invents a delicious new dish but also writes down the exact shopping list and cooking steps so you can actually make it in your own kitchen. By using real chemistry rules and a smart, global search strategy, MOZAIC helps scientists find new cures faster, cheaper, and with more variety than ever before.

Technical Summary

1. Problem Statement

Fragment-based drug discovery (FBDD) is a powerful strategy for exploring chemical space by combining small molecular fragments. However, existing computational methods for FBDD face three critical limitations:

• Synthetic Feasibility: Many algorithms (e.g., genetic algorithms, deep learning models) generate molecules that are chemically valid but difficult or impossible to synthesize in a laboratory because they lack defined synthetic pathways.
• Local Minima Traps: Population-based optimization methods (like standard genetic algorithms) often get trapped in local minima, failing to find the global optimum in the vast chemical space.
• Bias and Diversity: AI-driven approaches often inherit biases from their training sets, leading to the generation of repetitive analogues rather than structurally diverse chemotypes.
• Rigid Objectives: Many tools struggle to balance multiple, often conflicting, objectives (e.g., binding affinity vs. solubility vs. synthetic accessibility) without manual re-tuning.

2. Methodology: The MOZAIC Framework

The authors propose MOZAIC (Molecule Optimization via AutoDock Vina, in silico fragment reaction, and Conformational Space Annealing), a novel framework designed to generate synthesizable, drug-like compounds with optimized binding affinity.

Core Components:

1. Reaction-Based Fragment Growing:

   • Unlike methods that simply attach fragments to any available atom, MOZAIC uses SMARTS (SMILES Arbitrary Target Specification) rules derived from real organic chemistry databases (AutoClickChem, Robust Rxn, ChemoDOTS).
   • It employs 104 predefined reaction rules (89 bimolecular) to ensure that every generated molecule has a valid, documented synthetic route.
   • A fragment library of ~893,000 compounds (filtered from ZINC20) is mapped to these reaction rules.

2. Conformational Space Annealing (CSA):

   • Instead of a standard genetic algorithm, MOZAIC utilizes CSA, a powerful global optimization algorithm.
   • Mechanism: CSA maintains a "bank" of solutions (molecules). It iteratively generates "children" via crossover (combining reaction histories of two parents) and mutation (randomly re-selecting reaction components).
   • Distance Cutoff ($D_{cut}$): A key feature of CSA is the gradual reduction of a distance cutoff parameter. Initially, $D_{cut}$ is large to encourage exploration of diverse chemical space. As iterations proceed, $D_{cut}$ shrinks, focusing the search on refining the global optimum.
   • Selection: New molecules are accepted into the bank only if they improve the objective function score or if they are sufficiently diverse (distance > $D_{cut}$) from existing members.

3. Multi-Objective Function:

   • The fitness of a molecule is evaluated using a composite score:
     • Binding Affinity: Calculated via AutoDock Vina (lower $\Delta G$ is better).
     • Synthetic Accessibility (SA): Calculated via RDKit (lower score is better; 1–10 scale).
     • Drug-likeness (QED): Calculated via RDKit (0–1 scale; higher is better).
   • Flexibility: The objective function is modular. Users can swap components (e.g., replacing QED/SA with ESOL for solubility optimization) to target specific physicochemical properties.

4. Workflow:

   • Input: 3D protein structure, binding site residues, and a starting SMILES string.
   • Phase 1 (Initial Bank): Generate 1,000 candidate molecules via in silico reactions, then select 60 diverse seeds using the MaxMinPicker algorithm (based on ECFP4 fingerprints).
   • Phase 2 (CSA Iteration): Run three rounds of CSA. In each round, seeds are selected, crossed over/mutated, and the bank is updated based on score and diversity.
   • Output: A final bank of optimized molecules with docking poses, interaction profiles (via PLIP), and reaction history.

3. Key Contributions

• Synthesizable Design: By grounding fragment growth in real reaction rules (SMARTS), MOZAIC guarantees that proposed compounds are chemically synthesizable, bridging the gap between computational design and wet-lab execution.
• Global Optimization: The application of CSA to FBDD overcomes the local minima problem common in genetic algorithms, ensuring a more thorough exploration of chemical space.
• Diversity Generation: MOZAIC produces molecules with higher scaffold entropy and lower structural redundancy compared to existing tools, avoiding the "analogue trap."
• Customizable Optimization: The framework allows users to easily redefine the objective function to prioritize specific properties (e.g., solubility) without rewriting the core code.

4. Results and Benchmarks

The authors validated MOZAIC using three distinct benchmark scenarios:

A. PDE10A Optimization (Initial Bank Sensitivity)

• Task: Optimize a hit compound (Ki = 8.7 $\mu$M) toward the lead MK-8189 (Ki = 0.029 nM).
• Finding: MOZAIC generated a compound with a Vina score of -10.359 kcal/mol (superior to MK-8189's -7.770).
• Insight: The study demonstrated that using MaxMinPicker to select the initial bank (maximizing diversity) yielded significantly better final optimization results than selecting based on initial scores or random sampling.

B. TrmD Comparison (vs. CReM-Dock)

• Task: Compare MOZAIC against CReM-Dock (a genetic algorithm-based tool).
• Finding: MOZAIC produced compounds with stronger predicted binding affinities (mean -9.683 kcal/mol vs. -8.428 for CReM-Dock).
• Diversity: MOZAIC generated 45 unique scaffolds (entropy 3.59) compared to CReM-Dock's 32 (entropy 2.79). MOZAIC compounds showed lower Tanimoto similarity to the starting molecule (0.135 vs. 0.285), indicating broader chemical exploration.

C. Robustness (Receptor Structure Uncertainty)

• Task: Optimize ligands for the $\beta$2-adrenergic receptor (ADRB2) using both a crystal structure (2RH1) and an AlphaFold3 predicted structure.
• Finding: MOZAIC successfully identified high-affinity compounds (Vina score ~ -11.2) on both structures. The optimized compounds retained favorable docking scores even when the receptor conformation was uncertain, demonstrating robustness.

D. Custom Objective (PPAR$\gamma$ Solubility)

• Task: Optimize for Solubility (ESOL) and Binding Affinity simultaneously.
• Finding: Starting from a lipophilic ligand, MOZAIC successfully modified the structure to improve predicted logS (from -4.42 to -2.85) while maintaining or improving binding affinity. This confirmed the flexibility of the modular objective function.

5. Significance and Conclusion

MOZAIC represents a significant advancement in in silico drug design by effectively merging chemical synthesis constraints with global optimization.

• Practical Impact: It addresses the "synthesis gap" by ensuring that computational suggestions are chemically feasible, reducing the time and cost of experimental validation.
• Algorithmic Innovation: The integration of CSA with reaction-based growing provides a robust mechanism to escape local optima and discover novel chemotypes that traditional methods miss.
• Future Directions: The authors acknowledge limitations regarding the accuracy of AutoDock Vina (lack of explicit solvent/flexibility) and the fixed reaction library. Future work aims to integrate quantum mechanical scoring, broader reaction sets, and blind docking capabilities.

In summary, MOZAIC offers a versatile, synthesizable, and diverse platform for fragment-to-lead optimization, positioning itself as a valuable tool for modern drug discovery pipelines.