Discovery of resistance-resilient quinoline papain-like protease inhibitors through topology-constrained molecular generative design

Using an AI-driven topology-constrained molecular generative model combined with structure-guided optimization, researchers discovered GZNL-2016, a novel quinoline-based papain-like protease inhibitor that demonstrates potent antiviral activity against drug-resistant SARS-CoV-2 variants and favorable pharmacokinetic properties in vivo.

The Gist

🦠 The Problem: The Virus is a Master of Disguise

Imagine the SARS-CoV-2 virus (the one that causes COVID-19) is a notorious burglar. For a long time, we've had two main ways to stop it:

1. Vaccines: These are like "Wanted" posters showing the burglar's face (the Spike protein) so the police (your immune system) can recognize him.
2. Old Drugs: These are like specific locks on the burglar's favorite doors (viral enzymes).

The Catch: This burglar is a master of disguise. He keeps changing his face (mutations), making the "Wanted" posters useless. Worse, he keeps picking the locks on the doors we built. If a drug targets a specific part of the virus, the virus just mutates that part to break the drug's grip. This is called drug resistance.

🎯 The Target: The Virus's "Scissors"

Inside the virus, there is a pair of molecular scissors called PLpro (Papain-like protease).

• What it does: It cuts the virus's long protein chain into working pieces so the virus can replicate. It also snips off the "alarm bells" (immune signals) that human cells try to ring to warn the body of an infection.
• Why it's a good target: Unlike the "face" (Spike protein) that changes constantly, these scissors are very stable. They don't change much, making them a reliable target. However, the burglar has started changing the handle of the scissors (a spot called E167) to make old drugs slip right off.

🤖 The Solution: An AI Architect

The researchers didn't just guess new drugs; they used Artificial Intelligence (AI) to design them.

Think of the AI (called Tree-Invent) as a super-smart, 3D-printing architect.

1. The Blueprint: They gave the AI a picture of the virus's scissors and a "bad" drug (GZNL-P4) that worked well but broke down too fast in the liver (like a house built with rotting wood).
2. The Constraint: They told the AI: "Keep the handle and the blade the same, but completely redesign the roof (the head group) so it's stronger and doesn't rot."
3. The Generation: The AI didn't just pick random shapes. It built thousands of virtual molecules, testing them in a computer simulation to see which ones fit the scissors perfectly. It was like a digital game of Tetris, but the goal was to find the perfect molecular shape.

🔬 The Discovery: Finding the "Golden Key"

From the thousands of AI-generated designs, the team synthesized a few in the lab. They found a new family of drugs based on a quinoline structure (a specific chemical shape).

• The Star Player: They named the best one GZNL-2016.
• Why it's special:
  • The Grip: It fits the virus's scissors so tightly that it stops them from working almost instantly (nanomolar potency).
  • The Lockpick: Most importantly, when the virus tried to change its handle (the E167K mutation) to escape old drugs, GZNL-2016 didn't slip off. It held on tight. It's like a key that still works even if the lock has been slightly bent by a thief.
  • The Durability: Unlike the old drug, this one survives the liver's digestive process long enough to actually reach the virus in the body.

🐭 The Test: Saving the Mice

They tested this new drug on mice infected with a very tough version of the virus (Omicron variants).

• The Result: The mice given the drug lost very little weight and had significantly lower amounts of virus in their lungs compared to the untreated mice. It worked just as well as the current "gold standard" drug (Paxlovid) in the tests, but with a different mechanism.

🏆 The Big Picture

This paper is a victory for AI-driven drug discovery.

• Old Way: Try thousands of random chemicals in a lab, hoping one sticks. (Slow, expensive, hit-or-miss).
• New Way: Use AI to design the perfect chemical shape first, then build it. (Fast, precise, and smart).

In simple terms: The virus tried to change its locks to stop our old keys. This team used an AI architect to design a brand new "universal key" that fits the old locks and the new, tricky locks the virus invented. This gives us a powerful new weapon to fight future waves of the virus, even if the virus tries to mutate again.

Technical Summary

1. Problem Statement

The rapid emergence of SARS-CoV-2 variants has led to drug resistance against current antiviral therapies (e.g., Remdesivir, Nirmatrelvir) and the failure of monoclonal antibodies due to mutations in viral targets like the Spike protein, RdRp, and Main Protease (3CLpro).

• Specific Challenge: The Papain-like protease (PLpro) is a conserved viral target essential for replication and immune evasion. However, existing PLpro inhibitors (e.g., Jun12682, PF-07957472) suffer from two major limitations:
  1. Drug Resistance: Mutations at key residues, specifically E167K and Y268N, drastically reduce the potency of current inhibitors (up to 1000-fold loss in activity).
  2. Metabolic Instability: Previous lead compounds (e.g., GZNL-P4) showed poor liver metabolic stability, limiting their oral bioavailability and therapeutic potential.
• Goal: To discover a novel class of PLpro inhibitors with high potency, improved metabolic stability, and the ability to overcome resistance mutations, specifically the E167K variant.

2. Methodology

The study employed an AI-driven drug discovery (AIDD) strategy integrated with structure-based medicinal chemistry.

