On September 14, 2026, MoleculeMind announced via PR Newswire that research on its QuantaMind platform was published in Science Advances, demonstrating density-functional-theory (DFT)-level reactive simulations of complete enzyme catalytic cycles. The headline result: AI can «see» proton transfers, bond breaking and formation, and an entire catalytic cycle at atomic scale in large biomolecular systems—not merely predict static structures.
QuantaMind is presented as a transition-state-centered machine-learning force field (MLFF), trained on non-equilibrium conformations and reaction data, plus DFT method-classification embeddings to unify data from different functionals and basis sets. According to the company, a single time step of a 100,000-atom reactive system can run in ~0.25 seconds—performance they position as state of the art among leading MLFFs.
In the scientific work, a complete PETase cycle—key to PET plastic degradation—is described, including acylation and deacylation, with reported barriers on the order of ~22 and ~20 kcal/mol respectively, in a system of tens of thousands of atoms (technical coverage cited ~17,792 atoms). Those estimates carry uncertainty and broadly align with prior QM/MM work on concerted chemistry and moving-histidine proton transfer.
Beyond PETase, MoleculeMind highlights industrial applications in enzyme engineering and drug discovery. In a project on pH-sensitive antibodies with extended half-life, QuantaMind was used alongside other design models; one candidate showed a dissociation rate at pH 6.0 roughly 62 times faster than at pH 7.4 in experimental tests. The message is to move from «which molecule to design» to «how and why it reacts.»
Xu Jinbo, MoleculeMind’s founder, framed QuantaMind as infrastructure connecting molecular generation, mechanistic simulation, and experimental validation. The thesis: AI-powered molecular R&D has focused on appearance and design; process traceability is missing. By making mechanisms analyzable, the company seeks a more verifiable paradigm—especially when generators propose thousands of candidates that later fail in the lab.
The competitive context includes AlphaFold and waves of structure models, plus non-reactive dynamics platforms. Classical dynamics poorly capture bond breaking/formation; pure DFT does not scale to 10⁴–10⁵ atoms over nanoseconds. QM/MM requires delimiting reaction regions. Reactive MLFFs aim for the sweet spot: near-quantum accuracy with dynamics throughput. QuantaMind joins that race with an explicit emphasis on complete enzyme cycles.
For the plastics industry and circular economy, a PETase better understood at energy-barrier level can guide rational mutations instead of blind screens. That does not guarantee commercial enzymes tomorrow, but it shrinks search space. For pH-dependent antibodies—relevant to tumor release or endosomal recycling—understanding dissociation mechanism helps prioritize designs before wet lab.
Honest limits: Science Advances publication raises credibility, but barrier numbers carry uncertainty; 0.25 s/step benchmarks need independent reproduction; and the jump from simulation to industrial mutant still requires synthetic chemistry and assays. Much public communication also comes from MoleculeMind itself and press notes: the community will need to audit data and code when available.
The company-cited DOI (10.1126/sciadv.aeg3595) and related preprint work on cinematic substrate-to-product simulations with chemical accuracy reinforce that the effort had trajectory before 2026. September’s novelty is the peer-reviewed Science Advances stamp and industrial-platform framing, not only an academic paper.
For pharmaceutical labs, the promise is mechanistic debugging: when a design fails, was it a transient pocket, a protonation barrier, or bond instability? If QuantaMind delivers interpretable trajectories, it can shorten DMTA (design-make-test-analyze) cycles. If it fails to generalize beyond shown systems, it will remain an impressive but narrow demo.
On the 2026 «AI for science» map, QuantaMind aligns with physics-in-the-loop trends: fewer LLMs that merely «talk about molecules,» more learned simulators anchored to DFT. That direction coheres with other pieces this week—MutexaGPT in enzyme engineering, TranscriptFormer in cells—that seek to close the gap between generation and verification.
In sum, MoleculeMind published in Science Advances (14 September 2026) QuantaMind: a reactive MLFF simulating complete enzyme cycles (PETase) at DFT-like accuracy and industrial scale, with a pH-antibody example (~62× faster dissociation at pH 6 vs 7.4). The shift is from static prediction to process understanding. Neural Pulse will follow independent reproductions and measurable clinical/industrial cases.
The Science Advances publication also raises openness expectations: MLFF training datasets, simulation configs, and example trajectories should be auditable. Without that, the community will treat barrier numbers and the 0.25 s benchmark as vendor claims. MoleculeMind has incentives to share enough to attract pharma partners without giving away the whole crown. That balance will define how much «new paradigm» there really is.
In clinical and materials practice, adoption cycles will be long: even with clear mechanisms, making, formulating, and approving remain the bottleneck. QuantaMind accelerates understanding; it does not repeal experimental chemistry. The best near-term use is prioritizing which mutants or ligands deserve the next expensive assay—not declaring victory over traditional scientific method.
Independent researchers can start by reproducing the PETase cycle on reduced systems and comparing barriers with classic QM/MM. That partial reproduction would already build confidence before accepting 100,000-atom claims at 0.25 s. Science advances when platform marketing submits to replication.
