Pharmer: efficient and exact pharmacophore search

Abstract

Pharmacophore search is a key component of many drug discovery efforts. Pharmer is a new computational approach to pharmacophore search that scales with the breadth and complexity of the query, not the size of the compound library being screened. Two novel methods for organizing pharmacophore data, the Pharmer KDB-tree and Bloom fingerprints, enable Pharmer to perform an exact pharmacophore search of almost two million structures in less than a minute. In general, Pharmer is more than an order of magnitude faster than existing technologies. The complete source code is available under an open-source license at http://pharmer.sourceforge.net .

Figure 1

Pharmacophore features (spheres) are identified in library compounds using user-configurable SMARTs expressions. The collection of compound features is decomposed into coordinate-frame independent triangles. These triangles, along with associated molecular data including a Bloom fingerprint, are stored in a spatial index using a canonical ordering of the three lengths of the triangle.

Figure 2

A simple 2D example of the Pharmer KDB-tree. (a) The area surrounding the point set is divided with cutting planes (solid lines) and bounding boxes (dashed lines) to form (b) a search tree. Because of the explicit bounding boxes, a range query around q (white point), despite intersecting several cutting planes, only need traverse a single branch of the tree, shown as thick-lined nodes in (b), and the search terminates early when it is determined that the points of node c-d are completely contained within the query range.

Figure 3

The generation of a Bloom fingerprint. (a) The position of pharmacophore features relative to a given triangle l₁l₂l₃ are exactly represented using the distances to the triangle vertices (i.e., d₁d₂d₃) and a chirality bit, which specifies which side of the plane defined by the triangle the point is in. (b) This positional information and the feature type of the point are provided to a set of k hash functions, in this case three, that set k bits in a bitvector, the Bloom filter.

Figure 4

(a) A pharmacophore query derived from the kinase inhibitor of PDB 3K5U. A hydrogen acceptor (A, orange), two hydrophobic (B/C, green) and a hydrogen donor (D, white) feature were extracted by analyzing the crystal structure. Each feature is given a tolerance radius of 1.0Å. (b) The convex hull of the set of possible lengths for triangle ABC inscribed in its bounding box (lengths ± 2.0Å).

Figure 5

The pharmacophores for HSP90 (left) and FXIa (right) generated using MOE (top) and Pharmer (bottom). Hydrogen acceptor features are cyan in MOE and orange in Pharmer. Hydrogen donor features are magenta in MOE and white and Pharmer. Hydrophobic features are green and MOE and Pharmer position these features slightly differently. The hydrogen donor feature of HSP90 has a radius of 1.25Å, the hydrophobic feature of HSP90 has a radius of 2Å, the hydrophobic feature of FXIa has a radius of 1.5Å, and all other features have a radius of 1Å.

Figure 6

The pharmacophore search performance of Pharmer relative to MOE on the HSP90 and FXIa queries of Figure 5. (a) Pharmer is more than an order of magnitude faster and (b) its performance scales with the query complexity, resulting in a three order of magnitude speed up for very precise queries.