AI-Enforced Ultra-Large Virtual Screening Discovers Potent CD28 Binders

This study demonstrates that the AI-enforced PyRMD2Dock platform successfully overcomes computational bottlenecks in ultra-large virtual screening to discover potent, submicromolar small-molecule binders of the challenging immune receptor CD28, which effectively disrupt its interaction with CD80 and inhibit cytokine secretion in human models.

The Gist

Imagine you are a locksmith trying to pick a very tricky lock. Most locks have a deep, obvious keyhole where you can easily slide a key in. But the lock you are trying to open is the CD28 protein, a crucial part of our immune system. The problem? It doesn't have a deep keyhole. Instead, it has a flat, smooth surface with only a tiny, shallow dent.

For decades, scientists have struggled to find small molecules (our "keys") that can fit into this shallow dent to either turn the lock on or off. Traditional methods of searching for these keys are like trying to find a needle in a haystack by looking at one piece of hay at a time. It's slow, expensive, and often fails.

This paper describes a brilliant new strategy that uses Artificial Intelligence (AI) to speed up the search and successfully find working keys for the first time.

The Story of the "Super-Scanner"

Here is how the scientists did it, broken down into simple steps:

1. The Problem: The Flat Surface

Think of the CD28 protein as a flat table. Usually, drugs need a deep bowl on the table to sit in. But this table is flat. The scientists realized that right next to the main "meeting spot" on the table, there was a small, hidden nook (a cleft) formed by a few specific amino acids. They decided to try to jam their keys into this nook instead of the main flat surface.

2. The Old Way vs. The New Way

• The Old Way: To find a key, you would have to physically test every single molecule in a library of millions. It's like trying to find a specific person in a stadium of 50 million people by asking every single person, "Are you the one?" It would take forever and cost a fortune.
• The New Way (PyRMD2Dock): The scientists built a "Super-Scanner" (an AI tool called PyRMD2Dock).
  • Step A: They took a small sample of 2.4 million molecules and ran them through a computer simulation to see how well they fit into the nook.
  • Step B: They taught the AI to recognize the "shape" and "feel" of the molecules that fit well. The AI learned, "Ah, molecules with this specific curve and this specific chemical charge seem to stick here."
  • Step C: Once the AI learned the pattern, they let it scan the remaining 46 million molecules in seconds. The AI didn't need to simulate every single one; it just used what it learned to predict which ones would fit.

3. The Great Filter

The AI came back with a list of 20 million "maybe" candidates. That's still too many to test in a lab. So, the scientists applied a series of filters, like a sieve:

• The Energy Check: They only kept the ones that looked like they would stick tightly.
• The Shape Check: They made sure the molecules actually touched two specific "anchor points" inside the nook (like a key touching both the top and bottom of a lock).
• The Crowd Control: They grouped similar molecules together and picked the best representatives from each group to ensure they had a diverse set of keys to try.

After all this filtering, they went from 46 million down to just 232 top candidates.

4. The Lab Test: Do the Keys Work?

The scientists bought 150 of these top candidates and tested them in the real world.

• The Result: It was a massive success! About 7.3% of the tested molecules actually worked. In drug discovery, finding one working molecule out of 100 is a huge win; finding 11 is a miracle.
• The Champions: Two molecules, named 100 and 104, were the stars. They didn't just stick; they stuck very tightly (in the sub-micromolar range, which is like a super-strong magnet).

5. The Real-World Impact

Finding a key that fits the lock is only step one. The key needs to actually open the door.

• Breaking the Connection: The scientists showed that these keys successfully blocked CD28 from shaking hands with its partner protein (CD80). This is like putting a piece of gum in the lock so the key can't turn.
• Stopping the Alarm: In living cells, CD28 usually acts like a gas pedal for the immune system. The scientists showed that compounds 100 and 104 pressed the brake. They stopped immune cells from over-activating and releasing inflammatory chemicals (cytokines).
• The Final Test: They tested these compounds in complex 3D models that mimic human tumors and airway tissues. The compounds worked there, too, reducing inflammation in a realistic human environment.

The Big Picture

This paper is a game-changer because it proves that AI can solve problems that humans thought were impossible.

• Before: Scientists thought the flat surface of CD28 was "undruggable" (no key could fit).
• Now: By using AI to scan a massive library of 48 million molecules, they found potent new drugs that fit into the shallow nook.

The Analogy:
Imagine trying to find a specific grain of sand on a beach.

• Traditional Science: You walk the beach, picking up one grain at a time and checking it. You might spend your whole life and never find it.
• This Study: You hire a drone (the AI) that flies over the beach, learns what the "special grain" looks like from a small sample, and then instantly highlights the 200 most likely grains for you to pick up. You then test those 200, and boom—you find the treasure.

This study validates that AI-driven "virtual screening" is not just a theory; it is a powerful, real-world tool that can unlock new treatments for difficult diseases by finding needles in haystacks faster than ever before.

Technical Summary

1. Problem Statement

The study addresses two major challenges in computational drug discovery:

• Targeting Protein-Protein Interactions (PPIs): Small molecules historically struggle to inhibit PPIs because these interfaces are often broad, solvent-exposed, and lack deep, well-defined binding pockets (orthosteric cavities).
• Computational Scalability: Traditional Structure-Based Virtual Screening (SBVS) using molecular docking is computationally prohibitive when applied to ultra-large chemical libraries (billions of compounds), making it difficult to discover novel chemotypes for difficult targets.

Specific Target: The study focuses on CD28, a critical immune checkpoint receptor. CD28 is a homodimeric receptor with a flat, solvent-accessible extracellular $\beta$-sandwich surface that binds ligands CD80/CD86. The absence of a deep cavity makes it a "structurally demanding" model system for developing small-molecule modulators.

