Every branch answers its question at five levels. The levels are not “better and better” — they are a controlled increase in cost and physical content, run so that the shape of the improvement becomes visible. The interesting result of this repo is where each curve flattens.
L0 — null model. Deliberately embarrassing. Costs seconds, uses no simulation, and must be beaten by everything above it. A surprising fraction of published methods fail to beat their own null; that is the reason L0 is mandatory.
L1 — single structure, no sampling. One conformation, one energy evaluation or one geometric analysis. Docking scores, static electrostatics, normal modes. Cheap, interpretable, blind to everything dynamic.
L2 — short MD plus an endpoint estimator. Sampling enters, rigour does not. MM/GBSA, RMSF, contact maps, near-attack-conformer counting. Known to be badly calibrated for absolute numbers and often adequate for ranking — quantify that here.
L3 — rigorous classical free energy. Alchemical RBFE/ABFE with proper λ schedules and MBAR; metadynamics or umbrella sampling with convergence diagnostics; MSMs built from many trajectories. The expensive tier, and the one that should actually work.
L4 — beyond the classical force field. ML/MM or QM/MM for the region where the force field is wrong; ML-potential-corrected free energies. Expect this tier to be thin for a long time, and expect it to buy less than it costs for most systems.
| Branch | L0 | L1 | L2 | L3 | L4 |
|---|---|---|---|---|---|
| A binding affinity | ligand heavy-atom count | docking score | MM/GBSA on short MD | ABFE / RBFE with MBAR | ML/MM correction on the ligand |
| B conformational change | anisotropic network model | normal modes from one structure | RMSF, contacts, PCA on short MD | long or parallel MD, reversible transitions | MACE-OFF on a small peptide |
| C distributions | Boltzmann over crystal endpoints | static CV values | histogram from unbiased MD | metadynamics / umbrella + MSM | ML-potential FES for alanine dipeptide |
| D enzymatic properties | kcat of nearest characterised homolog | active-site geometry | NAC populations, field at the bond | PMF along the reaction coordinate | QM/MM barrier with xTB |
A level is not complete without all four:
results/<branch>.parquet matching the schema in CLAUDE.md.Rank correlation (Spearman) and pairwise ranking accuracy carry more weight than RMSE. Most of the real use of these methods is “is A better than B”, and a method with a 2 kcal/mol systematic offset but correct ordering is more useful than the reverse. Report all three, and report the null’s score beside every level.
For each branch, plot accuracy against log wall-clock across L0–L4. Name the level at which the curve flattens, and write it down in the branch’s intent file. That sentence — “for this system class, rigour above Ln did not pay” — is the product of this repo. A flat step is a finding, not a failure.