Expanding the Palette of Molecular Fragments for Small Molecule De novo Design through Isosteric Swapping

Abstract

Outcomes from fragment-based assembly of small organic molecules are highly dependent on the number and type of fragments available for growth. Here, we present a new chemical searching strategy for the DOCK6 de novo design (DOCK_DN) engine which logically expands the palette of available fragments compared to the current method. Termed DOCK_SWAP, the new routine employs principles of isosteric swapping and a customized library infrastructure (iso-libraries) in which topologically related sidechains, linkers, and scaffolds have been pre-aligned and rank ordered relative to a parent fragment (iso-tables). The primary objectives of the current work are: (I) introduce the DOCK_SWAP infrastructure and algorithms, (II) characterize iso-library expansion outcomes and impact of different alignment & ranking protocols, (III) assess search performance, molecular scores, heterogeneity, and growth path coverage and (IV) evaluate lead refinement. Depending on protocol, the new iso-libraries are roughly an order of magnitude larger (N=2,197 to 2,708) than the existing library (N=382). Large-scale simulations (N=3,420) across 57 protein-ligand systems, using three different iso-libraries each, confirm that the DOCK_SWAP code base and infrastructure is robust. Although the use of larger DOCK_SWAP iso-libraries come at an increased cost in terms of simulation time, the results highlight the many advantages over the existing DOCK_DN method for design of drug-like ligands with improved complementarity to the sites being targeted.

Graphical Abstract

A new isosteric swapping method for small molecule design termed DOCK_SWAP is reported in which chemical moieties can be replaced during ligand construction with topologically-related fragments. The method employs a fragment library comprised of a parent HEAD for which topologically related TAILs have been pre-aligned. The new infrastructure rationally expands the palette of fragments available for growth and comprehensive tests across 57 protein-ligand systems indicates DOCK_SWAP can lead to molecules with more favorable scores and properties compared to the existing DOCK method.

Introduction.

Drug discovery campaigns often include one or more computational screening techniques to aid in the identification/prioritization of ligands for experimental testing. One of the most widely employed methods is atomic-level virtual screening (VS),¹ employing principles of molecular docking, to computationally evaluate millions to potentially billions of small molecules known to exist (or could likely be synthesized) to a target of interest, typically a protein. Some representative and widely used VS programs include DOCK6² (for which we are developers), DOCK3³, Autodock⁴, Vina⁵, and Glide^6,7, among others. The primary outcomes of VS include 3D binding predictions (poses) for ligands complexed to a target and prioritized lists of molecules for further consideration. Fortunately, ligand libraries for virtual screening are freely available from online databases such as ZINC,^8-11 for which the current version contains an astounding 4.5 billion ready-to-dock 3D molecules.¹¹ However, for the majority of users, brute-force docking of millions, hundreds of millions, let alone billions of ligands presents considerable challenges.¹² Further, since the organization and connectivity of fragments (i.e. topologies) in commercially available ligand libraries have been pre-defined, compounds may not be optimal for the target of interest.

As an alternative to VS, de novo design approaches (abbreviated DN) have been developed that in principle reduce the need to perform costly ultra-large scale screening campaigns because molecules can be generated on-the-fly through the assembly of fragments directly in the context of the binding site.^13-19 Theoretically, by removing the need to adopt pre-defined topologies, DN could lead to more productive exploration of chemical space,^20-22 with greater computational efficiency, and should therefore be explored further as a complement to ultra-large scale docking. For example, DN methods could be employed to construct molecules customized for a specific binding site, and top scoring candidates could then be used to search ultra-large libraries to identify similar compounds for creation of more manageably sized bespoke libraries for VS. Some representative DN programs, focusing on those that primarily employ fragment-based construction methods, include OpenGrowth,¹⁴ BOMB,¹⁷ LUDI,¹⁸ and DOCK6,^15,16,19,23 among others.

