LondonBioEmu runs 100,000 GPU-hours of protein simulation in minutes.
BioEmu samples thousands of conformations by stability, showing which states a protein visits, not just its most stable shape.
Protein-structure competition CASP16 added conformational-ensemble tracks; BioEmu teams now target how drug binding reshapes the full state landscape.
Sources: Nature
THE OLD PARADIGM
For sixty years structural biology treated proteins as having one shape — the crystal structure. AlphaFold2's 2020 breakthrough predicted that single shape with near-experimental accuracy, and the field briefly declared the protein folding problem solved.
WHY ONE SHAPE ISN'T ENOUGH
Real proteins breathe. They flex between conformations on timescales from picoseconds to seconds, and biological function lives in those transitions — an ion channel opening, a receptor binding, an enzyme catalyzing. The crystal structure captures the most stable pose; biology happens at the others.
THE SIMULATION COST
Molecular dynamics simulates atoms one femtosecond at a time. Capturing a millisecond of protein motion — the timescale of most conformational changes — requires roughly a trillion steps. A single such trajectory on conventional hardware can consume months of GPU time.
WHAT BIOEMU DOES DIFFERENTLY
Instead of stepping through time, BioEmu is a generative model trained on simulation data — it samples conformations directly from the equilibrium distribution, weighted by stability. The reader gets a population of states with their probabilities, not a single trajectory.
WHY CASP MATTERS
CASP — the Critical Assessment of Structure Prediction — has run every two years since 1994. It blinds prediction teams to experimental results during the contest, then reveals them. CASP14 in 2020 was where AlphaFold2 stunned the field; adding ensemble tracks in CASP16 signals the field has moved past static structures as the frontier.
THE DRUG DESIGN PAYOFF
Roughly a third of approved drugs target G-protein-coupled receptors, which switch between active and inactive shapes. A binder that locks the inactive state can be a blockbuster; one that locks the active state can be a poison. Predicting the full landscape — not just the dominant pose — is the difference between rational design and screening millions of compounds.