Documentation

Everything you need to use a kernel

What the descriptors mean, what a σ-profile input looks like, and how to reproduce every number on this site from first principles.

The 9-descriptor feature vector

Every kernel draws from the same 9-dimensional vector, computed from one DFT SCF per molecule. No empirical additions — each descriptor is a physical quantity read off the σ-profile / COSMO surface.

cavity — cavity_area_aa2 — solvent-accessible surface area of the COSMO cavity (Å²)

g_polar — g_polar_kcal — polar screening (electrostatic solvation) energy (kcal/mol)

dipole — dipole_debye — molecular dipole moment (Debye)

sigma_variance — sigma_variance — spread of the σ-profile (polarity dispersion)

sigma_skewness — sigma_skewness — donor/acceptor asymmetry of the σ-profile

sigma_kurtosis — sigma_kurtosis — peakedness of the σ-profile

donor — hb_donor_mass — hydrogen-bond donor surface area (σ-degree classified)

acceptor — hb_acceptor_mass — hydrogen-bond acceptor surface area (σ-degree classified)

n_segments — n_segments — number of COSMO surface tessellation segments

Input format: the σ-profile (.mfsig)

Kernels take mfsig σ-profiles as input. A .mfsig is a signed JSON record carrying the σ-profile, the derived descriptors, and full provenance (recipe + vendor SHAs). Generate them at mfsig.com from a SMILES, or upload your own set.

{
  "provenance": { "recipe_sha": "16d3ca90", "vendor_sha": "56c80cea", ... },
  "chemistry_and_geometry": { "smiles": "...", "inchi_key_14": "..." },
  "cavity_and_sigma": { "cavity_area_aa2": 0.0, "sigma_profile": [ ... ] },
  "moments": { "sigma_variance": 0.0, "sigma_skewness": 0.0, ... },
  "hb_proxies": { "hb_donor_mass": 0.0, "hb_acceptor_mass": 0.0 },
  "energies": { "g_polar_kcal": 0.0 }
}

Reproducibility

Every kernel is reproducible from raw data. The fit is a closed-form Cholesky solve; the atlas regenerates for ~$10–21 of DFT compute in under two days. The rigor engine, the fitting engine, and the retrobench scripts are all versioned internally.

Fitting engine

Closed-form JAX ridge over the ¼-integer exponent grid; per-fold pair selection in nested CV.

Rigor verdict engine

Runs the 8-step gate and assigns A.IMPROVE / B.UNDERFIT with the 25% scaffold-blind lift.

Retrobench

Re-fits a kernel with a fixed form against a new/corrected atlas (Pass A) and with a free gate (Pass B).

Atlas

6,876 v0.91.1 σ-profiles, recipe SHA 16d3ca90, vendor SHA 56c80cea.

Next: browse the kernel catalogue, read the methodology, or upload your own dataset.