The well-validated DN method^15,16,19,23 in DOCK6 (abbreviated hereafter as DOCK_DN) enables users to construct new molecules from scratch, in a layer-by-layer fashion, from fragments (Figure 1a) with attachment points (APs) classified as sidechains (1 AP), linkers (2 APs), and scaffolds (3+ APs). As candidate fragments are added to a growing molecule, the would-be bond is inspected¹⁶ to ensure it was observed previously in the database originally employed in the generation of the fragment libraries themselves. This procedure ensures chemical soundness and helps promote physically reasonable outcomes. As outlined in Allen et al,¹⁶ the existing DOCK_DN method was extensively validated through re-construction of crystallographic binding poses, construction of new ligands, and application to a challenging protein-protein interface.

Figure 1.

(a) Standard DOCK6 de novo design method (DOCK_DN) yields a "single" molecule for a given accepted fragment. (b) DOCK6 isosteric swapping method (DOCK_SWAP) yields an "ensemble" of molecules by swapping out the accepted fragment with a group of pre-aligned and topologically-related fragments.

In the present work, we develop new infrastructure and algorithms collectively termed DOCK_SWAP (Figure 1b) which facilitates fragment-based isosteric swapping (also called isosteric replacement). The new method was inspired by the technique commonly employed by medicinal chemists,^24,25 in which chemical moieties on a ligand are systematically replaced with similar chemical groups, termed isosteres. At its core, isosteric swapping is a structure activity relationship (SAR) method geared to improve ligand binding to the target or alter pharmaceutically important properties. Although isosteres may be defined differently by different researchers, DOCK_SWAP isosteres are classified as related groups of "fragments" based on geometric and/or chemical similarities. While replacing peripheral functionality on a complete molecule with related groups to generate a congeneric series of analogs would be the most obvious use case for isosteric swapping, the method is also effective during ligand construction.

Importantly, DOCK_SWAP greatly expands the number of fragments available for growth which increases the overall search space. The default library of 382 fragments, derived by deconstructing ~13M drug-like compounds downloaded from ZINC^8-10 (termed ZINC13M) and retaining those occurring >13,000 times is only a fraction (0.57%) of the 67,579 total number of fragments obtained. DOCK_SWAP employs a newly designed library infrastructure which in the current work expands the library of searchable fragments by about an order of magnitude. For example, DOCK_DN searching involves the selection of one of 382 fragments followed by evaluation for compatibility with the molecule (Figure 1a), which, if successful results in a single completed chemotype. In contrast, if one of the 382 fragments is successfully attached using DOCK_SWAP (Figure 1b), the molecule enters isosteric swapping subroutines and up to 20 "similar" fragments are evaluated, potentially resulting in an ensemble of related chemotypes (analogs).

The primary objectives of the current work are: (I) introduce the DOCK_SWAP infrastructure and algorithms, (II) characterize iso-library expansion outcomes and impact of different alignment & ranking protocols, (III) assess search performance, molecular scores, heterogeneity, and growth path coverage and (IV) evaluate lead refinement. It is important to emphasize that this work focuses on comparing the new DOCK_SWAP method to the currently available DOCK_DN method, which for the purposes of discussion, is taken as ground truth. Overall, the outcomes suggest that the DOCK_SWAP method can effectively increase the number of fragments available for de novo design and will be a useful tool for those employing DN approaches for their research. All the algorithms described here have been incorporated into the current DOCK6 codebase and will be available to the community in a forthcoming release to coincide with publication of this manuscript.

Theoretical Methods

(I) DOCK_SWAP infrastructure and algorithms

DOCK_SWAP iso-library generation.

