← Về thư mục
📄 / / profiles / r-and-d / raw_evidence / OpenAI4S / skills / bioskills / bio-structural-biology-structure-validation / references / usage-guide.md

Structure Validation - Usage Guide

Overview

This skill decides whether a macromolecular model, or a specific region of it, is reliable enough to build on before any downstream measurement, docking, or mechanistic claim. It treats a coordinate file as an interpreted model whose reliability varies per-atom, so it reads the local signals (per-residue B-factor, real-space fit) rather than trusting the single headline resolution. It covers experimental validation (resolution, R-work vs R-free and the overfitting gap, B-factor sanity, MolProbity clashscore/Ramachandran/rotamer outliers, cis non-proline peptides, the wwPDB report and its percentile sliders) and predicted-model validation (pLDDT bands, PAE inter-domain confidence, and trimming with phenix.process_predicted_model before molecular replacement or docking). It also handles the cryo-EM global-vs-local resolution distinction (FSC 0.143 half-map vs 0.5 map-model) and NMR ensemble spread.

Prerequisites

pip install biopython numpy

MolProbity, phenix (phenix.molprobity, phenix.process_predicted_model), and DSSP (mkdssp) are separate CLI installs, not pip packages; install them independently and confirm they are on PATH. The MolProbity web service (http://molprobity.biochem.duke.edu) is an alternative to a local phenix install.

Quick Start

Tell the AI agent what you want to decide: - "Is this 2.8 Angstrom structure good enough to trust the active-site geometry?" - "Read the resolution and R-free from this mmCIF and tell me if it is overfit" - "Flag the Ramachandran and rotamer outliers and any cis non-proline peptides" - "Which residues in this chain have suspiciously high B-factors?" - "Validate this AlphaFold model before I dock a ligand into it" - "Should I trust the domain arrangement in this predicted structure?" - "Is the resolution uniform across this cryo-EM map or is the periphery a guess?" - "Report the per-residue spread of this NMR ensemble instead of using model 1"

Example Prompts

Judging a deposited experimental structure

"I want to measure a catalytic distance in this 3.1 Angstrom crystal structure. Read the resolution, R-work, R-free and the overfitting gap, then check whether the active-site residues have high local B-factors before I trust the measurement."

"Run the geometry validation on this model and tell me the clashscore, Ramachandran favored percentage and outliers, and poor-rotamer percentage. Is this worse than typical for its resolution?"

"Pull the wwPDB validation report for this PDB ID and tell me whether my region of interest is flagged as an RSRZ outlier."

Validating a predicted model

"Before I use this AlphaFold model for molecular replacement, trim the low-pLDDT residues, convert pLDDT to a pseudo-B, and split it into PAE-defined domains."

"This predicted structure looks confident everywhere. Read the PAE matrix and tell me whether the two domains have a trustworthy relative orientation or are independently placed."

"A long stretch of this model is low pLDDT. Is that a modeling failure or an intrinsically disordered region I should keep?"

Cryo-EM and NMR

"This is a 2.6 Angstrom cryo-EM structure of a large complex. Is that resolution uniform, or is the peripheral arm at much lower local resolution and effectively a docked hypothesis?"

"Compute the per-residue Ca RMSD across all models of this NMR ensemble and show me which regions are well-restrained versus flexible. Do not average the coordinates."

What the Agent Will Do

  1. Identify the experimental method (_exptl.method) to route to the correct validation logic (X-ray, cryo-EM, NMR, or predicted).
  2. Read resolution, R-work, and R-free from the mmCIF header via MMCIF2Dict and compute the R-free-minus-R-work overfitting gap against a resolution-scaled expectation.
  3. Sanity-check per-residue B-factors within the structure and flag the least-certain residues (never comparing B across structures without normalization).
  4. Screen backbone geometry for Ramachandran outliers and cis non-proline peptides in Python, then run phenix.molprobity for archive-calibrated clashscore, rotamer, and Ramachandran percentages when a real answer is needed.
  5. For predicted models, read pLDDT into bands, read the PAE matrix for domain segmentation, and process with phenix.process_predicted_model before docking or molecular replacement.
  6. For cryo-EM, direct the reader to the EMDB local-resolution map and distinguish the 0.143 (half-map) and 0.5 (map-model) FSC thresholds.
  7. For NMR, compute and report the per-residue ensemble spread instead of averaging or picking model 1.

Tips