Evolutionary exploration of drug-like chemical space utilizing generative AI and virtual screening

This paper presents an evolutionary AI framework that integrates generative models, reaction-based substructure searching, and iterative fine-tuning to efficiently navigate ultra-large chemical spaces, successfully identifying and validating novel, synthetically accessible ligands for the $\mu$-opioid receptor within the Enamine REAL Space.

The Gist

Imagine you are looking for a specific key to open a very complicated lock (a disease-causing protein). The problem is that the "key shop" has a catalog containing more keys than there are grains of sand on all the beaches on Earth. Trying to test every single key one by one is impossible; it would take longer than the age of the universe.

This paper describes a clever new way to find the right key without checking every single one. The researchers combined three powerful tools: Evolutionary Algorithms (nature's way of survival), Generative AI (a creative robot artist), and Virtual Screening (a super-fast computer simulator).

Here is how their "Evolutionary AI" works, explained through a simple story:

1. The Problem: The Impossible Library

In the past, scientists had physical boxes of chemicals (like a library of 1 million books) to test. Now, we have "virtual libraries" with trillions of potential molecules. But you can't physically build and test trillions of them. It's too expensive and slow.

2. The Solution: A "Survival of the Fittest" Robot

The researchers created a digital ecosystem where molecules compete to see which one fits the "lock" (the target protein) best. Here is the step-by-step process:

• Step 1: The Random Crowd (The Starting Population)
  Imagine a robot generates 1,000 random, weird-looking keys. It doesn't know if they work; it just makes them up.
• Step 2: The Tryout (Virtual Screening)
  The computer quickly tries to fit these 1,000 keys into the lock. It scores them: "This one fits okay, this one is terrible, this one is amazing."
• Step 3: The "Survival of the Fittest" (Selection)
  The computer picks the top 20% of keys that fit the best. The losers are thrown away.
• Step 4: The Creative Remix (Generative AI)
  This is the magic part. Instead of just mixing the winning keys together like a blender (which is how old computers did it), the researchers use a Generative AI.
  • Think of the AI as a master chef who has tasted the winning keys.
  • The chef looks at the winners and says, "I see a pattern here. They all have a specific shape on the left and a specific texture on the right."
  • The chef then cooks up a brand new batch of keys that are better versions of the winners, but with some random surprises (mutations) to keep things interesting.
• Step 5: The Reality Check (Synthesizability)
  Before the AI gets too crazy, the system checks: "Can we actually build this key in a real lab?" It breaks the new key down into Lego blocks to make sure the pieces exist in a real chemical store. If the key is impossible to build, it gets discarded.
• Step 6: Repeat
  The new batch of keys goes back to Step 2. The AI learns from the winners, makes them better, and the cycle repeats 20 times. With every round, the keys get closer to being the perfect fit.

3. The Special Twist: The "Temperature-Sensitive" Key

The researchers didn't just want any key; they wanted a key that only works in acidic environments (like inside a tumor or an inflamed area) but doesn't work in normal body conditions. This is crucial for drugs like painkillers (opioids) to stop them from causing addiction or side effects in the rest of the body.

• They told the AI: "Find keys that lock tight when it's 'sour' (acidic pH 6.5) but fall out when it's 'normal' (pH 7.4)."
• The AI learned to tweak the keys so that a tiny part of the key changes shape depending on the acidity, acting like a switch.

4. The Result: Real-World Success

After 20 rounds of this digital evolution, the AI proposed a few candidates. The researchers took the top 5, built them in a real lab, and tested them on actual human cells.

• The Outcome: It worked! One of the new molecules acted exactly as predicted. It bound strongly to the pain receptor in acidic conditions (good for pain relief) but barely stuck in normal conditions (good for safety).
• Why it matters: This proves that we don't need to test billions of chemicals physically. We can use a "smart robot" to evolve the best candidates in a computer, check if they are buildable, and then just build the very best ones.

The Big Picture Analogy

Imagine you are trying to find the perfect recipe for a cake that tastes great but has zero calories.

• Old Way: Bake 10 million random cakes, taste them, and throw away the bad ones. (Impossible).
• This Paper's Way:
  1. Bake 1,000 random cakes.
  2. Taste them and pick the top 200.
  3. Give the top 200 recipes to a Super Chef AI.
  4. The AI analyzes what made those cakes good, mixes the best ingredients, and invents 1,000 new recipes that are even better.
  5. The AI checks the pantry to make sure it can actually buy the ingredients.
  6. Repeat this 20 times.
  7. Bake the final 5 winners and serve them.

This approach allows scientists to explore the "universe" of possible medicines much faster and smarter than ever before, potentially leading to safer, more effective drugs with fewer side effects.

Technical Summary

1. Problem Statement

The identification of novel lead molecules in the vast drug-like chemical space (estimated at $10^{60}$ compounds) is a critical bottleneck in drug discovery.

• Limitations of Traditional Screening: High-throughput screening (HTS) relies on physical libraries (1–5 million compounds), which lack structural diversity and often miss potent hits.
• Limitations of Ultra-Large Virtual Screening: While virtual libraries like Enamine REAL Space have expanded to trillions ($>10^{15}$) of "make-on-demand" compounds, computationally screening every single molecule is prohibitive.
• Limitations of Pure Generative AI: While generative AI can design novel molecules de novo, ensuring these molecules are synthetically accessible (i.e., can be built from available building blocks) and physically tangible remains a major challenge.
• The Specific Challenge: The authors aimed to find pH-specific ligands for the $\mu$-opioid receptor (MOR). These ligands should bind strongly in acidic environments (e.g., inflamed tissue) but weakly in neutral pH to reduce side effects, a property difficult to optimize using standard screening methods.