The new DOCK_SWAP isosteric swapping method described in this work is built upon the previously described DOCK6 de novo design engine (DOCK_DN).^16,19 DOCK6 also includes a molecular evolution engine based on a Darwinian-inspired genetic algorithm (DOCK_GA),^15,26,27 in addition to the more well-known virtual screening engine (DOCK_VS)²⁸ used to dock pre-existing chemical libraries. The DOCK_SWAP infrastructure was developed around a data structure termed "iso-library" which is itself a collection of individual "iso-tables", each of which contains a group of topologically-related fragments aligned in 3D space and arranged as a rank ordered list. As shown in Figure 2, the generation of an iso-library involves three separate steps, which include alignment, collection, and ranking. The alignment step (Figure 2a) involves generation of iso-tables in which the first index is referred to as the HEAD fragment and the remaining indexes are referred to as the TAIL fragments. In the current work, the HEAD fragments were taken from the standard generic library (N=382). For each iso-table, candidate TAIL fragments are individually aligned to the HEAD (see Supporting Information for more details on alignment protocol, Figures S1-S3, Table S1, and accompanying text) based on alignment scores computed using the Hungarian Matching Similarity (HMS)²⁹ and Volume Overlap Heavy (VOH)³⁰ scoring functions in DOCK6. The goal is to identify and retain a curated group of TAIL fragments which are topologically similar to and geometrically aligned with each HEAD. In the current work up to 100 entries were retained although this number can be user-defined. For simplicity and efficiency, all fragments in a given iso-table have the same number of APs and their overlapped AP exit vectors are of the same bond type. The collection phase (Figure 2b) organizes individually created iso-tables into a group, however it is important to note that at this stage the TAIL fragments are not yet sorted (ranked) based on alignment scores. Ordering during the collection stage is dependent on the order of user-defined input fragments. The final ranking stage (Figure 2c) rank-orders the group of up to 100 TAIL fragments in each iso-table based on either the HMS or VOH scoring functions. Rank ordering using a 2D-based Tanimoto (TAN)³¹ scoring function was also explored.

Figure 2.

Iso-library construction involves three stages: (a) alignment, (b) collection, and (c) ranking. In step (a) candidate fragments for inclusion (termed TAILs) are aligned to the parent fragment (termed HEAD) and stored in an iso-table (up to 100 entries) but the entries are not rank ordered. Step (b) collects all iso-tables into a collection. The number of iso-tables is dependent on the number of HEAD fragments employed for the alignment step. Step (c) uses different rank-ordering functions to rank-order TAILs, relative to the relevant HEAD, to yield the final iso-library.

Each iso-library is stored in directory comprised of 382 individual iso-tables each containing the 3D atomic coordinates and bond information for the HEAD and TAIL fragments ordered by rank and saved in multi-mol2 format. In terms of nomenclature, an iso-library designated as HMS/HMS would refer to molecules being aligned using the HMS function (Figure 2a) followed by ranking with the HMS function (Figure 2c). In principle any similarity-based scoring function (or combinations) could be adapted for the alignment and/or ranking stages. The present work evaluates three different iso-libraries (HMS/HMS, VOH/VOH, VOH/TAN) for comparison with the standard generic library.

DOCK_SWAP search algorithm.

Figure 3 provides additional details with regards to the core DOCK_SWAP algorithm illustrated in Figure 1. For a given algorithmic move, an iso-table is randomly selected from the specified iso-library (Figure 3a) and the HEAD fragment is attempted to be attached to a target AP of the root molecule. Here, a root molecule could refer to a previously oriented "anchor" or partially grown molecule. The example root molecule in Figure 3 (yellow) is an ethyl sidechain with one unsatisfied AP (magenta dummy atom). If the attachment is allowable, DOCK_SWAP will sample dihedral angles (torsions) associated with the newly formed bond using the DOCK6 anchor-and-grow algorithm and retain only the conformation with the best overall score (Figure 3b). Users should note this is different than the DOCK_DN algorithm for which multiple torsions for each attached fragment may be saved (discussed further in Results & Discussions). When a HEAD is successfully attached the related group of TAILs is also evaluated for attachment (Figure 3c). Importantly, because the TAIL fragments have been pre-aligned to the HEAD the previously generated orientation matrix and torsion angle can be applied to the replacements which greatly reduces the number of calculations and ensures orientation of high geometric overlap. Finally, all newly formed molecules (provided the newly formed bonds are allowable) are subjected to a simplex-based energy minimization.^2,32

Figure 3.

Attachment of fragments to a root molecule using DOCK_SWAP. (a) For a given target AP on a root molecule an iso-table is selected at random, and the HEAD fragment is attempted to be attached. (b) If the HEAD attachment is successful, equivalent linear transformations are applied to the corresponding TAIL fragments (orientation matrix and torsion angle) to yield (c) a related group of molecules with attached iso-table TAILs.

