Even though the scientific idea in this publication seemed good, the extreme vagueness in the abstract was a turn off (this critique, while somewhat vague, is not a scientific paper). It's like they are channeling Stephen Hawking (or Larry Flynt, to be precise) and made it their mission to not quantify a single one of their statements (ergo, the Hawking reference), where a scientific paper is one of the few places where over quantifying makes sense. Please don't write like this:
- "hot topic"?
- "many efforts, ... many using"?
- "a benchmark"?
- "good acurracy"?
- "sometimes via very different proteins and chemicals"?
- "thousands of confident predictions"?
- "including structures ... lacking a structure"?
- "many new interactions"?
Predicting drug-target interactions is a hot topic, and many efforts have been undertaken to do this, many using large interaction networks. We take a novel approach using protein-chemical interactions derived from 3D structures. The basic premise is that two proteins sharing a common bound chemical will likely share others. We use protein and chemical superimpositions and physical tests of chemical-protein compatibility to identify the most likely candidates among the nearly one million potential interactions. We show for a benchmark that known protein-chemical structures are reconstructed with good accuracy and sometimes via very different proteins and chemicals. We make thousands of confident predictions, including structures for known protein-drug interactions lacking a structure (e.g. topoisomerase-2/radicicol) and many new interactions. The number of confident predictions grows faster than the number of known structures, suggesting that this approach will play a key role in completing the protein-chemical interaction repertoire.