2. Methodology

The authors developed an evolutionary algorithm framework that integrates molecular generative AI, reaction-based substructure searching, and iterative model fine-tuning. The workflow operates in a closed loop over 20 generations (iterations).

A. The Evolutionary Workflow

1. Initialization: A random starting population of 1,000 molecules is generated using a pre-trained generative model (REINVENT framework).
2. Evaluation (Objective Space):
   • Molecules are prepared for two conditions: pH 7.4 (neutral) and pH 6.5 (acidic).
   • Molecular Docking: Compounds are docked against two distinct MOR protein models (prepared at pH 7.4 and pH 6.5) using VirtualFlow and Quick Vina 2.
   • Scoring: A multi-objective optimization problem is converted into a single objective using a weighted average score. The goal is to maximize binding at pH 6.5 while minimizing binding at pH 7.4.
   • Formula used: $\bar{x}_w = \frac{1}{2} [x'(pH6.5) + (1 - x'(pH7.4))]$, where scores are min-max scaled and inverted.
3. Selection: The top 20% of performers (Pareto-optimal or best compromise) are selected as "parents."
4. Generative Variation (Mutation/Recombination):
   • The generative AI model is fine-tuned on the selected parent molecules.
   • To maintain diversity and prevent overfitting, ~4% random molecules ("noise") are added to the training set.
   • The fine-tuned model generates the next generation of 1,000 offspring.
5. Synthesizability Filter (Crucial Step):
   • Every AI-generated molecule is decomposed into fragments using BRICS reaction rules.
   • These fragments are matched against the Enamine REAL Space building block database (~1 million blocks).
   • Only molecules where all fragments exist in the database are retained, ensuring the output is synthetically accessible ("make-on-demand").

B. Experimental Validation

• Selection: From the final population, hits were filtered for:
  1. High predicted binding at pH 6.5 vs. low at pH 7.4.
  2. Specific protonation states (monoprotonated at pH 6.5, neutral at pH 7.4).
  3. Formation of a salt bridge with Asp149 (a key residue in MOR).
• Synthesis: Five top candidates were chemically synthesized.
• Assays:
  • Radio-labeled Displacement Assay (RDA): Measured competitive binding against [³H]-DAMGO at pH 7.4 and pH 6.0.
  • Molecular Dynamics (MD): 1 $\mu$s simulations to analyze binding stability and interaction networks (salt bridges, $\pi$-$\pi$ stacking) in protonated vs. deprotonated states.

3. Key Contributions

• Hybrid Framework: Successfully combined Evolutionary Algorithms (for global search and selection) with Generative AI (for intelligent mutation and recombination) and Virtual Screening (for fitness evaluation).
• Synthesizability Constraint: Integrated a fragment-matching pipeline that restricts the AI's search space to the Enamine REAL Space, ensuring all generated molecules are theoretically synthesizable from available building blocks.
• pH-Specific Optimization: Demonstrated a method to explicitly optimize for environmental sensitivity (pH-dependent binding), a complex property often missed by standard docking.
• Experimental Validation: Successfully moved from in silico prediction to in vitro validation, confirming the discovery of a novel pH-specific MOR ligand.

4. Results

• Iterative Improvement: Over 20 iterations, the mean weighted score of the top 100 ligands increased significantly ($p < 0.0001$).
• Differential Binding: The mean difference in binding free energy ($\Delta G_b$) between pH 6.5 and pH 7.4 increased, reaching ~0.33 kcal/mol for the top ligands.
• Protonation Shift: The proportion of molecules predicted to be neutral at pH 7.4 but monoprotonated at pH 6.5 rose from ~5% (Iteration 1) to ~60% (Iteration 20), indicating the AI learned the specific chemical feature required for pH sensitivity.
• Experimental Success:
  • Of 5 synthesized compounds, 4 showed competitive binding.
  • Compound 1 exhibited the desired profile:
    • Acidic pH (6.0): $IC_{50} = 167 \pm 179$ nM.
    • Neutral pH (7.4): $IC_{50} = 1424 \pm 541$ nM.
    • This represents a ~10-fold increase in potency at acidic pH.
• Mechanism of Action: MD simulations confirmed that at acidic pH, the protonated nitrogen of Compound 1 forms a stable salt bridge with Asp149 and $\pi$-$\pi$ stacking with Tyr328. At neutral pH, deprotonation breaks the salt bridge, causing the ligand to lose anchoring and conformational stability.

5. Significance

• Beyond HTS: This study demonstrates that AI-driven evolutionary workflows can explore chemical spaces orders of magnitude larger than physical libraries, identifying hits with superior initial activity and selectivity.
• Practical Applicability: By enforcing synthesizability constraints during the generative process, the approach bridges the gap between theoretical AI design and wet-lab reality, reducing the "synthesis gap."
• Targeted Discovery: The successful identification of pH-specific ligands proves the framework's ability to optimize for complex, multi-condition objectives (e.g., tissue-specific drug activation), which could lead to drugs with fewer systemic side effects.
• Future Outlook: The authors propose this as a new benchmark for early-stage drug discovery, potentially superseding traditional HTS. Future iterations could incorporate retrosynthetic analysis and receptor ensemble docking to further refine hit rates.