In the present work, for a given AP, DOCK_SWAP attempts to attach 20 different HEADs followed by an additional 20 TAILs each (values can be user defined) for up to 400 fragments per move. Additionally, DOCK_SWAP will attempt all possible AP permutations for fragments containing more than one AP (e.g linkers and scaffolds). Hypothetically, if all the selected fragments contained 3 APs the theoretical number of attempted attachments would be 1200 (20 heads x 20 tails x 3 APs). As high-volume attachments and efficient linear transformations work in tandem, searching through chemical space is more efficient and thorough.

Computational Details

Systems Studied.

The primary simulations employed in this work involved ligand growth initiated in 57 protein structures curated from Protein Data Bank³³ (57 unique pdb codes) representing eight distinct protein families (Table 1). For these simulations, the five most frequent fragments in the library from each fragment class (15 total) were selected as "anchors" to seed growth (Figure 4). It is important to note there is a distinction between the smaller number of anchors which are oriented into the binding sites to seed growth (in this case 15) versus the much larger number of fragments actually used during layer-by-layer assembly. For each system, four distinct simulation protocols were employed: DOCK_DN using the standard fragment library (control), and DOCK_SWAP using HMS/HMS, VOH/VOH, and VOH/TAN iso-libraries. Cumulatively, this resulted in a total of 3420 individual simulations (57 proteins x 15 fragments x 4 protocols). We also examined the utility of using DOCK_SWAP for sidechain refinement, employing the approved drug celecoxib (pdb 3LN1)³⁴ complexed with its target cyclooxygenase-2 (COX-2) protein as a representative example.

Table 1.

Protein families and pdb codes employed for simulations.

protein family (size)	pdb code
acetylcholinesterase (5)	1EVE, 1H22, 1J07, 1Q84, 1ZGC
cyclooxygenase (6)	1EQG, 1EQH, 1HT5, 1HT8, 1Q4G, 4COX
EGFR (5)	2ITP, 2ITT, 2ITY, 2RGP, 3BEL
HIV PR (12)	1AJV, 1DMP, 1HVR, 1MER, 1MES, 1MET, 1QBS, 2F80, 2F81, 2IDW, 2IEN, 2IEO
HIV RT (10)	1C1B, 1C1C, 1VRU, 2BE2, 2RKI, 2ZD1, 3BGR, 3DLE, 3DLG, 3DOL
IGF1R (4)	2ZM3, 3NW5, 3NW6, 3NW7
neuraminidase (10)	1BJI, 1F8B, 1F8C, 1F8D, 1F8E, 1MWE, 1NNB, 1NNC, 1XOE, 1XOG
streptavidin (5)	1DF8, 1SRG, 1SRI, 1SRJ, 2IZL
Total N = 57 systems

Figure 4.

Top 5 anchors from each fragment class (N=15 total) used to seed ligand growth in 57 protein systems showing 3D structure, bond types, fragment code, fragment name, and frequency derived from deconstructing the ZINC13M dataset.

System Setup Protocols.

The system setups employed in this work were based on protocols previously described for construction of our SB2010 and SB2012 docking test sets.^16,35 Briefly, for each system, raw pdb files were downloaded from the Protein Data Bank³³ and complexes were separated into individual protein, ligand, and co-factor (if present) files. Ligands and cofactors were protonated, visually examined, and assigned GAFF³⁶ force field parameters augmented by AM1-BCC^37,38 partial atomic charges (antechamber³⁹ program). Protein structures were protonated, assigned ff14SB⁴⁰ field parameters, and simulation ready complexes were generated using the program tleap included with the Amber⁴¹ program suite of tools (AmberTools). Prepared complexes were then briefly energy minimized with Amber16⁴² to relax the coordinates. Minimized protein structures were extracted and saved in mol2 format and used to generated molecular surfaces via the program DMS (1.4 Å radius probe).⁴³ For each molecular surface, a set of docking spheres was generated using the program sphgen⁴⁴ to aid in orientating the 15 ligand anchors (Figure 4) into each protein site. To speed up the calculations, docking grids were generated using the DOCK accessory program GRID,⁴⁵ which pre-computes van der Waals (6–9 Lennard-Jones exponents) and electrostatic (Coulombic interactions scaled by a distance-dependent dielectric constant = 4r) contributions from the protein. The grids, in conjunction with a docked ligand pose, forms the basis of the DOCK6 single grid energy (SGE) scoring function used to compute the intermolecular non-bonded van der Waals (SGE_VDW) plus electrostatic (SGE_ES) interactions between ligand and protein.

