One GPU Card to Film an Atomic Movie: Molecular Heart's QuantaMind Pushes Reactive Molecular Dynamics to Scale

A molecular structure diagram looks like a still photo. Real molecules are not still. In solution they constantly collide and rotate, chemical bonds break and reform, and protons hop between atoms.

A single bond-breaking event happens on the femtosecond scale — one quadrillionth of a second — while a complete enzyme-catalyzed reaction can take milliseconds. To see how atoms move at every step in between, a simulation must cover roughly twelve orders of magnitude in time. Two further barriers stand in the way: accuracy approaching first-principles calculations, and scale large enough to hold a biological system. QuantaMind, from Shanghai-based Molecular Heart (Molecular Heart), recently published a set of numbers in *Science Advances* attempting to push all three forward simultaneously.

The Trade-Off That Could Not Be Won

The difficulty of computational chemistry is a three-way trade-off:

• Accuracy: a quantum-mechanical description, because bond breaking and forming is fundamentally electron rearrangement — classical force fields model bonds as springs and cannot capture this.
• Speed: long enough trajectories to see a full reaction pathway.
• Scale: enough atoms that a system can contain a patch of cell membrane or a complete catalyst active site — on the order of 100,000 atoms, not a few thousand.

For decades, reactive molecular dynamics has only been practical for small systems over short time windows. The scenarios that actually need it — how a drug molecule completes a chemical transformation in an enzyme's active site, or how a proton hops along a hydrogen-bonded chain — have mostly fallen outside its reach.

Three Numbers

The core figures from the paper:

• Inference speed of 4.2 × 10⁻⁶ seconds per atom per step on a single NVIDIA A100 80GB GPU, which the team says matches current state-of-the-art machine learning force fields. Extended tests pushed this to 2.1 × 10⁻⁶ s/atom/step.
• Converted, this means a ~100,000-atom reactive system takes about 0.25 seconds per time step on one card.

One caveat: the number of time steps depends on the timestep size, which the company has not disclosed. If one assumes the 1-femtosecond timestep common in classical molecular dynamics, running 100 nanoseconds would require 10⁸ steps — over half a year on a single card. This inference does not contradict the original claim; it simply means the wall-clock time behind the "hundred-nanosecond" figure cannot be derived from public information.

The independently benchmarked part is the protein system: the paper reports QuantaMind validated efficiency on a complete 17,792-atom protein system — an order of magnitude below 100,000 atoms, but a real biomolecule rather than a cut-out fragment.

| Metric | Paper | Extended tests | |---|---|---| | Inference speed | 4.2×10⁻⁶ s/atom/step | 2.1×10⁻⁶ s/atom/step | | Validated system | 17,792-atom complete protein | ~100,000-atom reactive system | | Time scale | Tens of nanoseconds | Hundreds of nanoseconds | | Per-step cost | Not disclosed | ~0.25 s (100k atoms) | | Hardware | Single NVIDIA A100 80GB | Single A100 80GB |

What "Reactive" Costs

Machine learning force fields have advanced rapidly — they can fit spectra, compute energies, and run large systems. Their shared weakness is that they are non-reactive: connectivity between atoms is fixed in the topology, so once a bond must break or form, the model has no description for it.

QuantaMind is positioned as a reactive machine learning force field. It can track proton transfer, chemical bond formation and breakage, and the emergence of reaction intermediates, spanning small molecules, drug-like molecules, and complex biological systems like proteins and enzymes.

This distinction matters practically in drug discovery. How atoms move after a candidate binds a target, why two visually similar molecules show different binding and reaction behavior, how a single proton transfer or bond breakage changes the outcome — these are essentially black boxes in non-reactive models. Tracking mechanism means researchers can see the "why" computationally, not just receive a "can/cannot" score.

Rendering Trajectories into Watchable Films

The real usability test is visualization. In one demonstration, a catalytic reaction — hydrogen molecules adsorbing onto a catalyst surface, breaking into two hydrogen atoms, and desorbing as water — was rendered as an animation with femtosecond-level time resolution.

For a computational chemist, the value lies in verification cost, with nice animations as a byproduct. Judging whether a reaction path is reasonable used to require frame-by-frame inspection of coordinates and energies; now it can be watched directly, making anomalous frames easier to spot. A side benefit: non-computational experimentalists can understand trajectories too, a real communication saving in interdisciplinary teams.

Others on the Same Track

Training ML models on large-scale quantum chemistry data, then replacing expensive first-principles calculations with inference — nearly everyone on this route is a giant: Meta's UMA, Microsoft's MatterSim, and Google DeepMind's GEMS. The differences lie in reactivity coverage and system types.

Molecular Heart founder Xu Jinbo developed the world's first effective AI protein-structure-prediction algorithm in 2016, later widely credited as inspiring the first generation of AlphaFold. The company was founded in 2022, completed Series A financing exceeding $100 million cumulative in June 2026, and opened its bio-economy operating system MoleculeOS in July.

QuantaMind fills the last link in the matrix: MoleculeOS previously covered "generate–design–predict" (what a molecule looks like, what it should be designed as); QuantaMind adds "mechanistic understanding" — why a molecule behaves the way it does.

Four Things Not Yet Verified

1. No independent reproduction yet. 2. "Near-DFT accuracy" needs unpacking. ML force fields usually perform well within their training distribution; how fast errors grow out-of-distribution is the key question for real R&D use. The paper gives system size and timing but no public error distribution. 3. Do hundreds of nanoseconds and 100,000 atoms hold simultaneously? 4. Real cellular environments remain orders of magnitude away. A cell contains billions of atoms; 100,000 atoms is roughly one membrane patch or one active site. Each additional order of magnitude raises the training-data coverage difficulty nonlinearly.

What Happens When the Barrier Drops

The ultimate scenario is concrete: an ordinary lab that once queued for supercomputing time and needed a computational-chemistry postdoc could, if a single A100 delivers usable 100,000-atom results, do this work on a workstation.

What changes is not one team's capability ceiling, but who is qualified to ask the question. Questions once reserved for well-resourced institutions become accessible to small teams — a shift that has recurred throughout the history of science, with different consequences each time.

So when simulation moves from "predicting outcomes" to "reconstructing processes," will researchers spend more time designing experiments — or verifying the mechanisms AI hands them?

Sources

1. Molecular Heart QuantaMind research, *Science Advances*, 2026-09 — Science and Technology Daily: https://www.stdaily.com/web/gdxw/2026-09/17/content_583085.html 2. Science and Technology Daily, 2026-09-16: https://www.stdaily.com/web/gdxw/2026-09/16/content_582193.html 3. QbitAI report, 2026-09-15, via BAAI Link: https://link.baai.ac.cn/@liangzw/117280329474008819 4. Sohu Tech, 2026-09: https://www.sohu.com/a/1076656175_121421892 5. 6th "Haiju Yingcai" Global Innovation and Entrepreneurship Competition AI4S track coverage, 2026-09-17: https://news.qq.com/rain/a/20260917A05Q9N00