BioKinema

BioKinema is a generative model for predicting all-atom biomolecular dynamics, introduced in a February 2026 bioRxiv preprint from the International Digital Economy Academy (IDEA) in Shenzhen. Where conventional molecular dynamics (MD) simulations resolve kinetic pathways one tiny time step at a time at enormous computational cost, BioKinema learns to generate continuous-time, all-atom trajectories directly, aiming to reproduce the kinetics of conformational transitions at a fraction of that cost.

Deep learning has already transformed static structure prediction and equilibrium ensemble sampling, but simulating how biomolecules move between states over time — the kinetics — has remained a harder, less solved problem. BioKinema targets this gap with a scalable diffusion architecture that uses temporal attention mechanisms derived from Langevin dynamics, grounding the generative process in the physics of molecular motion rather than treating it as an unconstrained sequence-generation task.

A central challenge in generating long trajectories is error accumulation, where small per-step mistakes compound over many steps. BioKinema addresses this with a hierarchical forecasting-and-interpolation strategy, enabling longer-horizon generation while keeping trajectories physically stable. The model is positioned as a complement and potential alternative to MD for exploring kinetic landscapes in structural biology and drug discovery.

#Key Features

• Continuous-time all-atom generation: Produces all-atom trajectories in continuous time, capturing dynamics rather than only static structures or equilibrium ensembles.
• Physics-grounded diffusion: Uses a scalable diffusion architecture with temporal attention derived from Langevin dynamics, anchoring generation in molecular-motion physics.
• Hierarchical forecasting and interpolation: Mitigates error accumulation over long horizons, supporting stable extended-trajectory generation.
• Conformational and allosteric transitions: Captures functionally relevant motions, including induced-fit changes and allosteric responses in protein-ligand complexes.
• Rare-event prediction: Leverages enhanced-sampling data to predict rare kinetic events such as ligand unbinding pathways.

#Technical Details

BioKinema is a diffusion-based generative model employing temporal attention mechanisms derived from Langevin dynamics, paired with a hierarchical forecasting-and-interpolation strategy to combat the error accumulation that typically degrades long-horizon trajectory generation. The authors report that BioKinema produces physically stable and dynamically accurate trajectories suitable for downstream analysis, captures key conformational transitions tied to protein function, and for protein-ligand complexes elucidates mechanisms including induced-fit conformational changes and allosteric responses. Trained or conditioned with enhanced-sampling data, it can also estimate ligand unbinding pathways as rare kinetic events. As a recent preprint, no public code or model weights are referenced in the manuscript.

#Applications

BioKinema is intended for structural biology and drug discovery workflows that need dynamic, not just static, structural information. By generating trajectories far faster than conventional MD, it could support high-throughput exploration of conformational and kinetic landscapes, mechanistic studies of allostery and induced fit, and estimation of ligand binding and unbinding pathways relevant to assessing drug efficacy and residence time.

#Impact

BioKinema contributes to an emerging line of work that extends deep-learning structure prediction from static and equilibrium settings into kinetics, one of the field's harder open problems. Its physically grounded diffusion formulation and long-horizon stability strategy offer a template for scalable trajectory generation. As an unreviewed preprint without a referenced code or weight release, its accuracy relative to gold-standard MD and experiment will require independent benchmarking before broad adoption.