DOCK6 parameters.

The de novo design engine currently contains three DOCK_DN methods which control how fragments are prioritized for attachment: (i) random selection, (ii) graph-based selection, and (iii) exhaustive selection.¹⁶ The work described here adds an additional method based on (iv) isosteric selection. The dn_sampling_method keyword (random, graph, exhaustive, isoswap) controls which of the methods are used. Table 2 summarizes other key parameters employed in the study. For most of the experiments, 20 fragments were selected at random using DOCK_DN with the generic library, or 20 HEAD fragments using DOCK_SWAP with one of the three iso-libraries (dn_num_random_picks, dn_num_rand_head_picks). For DOCK_SWAP, for each successfully attached HEAD (dn_iso_skip=yes), the 20 top-ranked TAILs (dn_num_iso_picks, dn_iso_pick_meth) in the iso-table were also attempted. During the simulations, partially grown molecules were pruned (see Table 2 for values) if they did not conform to specific molecular properties (MolProp), including Molecular Weight (MW), number of Rotatable Bonds (#ROT), and Formal Charge (FC), ranges defined by the user (dn_upper_constraint_mol_wt, dn_lower_constraint_mol_wt, dn_mol_wt_std_dev, dn_constraint_rot_bon, dn_constraint_formal_charge). The Hungarian RMSD (hRMSD) function was used to prune molecules which were too similar in terms of conformations (dn_heur_matched_rmsd). The size limitations (dn_max_root_size and dn_max_layer_size) on the Layer and Root vectors were both set to be 25 molecules. Molecules selected for propagation to the next layer of growth were based on the best energy scoring candidates (dn_random_root_selection) and the maximum number of growth layers was set to 9 (dn_max_grow_layers). For lead refinement however the dn_max_grow_layers was set to 1, DOCK_DN employed exhaustive sampling, and DOCK_SWAP upped the dn_num_iso_picks value from 20 to 3,446 fragments (which is equivalent to "exhaustive"). For additional information on DOCK_DN interested readers should consult Allen et al¹⁶ and the DOCK6 manual (dock.compbio.ucsf.edu).

Table 2.

Key DOCK6 searching parameters employed in this work.

Parameter	Description	Values Used
dn_sampling_methoda	Fragment search methods available (random, graph, exhaustive, isoswap)	random, exhaustive, isoswap
dn_num_random_picks	N fragments randomly selected for random	20
dn_num_iso_head_picks	N fragments randomly selected for isoswap	20
dn_iso_pick_meth	TAIL fragment selection method for isoswap (top_rank, random, worst_rank)	top_rank
dn_iso_skip	Skip TAIL attempts if HEAD frag fails to attach	yes
dn_num_iso_picks	N TAIL fragments randomly selected	20, 3446
dn_random_root_selection	Randomly select molecules propagated to next layer	no
dn_mol_wt_cutoff_type	Molecular weight filtering method (hard, soft)	soft
dn_upper_constraint_mol_wt	Upper limit for MW filter	550
dn_lower_constraint_mol_wt	Lower limit for MW filter	0
dn_mol_wt_std_dev	Standard deviation for MW filter	35
dn_constraint_rot_bon	Max allowed number of rotatable bonds (#ROT)	15
dn_constraint_formal_charge	Max allowed formal charge of molecule (FC)	2
dn_heur_matched_rmsd	Max RMSD of matching heavy atoms of molecules	2.0
dn_max_grow_layersb	Max N of layers for growth starting from an anchor	9, 1
dn_max_root_size	Max N of partially grown molecules passed to next layer	25
dn_max_layer_size	Max N of molecules allowed from a given growing root	25
dn_max_successful_att_per_root	Max N of successful attachments allowed per root	50,000

^a,bLead refinement employed exhaustive searching method with 1 growth layer.

Simulation analysis: Molecular output details.

