Risk Minimization: Upstream, Not Downstream.
Traditional discovery waits for failure to find faults. At FoldForce, we bring risk assessment to the very beginning. By integrating physical constraints and dynamic interactions early, we transform the generative process.
Efficiency
Eliminate dead-end candidates before they consume resources.
Reliability
Move from static snapshots to dynamic biological reality.
Productivity
dramatically reduce waste and accelerate the path to clinics and market.
WHAT YOU GET
Three decision-grade outputs across all phases.
Ranked candidates
Variants, ligands, partners, assemblies-prioritized for what matters.
Calibrated metrics
Stability, dynamics, binding signals-when available-tied to real conditions.
Key Differentiator
Risk Profiles
Probability of failure under stated conditions and a time horizon.
A Contract-First Pipeline
Data → Dynamics → Calibration → Decision
Phase 1
Scaffold Refiner
Ingests raw structures, performs cleaning and refinement, and produces refinement-ready scaffolds with consistent downstream-safe featurization.
Phase 2
Data Foundry
Builds deterministic, manifest-backed datasets with QC gates, recorded metadata, and reproducible splits.
Phase 3
Learn Dynamics
Trains an SE(3)-aware DynamicField model using multi-lag objectives, with rollout diagnostics to assess physical plausibility over time.
Phase 4
Stability Profiler
Combines hybrid physics (classical force fields + learned residuals) with spectral and ensemble features, calibrated to experimental stability under explicit conditions.
Three Phases, One Coherent System
Phase 1
Working Prototype
Scaffold Refiner implemented and validated
Deterministic contracts and pipeline discipline established
Phase 2
Training-Ready Stack
Manifest-backed dataset generation implemented
Dynamics training notebook ready up to model training and rollouts
Phase 3
Roadmap
Environment-conditioned Stability Profiler
Rankings, calibrated metrics, and probabilistic risk profiles
BENCHMARK RESULTS
Hard evidence that the system works - validated on real structural data.
68%
RMSD reduction
at 7 Å perturbation
73–90%
Reduction in covalent bond
strain across 3–7 Å noise
87–93%
Near-clash-free structures
at moderate noise levels
~1 Å
Global packing restored
of native compactness
Even from severe distortions, Phase 1 restores Cα–Cα bond lengths to within ~0.02–0.09 Å of the ~3.80 Å native baseline.
KEY RESULTS
Built to Be Reproducible and to Fail Loud
Deterministic datasets + recorded artifacts
QC gates (no silent deletion)
Rollout diagnostics (behavior over time)
Traceable runs for audit-ready decisions
FoldForce Refine (Structure Refiner)
Perfecting the Biological Blueprint in 0.53 Seconds
Raw AI predictions are often geometrically impossible. We relax structures into physically valid states, turning rough guesses into reliable assets-averaging 0.53 seconds per protein.
Why It Matters
Rescue Targets
clarify cryo-EM/low-res structures to reveal binding sites.
Boost Screening
dock into physically plausible pockets to reduce false negatives.
Publication Quality
convert raw predictions into experiment-grade models.
Proof Metrics (Smoke Test)
Validated on N=30 chains per noise level (3–7 Å RMSD).
Note: Median clash returns to ~0 at all noise levels; residual overlaps are localized to a minority of hard cases.
High-Speed Molecular Dynamics
MD Accuracy. ML Velocity.
Traditional MD is insightful but slow. FoldForce targets MD-level accuracy at ML speeds so you can run dynamics across every candidate-not just finalists.
Capabilities
Dynamic Docking
flexible binding poses beyond static docking.
Unlock Cryptic Pockets
transient sites revealed by motion.
Full Trajectories
association/dissociation paths for mechanism.
Why It Matters
Universal Screening
simulate the whole library.
True Biological Context
molecules behave in motion, not snapshots.
Validation (what we measure)
Thermodynamics, Kinetics, and DMPK Risk
From motion to biophysics: binding energy, on/off rates, and risk.
Don’t just predict “binds.” Predict how strong, how long, and how risky.
Capabilities
Binding free energies
(ΔG) and affinity ranking
On/off rates
(k_on / k_off) and residence time
Multi-body interactions
PPI, antibody-antigen, protein-nucleic acid
DMPK risk profiles
transporter/CYP susceptibility signals
Why It Matters
Fail earlier
fewer late-stage surprises
Protect patent life
faster convergence on viable series
Reduce waste
fewer doomed compounds enter expensive stages
The Horizon: Digital Twins
Scaling Thermodynamics to Life Itself
This is our moonshot: scale from single binding events to multi-body interactions and pathways—forming the basis of Cellular Digital Twins.
Multi-body interactions
simulate assemblies and interfaces at scale
Cellular digital twins
extrapolate from interactions to pathways and subcellular systems
Key Differentiator
Uncharted territory
find bottlenecks, design de novo enzymes, explore “undruggable” disease space
"We are not just modeling drugs; we are building the simulation layer for biology itself—extrapolating what we know to predict what we have yet to discover."
FAQ
Make risk minimization upstream, not downstream.
Screen more. Waste less. Move faster—with physics-aware reliability.
