A Beijing-based AI platform has achieved something drug researchers have struggled with for decades: simulating exactly how enzymes work at the atomic level, including the moment bonds break and form. The breakthrough, published in Science Advances on September 14, addresses a fundamental problem in AI-powered drug discovery: generative models can design promising drug candidates, but they can't explain why those candidates succeed or fail when they encounter real enzymes inside the body.

What's the Gap Between Designing Drugs and Understanding How They Work?

Imagine designing a key that looks perfect on paper but doesn't actually turn the lock. That's the problem facing AI drug discovery today. Generative AI models have become remarkably good at proposing new molecules and optimizing their structures, but generating a candidate is only the first step. The real question is whether that molecule actually behaves correctly when it encounters an enzyme, solvent, and the thermal chaos of a real biological environment.

Conventional molecular dynamics simulations, the standard tool for studying how proteins move and molecules bind, treat atoms like balls connected by springs. They can show how a protein flexes but not how a proton jumps between amino acids, how a chemical bond snaps, or how a catalytic cycle completes. Quantum mechanical methods like Density Functional Theory (DFT) can handle those reactive events, but they're so computationally expensive that they're limited to simulating just dozens of atoms, far too small to capture a realistic enzyme environment.

How Does QuantaMind Solve This Problem?

The platform, developed by MoleculeMind, uses machine learning force fields (MLFFs), which are AI models trained on quantum chemical data that can run at speeds approaching classical simulation. But most existing MLFFs share a critical weakness: they're trained primarily on molecular structures at their equilibrium states, the "before" and "after" snapshots of chemistry. They haven't been trained to navigate the strained, high-energy configurations that molecules pass through as a reaction actually unfolds.

QuantaMind addresses this directly with two key innovations:

• Transition-state-centered training: The model is trained explicitly on non-equilibrium conformations and the critical high-energy configurations along reaction pathways, teaching it what chemistry looks like as it unfolds rather than only at the start and end.
• DFT method classification embeddings: Because quantum chemistry training data comes from different density functionals and basis sets, each with different accuracy characteristics, QuantaMind encodes which method produced each data point, allowing heterogeneous training data from multiple quantum chemistry protocols to be unified without sacrificing consistency.
• Massive-scale simulation capability: The platform achieves DFT-level accuracy while enabling reactive molecular dynamics simulations of systems containing 10,000 atoms at nanosecond timescales, a regime that has been inaccessible to quantum methods.

At industrial scale, MoleculeMind reports extending the platform to reactive systems with hundreds of thousands of atoms, with a single time-step simulation of a 100,000-atom reactive system completing in 0.25 seconds.

What Real-World Problem Did QuantaMind Solve First?

The research team demonstrated the platform's capabilities by simulating PETase, a plastic-degrading enzyme first isolated in 2016 from bacteria found at a PET plastic recycling facility in Japan. PETase breaks down polyethylene terephthalate, the most common thermoplastic used in food and beverage containers and textiles, and engineering improved variants for industrial plastic remediation is an active research priority.

QuantaMind simulated the complete PETase catalytic cycle in a system containing 17,792 atoms, including the full enzyme (257 amino acid residues), the substrate molecule, and nearly 5,000 water molecules, over 400 picoseconds of total simulation time across four reaction steps. The simulation captured the specific sequence of proton transfer events that had been disputed in prior studies: a proton jumping from serine residue S165 to histidine H242 during the acylation step, the subsequent transfer of that proton to the product, a water molecule entering the active site and attacking during deacylation, and the final restoration of the catalytic triad.

The estimated free energy barriers, approximately 22 kilocalories per mole for acylation and 20 kilocalories per mole for deacylation, are within approximately 3 kilocalories per mole of the experimental value and close to prior reference calculations. This level of accuracy is significant because it means the simulation is capturing real chemistry, not just mathematical approximations.

"AI-powered molecular R&D has largely focused on what molecules look like and what we can design. QuantaMind goes further by addressing how molecules move, how they react, and why these processes influence outcomes," stated Jinbo Xu, founder of MoleculeMind.

Jinbo Xu, Founder of MoleculeMind and Tenured Full Professor at the Toyota Technological Institute at Chicago

How to Evaluate AI Drug Discovery Tools for Your Research

• Verify transition-state accuracy: Ask whether the tool has been trained on high-energy configurations along reaction pathways, not just equilibrium structures. This determines whether it can predict what actually happens during a chemical reaction.
• Check system size limitations: Confirm the maximum number of atoms the platform can simulate. Tools limited to dozens of atoms cannot capture realistic enzyme environments; systems with 10,000 or more atoms are more likely to produce biologically relevant results.
• Assess independent validation: Look for peer-reviewed publications in high-impact journals that demonstrate the tool's accuracy against experimental data. The TEA Challenge 2023, a rigorous blind evaluation of leading machine learning force fields, found that model performance depends strongly on training data quality rather than architecture choice.
• Understand computational speed: Evaluate whether the tool can complete simulations in reasonable timeframes. A platform that can simulate a 100,000-atom system in 0.25 seconds per time-step is significantly faster than quantum mechanical methods that would take weeks or months.

Why Does This Matter Beyond Plastic-Eating Enzymes?

The implications extend far beyond plastic remediation. The same capability to simulate enzyme catalysis at quantum accuracy applies to pharmaceutical research, where understanding why a drug candidate fails inside an enzyme is as valuable as knowing it should work. It applies to biology, where researchers need to understand how natural enzymes function. And it applies to environmental science, where engineered enzymes might address pollution or industrial processes.

This breakthrough also reflects a broader shift in how AI is being applied to molecular science. Earlier this week, DeepMind published AlphaGenome Atlas, which contains predictions for the effects of around 9 billion possible single-letter changes in human DNA. Testing 9 billion mutations individually in a laboratory is practically impossible, but AI can help researchers rank which variations deserve attention, including in work on rare diseases and understanding genetic changes associated with disease.

The research team candidly identified current limitations: the free energy estimates carry large error bars stemming from simulation time constraints, and more training data specifically targeting transition states in serine hydrolases would improve accuracy. But the fact that this is the first complete enzyme simulation of its kind enabled by a deep learning-based machine learning force field represents a significant step forward in closing the gap between what AI can design and what AI can understand about how those designs actually work in biology.

" }