As molecules are advanced through successive layers of growth using DOCK_DN or DOCK_SWAP, multiple pruning stages are employed to prevent combinatorial explosion and promote drug-likeness. Table 3 provides an overview in terms of growth outcomes, pruning outcomes, and final outcomes for molecules constructed during a typical simulation. Briefly, growth outcomes for molecules can be categorized as "finalized", "root candidate", or "propagated". Molecules become finalized (Table 3, index 1-2) when all possible APs are occupied with a compatible sidechain. Finalized molecules are retained and written to the primary output file (Table 3, index 1) provided that ligand molecular properties (MolProp) are within desired ranges. Otherwise, they are pruned (pruning outcomes), but if desired they can be written out to auxiliary output files (Table 3, index 2). Although not discussed here, additional pruning stages using cheminformatics-based descriptors which complement those listed in Table 3, can be performed using our recent integration¹⁹ of DOCK6 with RDKit. At successive layers of growth, as unsatisfied APs are engaged with compatible fragments, some molecules will incorporate linkers and/or scaffolds which introduce new APs. Such molecules will be potential root candidates for the next layer of growth.

Table 3.

Growth, pruning, and final outcomes for molecules constructed during de novo design.

	Growth outcomes a		Pruning outcomes b				Final outcomes c
Index	Type	Action	Lsize	MolProp	hRMSD	Rsize	Fate	Filename label
1	finalized	APs filled	n/a	pass	n/a	n/a	Retained	DN build (primary)
2	finalized	APs filled	n/a	fail	n/a	n/a	Pruned	filtered.i (auxd)
3	root candidate	APs capped	fail	fail	n/a	n/a	Pruned	prune root.i (aux)
4	root candidate	APs capped	n/a	n/a	fail	n/a	Pruned	prune rmsd.i (aux)
5	root candidate	APs capped	n/a	n/a	n/a	fail	Pruned	root ignore.i (aux)
6	propagated	APs remain	pass	pass	pass	pass	Advanced	root.i (aux)

^aGrowth outcome types include "finalized", "root candidate", or "propagated".

^bPruning outcomes depend on growth outcome types and include removal when molecules fail to adhere to parameter values or ranges specified in the input file with regards to user-defined layer size (Lsize), molecular properties (MolProp), Hungarian-based RMSD (hRMSD), and root size (Rsize).

^cFinal outcomes involve being "Retained" and written to the primary output file (multi-mol2 format), "Pruned" and written to one of several auxiliary output files (if desired) using a layer-based index i, or "Advanced" as roots for the next layer of growth.

^dShorthand for auxiliary.

It is important to emphasize candidates may also be pruned based on the maximum allowable values for two intermediary vectors (Table 3, index 3 and 5) which store molecules pertaining to (i) the attachment of new fragments (Lsize defined by the dn_max_layer_size parameter), and (ii) the transition between growth layers (Rsize defined by the dn_max_root_size parameter). Root candidates may also be pruned based on MolProp (Table 3, index 3), as well as an additional hRMSD (Table 3, index 4) similarity check which removes geometrically redundant poses. Depending on the step, molecules flagged for pruning may be capped with hydrogen atoms (if physically reasonable) and written to second group of auxiliary output files (Table 3, index 3-5). Finally, root candidates to be "Advanced" to the next layer may be selected randomly or based on favorable fitness (defined by dn_random_root_selection parameter, Table 2), and for tractability can be written out to a third auxiliary output file (Table 3, index 6). Readers should note that all molecules written to the primary and auxiliary output files (Table 3, index 1-5) are valid in the sense that the aforementioned allowable torsion check has been applied to verify the reasonableness of all newly formed bonds and the binding site poses are energy-minimized. For most users, the primary DOCK_DN or DOCK_SWAP output file labeled Retained (Table 3, index 1) containing molecules in mol2 format that conform to all user-defined parameters would be the most natural group of results to examine first. The bulk of the analysis results employed this group. To facilitate the analysis of simulation outcomes, in terms of the topological and geometric similarity of constructed molecules, we implemented a best-first clustering approach into DOCK6 as a new utility function (see Supporting Information).

Results & Discussion

(II) Iso-library expansion outcomes and impact of different alignment & ranking protocols

Iso-libraries expand the palette of theoretical fragments available for chemical searching.

