TopPDBLX | Crystallisation condition index

DOI Source marc@marcdeller.com

#######               ######  ######  ######  #       #     #
   #     ####  #####  #     # #     # #     # #        #   #
   #    #    # #    # #     # #     # #     # #         # #
   #    #    # #    # ######  #     # ######  #          #
   #    #    # #####  #       #     # #     # #         # #
   #    #    # #      #       #     # #     # #        #   #
   #     ####  #      #       ######  ######  ####### #     #

An index of what crystallised proteins in the Protein Data Bank, and how little of it is specific to any one protein.

Formulation index · v1.0.0

What everyone tries

How often each reagent appears across every parsed condition. This is the query-independent frequency prior, and on this corpus it beats homology retrieval at every level tested.

Read this as popularity, not efficacy. PEG 3350 leads because it is in everyone's screens, not because it works best. The Protein Data Bank records only successes, so nothing here says what would have failed.

Filter reagents

Chemical class

Reagent	Class	Conditions	Share

What a reagent is used with

Pick a reagent to see which others appear in the same condition. Co-occurrence is computed over the 40 most frequent reagents.

Reagent

Appears with	Conditions	Share

Against the prior

Bars compare a partner's frequency inside these conditions with its frequency across the whole corpus. Similar heights mean the pairing tells you nothing.

Precipitant classes

Every condition is placed in one of seven classes, or left Unclassified with the reason recorded. Refusing to guess is deliberate.

Distribution

Conditions per class across the whole corpus.

Counts

Unclassified is a recorded outcome, not a gap in the data.

Class	Conditions	Share

Commercial screens

Conditions whose component set matches a published screen well. Formulations were transcribed verbatim from vendor support materials.

A match is not proof the screen was used. Many deposited conditions are optimised from a hit rather than the hit itself, so a component-set match with different concentrations is flagged rather than claimed.

Screen	Matched conditions	Share of matches

What this is, and what it is not

An exploration tool over a parsed dataset. It is deliberately not a recommender.

Why there is no "suggest a condition" button

Homology retrieval was tested against a query-independent frequency prior at three levels of granularity, on two split thresholds, under three definitions of ground truth. The prior won every time. Homologues get screened with the same popular screens as everything else, so there is little protein-specific signal to retrieve. A recommender here would dress up "what everyone tries" as a prediction.

Where the numbers come from

Source	Components	What it is

Every component in the released data carries a parser field, so the deterministic subset can be recovered exactly.