FLock.io has had two research papers about training generative models of protein on molecular dynamics accepted at the world’s premier research conference in data mining, IEEE ICDM 2026. The research was done in collaboration with researchers from the University of Oxford and Korea University. It will be held November 12-15 in Shenyang, China.
Understanding the way proteins move in the human body is pivotal to drug discovery and molecular design. For example, capturing how an enzyme opens up to accept a medicine can lead to a breakthrough, but is extremely rare and difficult. Capturing these movements using traditional computer simulations takes enormous computing power and can take months.
These two papers 1) present an AI system that predicts the full range of movements a protein naturally makes without confusing fast vs. slow motions, and 2) an AI video generator for molecular changes that speeds up simulation by over a million times. Together, this makes simulating biological processes faster, cheaper and more accurate, allowing scientists to test how proteins behave and design better targeted treatments in a fraction of the time.
The International Conference on Data Mining (ICDM) provides an international forum for sharing original research results, as well as for exchanging and disseminating innovative and practical development experiences. It covers aspects including algorithms, software, systems and applications. It draws researchers, developers and practitioners from a range of data mining-related areas, such as big data, deep learning, pattern recognition, statistical and machine learning, databases, data warehousing, data visualisation and knowledge-based systems.
The papers were authored by FLock.io’s Chief AI Scientist Zehua Cheng, Wei Dai and FLock.io CEO Jiahao Sun. Competition was fierce. The ICDM received almost 2,000 submissions to its Research and Applied tracks and accepted 19.9%, each one getting triple-reviewed.
- “Timescale-Disentangled Generative Models of Protein Dynamics (TDEG)”
- “Transition Path Diffusion for Protein Reactive Trajectories (TPD)”
“Timescale-Disentangled Generative Models of Protein Dynamics (TDEG)”
The first paper tackles the trade-off faced by molecular dynamics: we have abundant short trajectories capturing fast local motions, but rare and expensive long trajectories capturing slow, metastable transitions. Existing generative models treat conformations as static distributions, blending these distinct timescales together.
TDEG factorises the model’s latent space along the physical timescale hierarchy:
- Fast subspace: An SE(3)-invariant local graph network (8 Å cutoff) pre-trained on the ATLAS archive captures transferable, residue-local fluctuations.
- Slow subspace: A global Graph Transformer trained on D. E. Shaw long trajectories uses a TICA-inspired loss to encode slow, metastable modes.
- Hierarchical decoder: Composes coarse backbone dihedrals with a fine sidechain refiner to generate complete heavy-atom structures.
The benefit is firstly higher realism. TDEG separates fast local wiggles from slow major shape changes. This results in much more physically accurate 3D protein structures with fewer structural errors (like overlapping atoms). Secondly, it works on unseen proteins. Because local wiggles follow universal rules across many proteins, TDEG can accurately predict movements for brand-new proteins it has never seen before.
“Transition Path Diffusion for Protein Reactive Trajectories (TPD)”
The second paper addresses how rare reactive events, such as domain closures or protein folding, occupy a tiny fraction of equilibrium dynamics. Traditional rare-event methods like Transition Path Sampling (TPS) require weeks or months of wall-clock simulation time per system.
TPD models entire 64-frame transition paths simultaneously as single diffusion objects:
- Trajectory denoiser architecture: Couples intra-frame spatial GNNs, inter-frame temporal multi-head self-attention, and geometrically grounded coordinate updates.
- Composite physics loss: Integrates an RMSD continuity bound, a coarse-grained energy surrogate, and a differentiable committor surrogate with soft monotonicity regularization.
- Adaptive arc-length reparameterisation: Concentrates generated trajectory frames near critical energy barriers.
The benefit is firstly that instead of predicting a move frame-by-frame, TPD generates the entire transition path all at once using diffusion AI (similar to how modern AI image generators work, but for 3D trajectory frames). Secondly, a simulation that used to take months of heavy computing can now be generated accurately in about 45 seconds.
We look forward to presenting both our papers in Shenyang in November!
More about FLock.io
FLock.io is an AI research and infrastructure company pioneering enterprise-grade federated learning and distributed AI solutions. Its decentralised federated learning architecture and production-ready platforms (AI Arena, FL Alliance, and FLock API Platform) enable organisations to train and deploy their own custom AI models on local hardware while maintaining full data privacy, model ownership, and regulatory alignment by design.
FLock.io is internationally recognised for its academic research, including the NeurIPS award-winning paper “FLock: Defending Malicious Behaviors in Federated Learning with Blockchain” and sponsors computer science PhD students at the University of Oxford.