As described in Methods, DOCK_SWAP iso-libraries are comprised of a group of iso-tables each containing an ensemble of topologically related TAIL fragments geometrically aligned to a parent HEAD fragment. Table 4 shows size statistics for the three different iso-libraries examined here for comparison with the two baseline libraries (generic library, ZINC13M). Here, "overall" refers to the total number of entries in each library, "uniq" refers to the number of unique non-redundant entries, and "rec" refers to the number of fragments actually recovered by deconstructing molecules generated by DN simulations initiated in 57 proteins seeded with 15 anchors each. For completeness, in most cases, results and analysis in this manuscript compare DOCK_SWAP results using all three iso-libraries to DOCK_DN with the generic library, given that there can be considerable differences depending on the metrics being examined.

Table 4.

Iso-library composition in terms of number of fragments and torsions compared to baseline libraries.

	iso-lib size	iso-library type									baseline library type
	all types	HMS/HMS			VOH/VOH			VOH/TAN			generic		ZINC13M
	overalla	uniq-allb	uniq-20b	rec-20c	uniq-all	uniq-20	rec-20	uniq-all	uniq-20	rec-20	uniq	rec	uniq
sid	21,519	2,637	1,429	1,418	2,551	1,352	1,347	3,154	1,689	1,668	217	217	38,949
lnk	9,874	1,247	800	794	1,220	729	727	1,281	903	894	146	146	20,892
scf	211	136	115	108	136	116	111	136	116	1113	19	19	7,738
total	31,604	4,020	2,344	2,320	3,907	2,197	2,185	4,571	2,708	2,675	382	382	67,579
tors	10,844	10,844	10,844	5,695	10,844	10,844	6,121	10,844	10,844	5,555	10,844	4,412	10,844

^aTotal number of sidechains, linkers, scaffolds, and torsions in the three iso-library types is identical.

^bNumber of unique (uniq) sidechains, linkers, scaffolds, and torsions (abbreviated tors) in each iso-library for all (up to 100 max) and the top 20 fragments.

^cNumber of recovered (abbreviated rec) sidechains, linkers, scaffolds, and torsions using different iso-libraries (using the top uniq-20 fragments) for de novo design from simulations across 57 proteins (15 seed anchors each).

It is important to emphasize that all three iso-libraries (HMS/HMS, VOH/VOH, VOH/TAN) contain the exact same number of sidechains (21,519), linkers (9,874), and scaffolds (211) which sums to 31,604 total fragments (Table 4, column labeled "iso-lib size") and they each employ the same companion group of 10,844 allowable torsion types. It is important to emphasize that this is because all three are derived from a common alignment procedure. However, each iso-library employed a distinct ranking protocol (see Supporting Information, Table S1), thus, the ordering of TAILs will be different across the different iso-libraries. Further, since the current DOCK_SWAP protocol attempts to attach only the 20 top-ranked TAILs for a given move, the specific fragments selected for attachment will be different when different iso-libraries are employed. The discrepancy between the theoretical maximum (382 x 100 = 38,200) and the observed (31,604) iso-library sizes arise from some iso-tables not containing 100 entries. For example, the lnk.102 HEAD fragment contains a saturated heterocyclic pentane with two AP exit vectors pointing in different directions. Although the observed frequency of the lnk.102 (HEAD) is relatively high (77,616 in the ZINC13M library), and well-above the frequency cutoff of >13, 000 used to generate the generic library, the number of topologically-related and eligible TAIL candidates for the lnk.102 iso-table was only 6 fragments using the current protocols.

