G GENESIS / THE DECISION ENGINE

How the Math Chooses

BindCraft generates hundreds of peptide designs. Simulation and docking evidence guide selection.

A mathematical decision engine combines three evidence streams to reveal three illustrative candidates from a previously identified shortlist. Hundreds of peptides are first generated by BindCraft; their design metrics, GROMACS simulations and AutoDock Vina docking results feed the decision engine. Their evidence flows into a Bayesian decision engine. Press Start the engine to animate three named candidates emerging. All three displayed candidates use uploaded molecular structures. BINDCRAFT-GENERATED PEPTIDES THREE SOURCES OF EVIDENCE COMBINE EVIDENCE WITH MATH THREE ILLUSTRATIVE PEPTIDES 30 examples from hundreds designed Designed by BindCraft Hundreds of peptide candidates Time and resources are limited BindCraftDesign + structural confidenceBi GROMACSDynamics + stability over timeMi AutoDock VinaLipid A docking + pocket accessVi MATHEMATICAL ENGINE Bayesian decision model ∑ P(success | evidence) + learning value + diversity Run selection Promising. Informative. Diverse. A few candidates for wet-lab study THE OUTPUT PILE Press Run to release peptides Overlapping cards · click to inspect Illustrative selection from a wider shortlist
1
Gather evidence

Design, dynamics and docking contribute different signals.

2
Make a mathematical decision

Balance predicted success with what an experiment can teach us.

3
Choose a small batch

Three example candidates emerge and form a physical stack.

Press Start the engine to play the animated selection sequence.0:00 / 0:18
CONCEPTUAL MODEL · NOT EXECUTED BY THIS PAGE

The decision is more than a docking score.

We use computational evidence to estimate which experiments are promising and informative. The animation replays our existing shortlist; it does not fit a model, import score tables, or run the three scientific tools.

01

Represent each candidate with evidence

xi = (Bi, Mi, Vi)

B: BindCraft design and confidence features. M: GROMACS dynamics features. V: AutoDock Vina docking-derived features. These are separate evidence streams, not an assertion that all tools run independently or in this pictured order.

02

Average over uncertainty

pi = ∑m P(yi=1 | xi, D, m) P(m | D)

D denotes observed outcomes and m a candidate hypothesis. Bayesian model averaging combines predictions across hypotheses rather than treating one explanation as certain.

03

Value informative experiments

Ii = H(Eθ[pi,θ]) − Eθ[H(pi,θ)]

This BALD-style information gain measures disagreement across posterior predictions. H is binary entropy; θ is a posterior model draw. It is not just a reward for noisy or ambiguous outcomes. [4]

04

Select under a resource constraint

Ai = wppi + wIIi + wddi

A schematic acquisition score balances predicted success, information and diversity relative to the chosen batch. The animation illustrates three outputs, not a prescribed experimental batch size. This explanatory formula is not a substitute for the project's full scoring implementation; this page supplies no fitted weights or numerical acquisition scores.

Interpretation. Predicted binding confidence, molecular stability, and docking geometry do not by themselves establish inhibition of TLR4 signaling. The modeled receptor composites are structural visualizations, not experimentally resolved complexes. The three outputs are illustrative members of an earlier shortlist, not three newly calculated winners or experimentally proven successes.

Literature and visualization references

  1. Pacesa M et al. (2025). One-shot design of functional protein binders with BindCraft. Nature 646, 483–492. doi:10.1038/s41586-025-09429-6.
  2. Abraham MJ et al. (2015). GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers. SoftwareX 1–2, 19–25. doi:10.1016/j.softx.2015.06.001.
  3. Eberhardt J, Santos-Martins D, Tillack AF, Forli S (2021). AutoDock Vina 1.2.0: New Docking Methods, Expanded Force Field, and Python Bindings. J. Chem. Inf. Model. 61, 3891–3898. doi:10.1021/acs.jcim.1c00203.
  4. Houlsby N, Huszár F, Ghahramani Z, Lengyel M (2011). Bayesian Active Learning for Classification and Preference Learning. arXiv:1112.5745.
  5. Park BS et al. (2009). The structural basis of lipopolysaccharide recognition by the TLR4–MD-2 complex. Nature 458, 1191–1195. doi:10.1038/nature07830. Receptor: PDB 3FXI.
  6. Rose AS et al. (2018). NGL viewer: web-based molecular graphics for large complexes. Bioinformatics 34, 3755–3758. doi:10.1093/bioinformatics/bty419.
  7. Kabsch W (1976). A solution for the best rotation to relate two sets of vectors. Acta Crystallographica A 32, 922–923. doi:10.1107/S0567739476001873.