2. Methodology

The authors employed a novel, AI-enforced workflow called PyRMD2Dock, integrated within the PyRMD Studio suite. This approach bridges Ligand-Based Virtual Screening (LBVS) and Structure-Based Virtual Screening (SBVS).

The Workflow:

1. Target Definition: Instead of the canonical flat interface, the team identified a distinct, adjacent cleft on CD28 (delimited by residues H38, F93, K95, D106, N107, K109, S110) based on structural homology with CTLA-4 and its antibody interaction (tremelimumab).
2. Training Set Generation:
   • A chemically diverse subset of 2.4 million compounds was selected from the Enamine REAL Diversity Space (ERDS) using RDKit-based diversity picking (ECFP6 fingerprints).
   • These compounds were docked into the CD28 cleft using AutoDock-GPU (AD4-GPU).
3. Machine Learning Model Training:
   • Docking scores ($\Delta G_{AD4}$) were used to classify compounds as "active" or "inactive" based on specific energy thresholds.
   • 672 classification models were generated by varying activity/inactivity thresholds and $\epsilon$ cutoffs (projection distance parameters).
   • The models utilized MinHash fingerprints (MHFP) and a rigorous StratifiedGroupKFold splitting strategy (coupled with Butina clustering) to prevent data leakage and ensure generalization.
   • The optimal model was selected based on the best trade-off between True Positive Rate (TPR), False Positive Rate (FPR), Precision, and F-score.
4. Ultra-Large Scale Screening:
   • The optimized model screened the remaining ~46 million compounds in the ERDS.
   • The top 20 million predicted binders were filtered down to 650,000 based on high confidence scores.
5. Refinement and Prioritization:
   • The 650,000 compounds were re-docked.
   • Interaction Filtering: Compounds were required to form simultaneous contacts with key residues K109 (ionic anchor) and C94 (bottom of the pocket).
   • SASA Analysis: Solvent-Accessible Surface Area calculations ensured steric fit.
   • Clustering: Molecules were clustered to ensure chemical diversity, resulting in a final pool of 232 prioritized ligands.
6. Experimental Validation:
   • 150 compounds were purchased and synthesized on demand.
   • Tiered Validation:
     • Dianthus TRIC Assay: Primary screening for signal perturbation.
     • Microscale Thermophoresis (MST): Quantification of direct binding affinity ($K_d$).
     • Competitive ELISA: Assessment of CD28-CD80 disruption ($IC_{50}$).
     • Cellular Reporter Assays: Luciferase-based blockade of CD28 signaling.
     • Physiological Models: 3D Tumor-PBMC co-cultures and Human PBMC-Airway Tissue co-cultures to measure cytokine suppression (IFN-$\gamma$, IL-2, TNF-$\alpha$).

3. Key Contributions

• First Prospective Application: This is the first real-world, prospective application of the PyRMD2Dock platform, demonstrating its viability for ultra-large scale screening.
• AI-Enforced Efficiency: The study proves that training ML models on a small, docked subset (2.4M) can effectively predict binding for ultra-large libraries (48M+), achieving a 3.3-fold speedup over previous software versions while maintaining accuracy.
• Novel Targeting Strategy: Successfully identified a non-canonical, adjacent cleft on CD28 as a viable binding site for small molecules, overcoming the "flat interface" challenge.
• High Hit Rate: Achieved a remarkable experimental hit rate, identifying multiple direct binders from a pool of 150 candidates.

4. Key Results

• Hit Identification: Out of 150 tested compounds, 12 primary hits were identified in the initial screen.
• Lead Compounds: Two compounds, 100 and 104, emerged as the most potent leads:
  • Compound 100: $K_d = 343.8$ nM (MST); $IC_{50} = 231.4$ nM (ELISA); $IC_{50} = 748.9$ nM (Cellular).
  • Compound 104: $K_d = 407.1$ nM (MST); $IC_{50} = 840.6$ nM (ELISA); $IC_{50} = 1.13$ $\mu$M (Cellular).
• Mechanism of Action: Both compounds demonstrated potent competitive disruption of the CD28-CD80 interaction and effectively blocked downstream costimulatory signaling.
• Binding Modes: Computational analysis revealed that 100 and 104 adopt distinct binding orientations (Cluster 3 and Cluster 1, respectively) within the same cleft, yet both engage key residues (H38, K95, K109, K2, D106) via halogen bonds, cation-$\pi$ interactions, and hydrogen bonds.
• Physiological Efficacy:
  • In Tumor-PBMC 3D co-cultures, both compounds dose-dependently reduced IFN-$\gamma$, IL-2, and soluble CD69 secretion.
  • In Human PBMC-Airway Tissue models, Compound 100 significantly attenuated cytokine release (IFN-$\gamma$, IL-2, TNF-$\alpha$), mimicking the effect of the clinical-stage antagonist FR104.

5. Significance

• Validation of AI-Driven SBVS: The study validates that AI-enforced workflows can successfully navigate the computational bottlenecks of ultra-large chemical space screening, making the discovery of PPI inhibitors feasible.
• Expanding the Druggable Proteome: It demonstrates that shallow, solvent-exposed immune receptor interfaces, previously considered "undruggable" by small molecules, can be successfully targeted.
• Therapeutic Potential: The discovery of sub-micromolar CD28 antagonists (100 and 104) offers new chemical starting points for treating autoimmune diseases and modulating immune responses in cancer, potentially with improved pharmacokinetic profiles compared to antibody-based therapies.
• Scalable Protocol: The PyRMD2Dock protocol provides a scalable, democratized framework (via GUI) for researchers to mine massive libraries for challenging targets without requiring advanced coding expertise.