• Generative Model (Tree-Invent):
  • The researchers utilized Tree-Invent, a topology-constrained molecular generative model based on graph neural networks.
  • Scaffold Hopping: Starting from a known lead (GZNL-P4), the model was constrained to keep the "linker" and "tail" groups (which interact with the P2/P3 sub-pockets) fixed while generating new "head" groups to replace the unstable tricyclic dihydroacenaphthylene (DHAN) scaffold.
  • Constraints: The model was restricted to generate 5,6- or 6,6-fused bicyclic aromatic systems to mimic the $\pi$-$\pi$ stacking interactions observed in the binding pocket.
  • Reinforcement Learning (RL): RL was integrated to optimize molecules for high docking scores against the PLpro crystal structure (PDB: 7JRN).
• Virtual Screening & Synthesis:
  • Generated molecules were filtered by docking scores and Lipinski rules.
  • Nine top candidates were synthesized and tested for enzymatic inhibition.
• Structure-Guided Optimization:
  • Co-crystal structures were solved to understand binding modes.
  • A virtual library of ~2,300 compounds was designed to optimize the R2 position (tail group) to enhance salt-bridge interactions with the E167 residue.
• Biological Evaluation:
  • In Vitro: Enzymatic assays (IC50), biophysical binding assays (ITC, BLI), liver microsome stability (rat/human), CYP450 inhibition, and hERG channel inhibition.
  • Cellular: Antiviral activity against Wild-type, Omicron BA.5, and XBB.1 variants in Vero E6 cells; cytotoxicity in HEK293T cells.
  • In Vivo: Pharmacokinetic (PK) profiling in mice and efficacy studies in K18-hACE2 transgenic mice infected with SARS-CoV-2 EG.5.

3. Key Contributions

• AI-Enabled Scaffold Hopping: Successfully used a topology-constrained generative model to identify a novel quinoline-based scaffold that replaced the unstable DHAN scaffold of previous leads.
• Overcoming Drug Resistance: Identified a lead compound (GZNL-2016) that retains significant potency against the clinically relevant E167K mutation, a mutation that renders existing inhibitors (Jun12682, PF-07957472) ineffective.
• Mechanistic Insight: Resolved co-crystal structures revealing that the quinoline ring engages in $\pi$-$\pi$ stacking with Y268, while a specific water-mediated hydrogen bond network involving D164, E167, and the inhibitor contributes to the superior stability and potency of the new series.
• Lead Optimization: Systematically optimized the tail group to maximize the salt bridge with E167, leading to a compound with nanomolar potency.

4. Key Results

• Discovery of GZNL-2016 (Compound 16):
  • Enzymatic Potency: IC50 = 10.5 nM against wild-type PLpro.
  • Resistance Profile: Retained activity against E167K (IC50 = 480.2 nM; Ki = 439.3 nM). In contrast, reference inhibitors Jun12682 and PF-07957472 showed Ki values >10 $\mu$M against this mutant.
  • Binding Affinity: Kd values of 68.6 nM (ITC) and 78.9 nM (BLI) for wild-type PLpro.
• Metabolic Stability:
  • GZNL-2016 showed a dramatic improvement in liver stability compared to the previous lead GZNL-P4.
  • Rat liver microsome half-life ($t_{1/2}$): 120.12 min (vs. 7.95 min for GZNL-P4).
  • Low intrinsic clearance (CLint) and favorable human liver microsome stability.
• Antiviral Activity:
  • EC50: 0.49 $\mu$M (WT), 0.62 $\mu$M (BA.5), and 0.62 $\mu$M (XBB.1).
  • Selectivity: High safety window (CC50/EC50 > 379-fold).
  • PK Properties: Oral administration in mice (20 mg/kg) resulted in an AUC of 1862 h·ng/mL and rapid absorption ($T_{max}$ = 0.67 h).
• In Vivo Efficacy:
  • In a mouse model infected with SARS-CoV-2 EG.5, oral administration of GZNL-2016 (25 mg/kg) significantly reduced pulmonary viral titers at 3 days post-infection, comparable to the positive control PF-07321332 (Nirmatrelvir) at 500 mg/kg.
• Safety Profile: Low inhibition of CYP450 isoforms and hERG channels at tested concentrations.

5. Significance

• Addressing Resistance: This study provides a critical solution to the growing threat of PLpro-mediated drug resistance, specifically targeting the E167K mutation which has rendered current clinical candidates obsolete.
• AI in Drug Discovery: It demonstrates the practical utility of topology-constrained generative models (Tree-Invent) for "scaffold hopping," successfully navigating chemical space to find novel, active chemotypes that traditional methods might miss.
• Therapeutic Potential: GZNL-2016 represents a highly promising lead compound with a balanced profile of high potency, oral bioavailability, metabolic stability, and safety, suitable for further development as a next-generation antiviral for COVID-19 and future coronavirus outbreaks.
• Structural Biology: The elucidation of the binding mode, particularly the water-mediated interaction network and the specific orientation of the quinoline scaffold, offers a blueprint for designing future inhibitors against conserved viral proteases.