DOGS: reaction-driven de novo design of bioactive compounds

Abstract

We present a computational method for the reaction-based de novo design of drug-like molecules. The software DOGS (Design of Genuine Structures) features a ligand-based strategy for automated 'in silico' assembly of potentially novel bioactive compounds. The quality of the designed compounds is assessed by a graph kernel method measuring their similarity to known bioactive reference ligands in terms of structural and pharmacophoric features. We implemented a deterministic compound construction procedure that explicitly considers compound synthesizability, based on a compilation of 25'144 readily available synthetic building blocks and 58 established reaction principles. This enables the software to suggest a synthesis route for each designed compound. Two prospective case studies are presented together with details on the algorithm and its implementation. De novo designed ligand candidates for the human histamine H₄ receptor and γ-secretase were synthesized as suggested by the software. The computational approach proved to be suitable for scaffold-hopping from known ligands to novel chemotypes, and for generating bioactive molecules with drug-like properties.

Figure 1. Encoding of reactions.

Example of a Paal-Knorr pyrrole reaction encoded as Reaction-MQL expression (top). Reactant substructure descriptions (left part) are separated by ‘++’. The product (right part) is separated from the reactants by ‘>> ID >>’ where ID is an arbitrary identifier of the reaction. A direct structural representation of the line notation description including atom identifiers is shown in the center. The conventional structural representation of the reaction (bottom) denotes variable parts of molecules by R-groups (R^x).

Figure 2. Two-step procedure of an extension cycle.

Step 1 (left) selects the reaction by scoring generated dummy products. In the example, only two reactions can be applied (Suzuki coupling and amide coupling), and the amide dummy product scores favorable. In Step 2 (right), all reactants from the building block library exhibiting a suitable amine are added to the growing molecule via amide bond formation. The top-scoring product represents the extended intermediate product and is selected for the next design cycle.

Figure 3. Flowchart of the molecule design algorithm.

(A) The stop criterion controlling the maximum number of reaction steps is excluded from the flowchart for simplification. (B) Detailed description of flowchart element B (grey circle). It comprises the key steps taken to extend intermediate product Z and yield the top-scored intermediate product Ž ( = Z grown by an additional fragment) by applying in silico reactions.

Figure 4. Reduced graph representation.

(A) An example of a reduced graph representation. Dashed lines connect atoms or rings of the molecule (left) with their corresponding vertex of the reduced graph (right). For clarity only some lines are shown. (B) Examples of polycyclic (‘amalgamated’) substructures translated to a single vertex in the reduced graph. (C) Edges of order two are used to connect fused rings (bottom) in order to distinguish the shown cases of neighbored rings in reduced graph representation.

Figure 5. Known trypsin inhibitors.

Five trypsin inhibitors served as reference compounds for DOGS design runs (Camostat , NAPAMP , Efegatran , Patamostat , , UK-156406 [50]).

Figure 6. Property distributions.

Comparison of property distributions between compounds designed by DOGS (left) and the reference compounds (right). ‘Rule of 5’ violations (A) and logP(o/w) values (C) were calculated using MOE. Dug-likeness scores (B) were computed by a trained neural network classifier (1 = high drug-likeness).

Figure 7. Bioisosteric replacement.

Side-chains addressing the S1 pocket present in the reference compounds (left) and surrogates suggested by DOGS found in top-scored 200 designs (right).

Figure 8. Side-chains addressing the S1 pocket of trypsin.

Known inhibitors of trypsin exhibiting pyrimidin-2-amine (left) and the pyridin-2-amine (right) side-chains (grey circles). These moieties were also suggested by DOGS as bioisosters for side-chains of the reference ligands addressing the S1 pocket of trypsin.

Figure 9. Suggested trypsin inhibitors.

Compounds 1 and 2 were proposed by the software as potential trypsin inhibitors. Reference ligands (Efegatran , Camostat [45]) and suggested synthesis pathways are presented for both candidate structures.

Figure 10. Automated design of γ-secretase modulators.

Compounds 3 and 4 were proposed by DOGS as potential modulators of γ-secretase. Synthesis plans were suggested by the software and successfully pursued. Molecules 3 and 4 originate from distinct runs based on different reference ligands .

Figure 11. Automated design of H₄ ligands.

Compounds 5 and 6 were proposed by DOGS based on an inverse agonist of hH₄R (A). Compound 7 is a design originating from the hH₄R antagonist JNJ7777120 (B).

Figure 12. Structural features of H₄R ligands.

Highlighted features of known H₄R ligands (compound 8 : central triazole ring; compound 9 : ether alkyl linker) are combined in designed compound 7. None of these features is present in the reference ligand underlying the design of 7.