Table 4 also highlights that the number of "unique" fragments in the different iso-libraries is dramatically less by about an order of magnitude than the "overall" number of fragments. For example, the overall number of sidechains derived by summing across all 382 iso-tables is 21,519 while the number of unique sidechains in HMS/HMS = 2637, VOH/VOH = 2551, and VOH/TAN = 3154. These dramatic differences arise because a given fragment can be incorporated into multiple iso-tables as illustrated in Figure 5 which highlights two representative examples taken from the HMS/HMS iso-library. Here, the significant geometric similarity between methylene (lnk.0), ether (lnk.5), and thioether (lnk.17) HEADs (Figure 5 top) leads to nearly the same group of TAILs being incorporated into each of the three respective iso-tables as shown by the accompanying Venn diagram intersection for all entries (Figure 5a). Importantly, and as would be expected, the intersections of the Venn diagrams based on top 50 (Figure 5b, N=34) or top 20 (Figure 5c, N=12) ranked entries systematically show less overlap. An analysis of benzene (sid.7), p-fluorobenzene (sid.24), and p-methylbenzene (sid.30) iso-tables yields similar trends (Figure 5 bottom) with progressively less sidechain overlap being observed as the number of top-ranked fragments becomes less: top 100 or max (Figure 5d, N= 27) top 50 (Figure 5e, N= 10) top 20 (Figure 5f, N= 1). The intersection values for the three "pairwise sets" (benzene∩p-fluorobenzene, p-fluorobenzene∩p-methylbenzene, and benzene∩p-methylbenzene) for these three sidechains followed the same trend: top 100 (27, 6, 23) > top 50 (12, 1, 11) > top 20 (3, 0, 3). Overall, while variety is important, we also believe that some redundancy between iso-tables is beneficial because it increases the chances that similar move sets (i.e. fragment picks) can be attempted for topologically-related fragments, analogous in spirit to how a medicinal chemist would likely approach the problem.

Figure 5.

Topological comparisons for three related linkers (methylene, ether, thioether) and three related sidechains (benzene, p-fluorobenzene, p-methylbenzene) and associated Venn diagrams showing the intersections of fragments in the accompanying iso-tables for the top 100 (or max), top 50, and top 20 ranked entries from the HMS/HMS iso-library.

The number of unique fragments represents an upper bound in terms of those being "theoretically available" for molecular growth but does not denote how many might get picked and survive in practice. To quantify how many fragments become incorporated using different DN protocols we deconstructed the molecules designated as Retained (Table 3 index 1), from simulations initiated in 57 different protein systems (see Computational Details), and counted if a given sidechain (sid), linker (lnk), or scaffolds (scf) type was actually recovered as shown in Table 4 (recovery rate = "rec" or "rec-20" divided by "uniq" or "uniq-20" as appropriate). For the DOCK_DN simulations, with the small 382 fragment library, there was a 100% fragment recovery rate (382/382) and 40.69% torsion recovery rate (4,412/10,844). Notably, the DOCK_SWAP simulations led to a 99% fragment recovery rate using the protocol in which up to the top 20 TAILs were considered (2,320/2,344; 2,185/2,197; 2,675/2,708) and a 51-56% torsion recovery rate (5,695/10,844; 6,121/10,844; 5,555/10,844) depending on the iso-library employed. These are important outcomes because it shows that despite the larger sizes of the uniq-20 libraries relatively to the generic library (6-7 fold for sidechains, 5-6 fold for linkers, and 6-fold for scaffolds) there is essentially complete (i.e. 99%) sampling/retainment of fragments. The concomitant increases in torsion recovery rates (51-56% using iso-libs vs 40.69% generic) shows increased chemical searching from a different perspective. To more fully explore the composition between different iso-libraries we also computed "congruence indices" between pairs of iso-tables. As shown in Supporting Information (see Figure S4 and accompanying text), iso-library composition is dependent on which fragment class type is employed (sidechains, linkers, or scaffolds).

Top-ranked TAILs across different iso-libraries show considerable variety for a given HEAD fragment.

Figure 6 provides a more in-depth visual representation of TAIL fragments in the different iso-tables for the benzene sidechain HEAD fragment (sid.7) in the three iso-libraries (6a) HMS/HMS, (6b) VOH/VOH, and (6c) VOH/TAN. For simplicity, only the top 9 ranked TAILs (entries 2-10) in each group are shown. Displayed below each fragment are the fragment ranking scores and their accompanying frequencies observed in the ZINC13M library. The aligned group of fragments shown at the bottom of each plot highlights the geometric overlap. A point to emphasize regarding the fragment ranking scores in Figure 6 is that the HMS/HMS and VOH/VOH scores are continuous while the VOH/TAN scores may be quantized. As an example, all top 9 TAIL fragments in the VOH/TAN group in Figure 6c differ only from the benzene HEAD by a single heavy atom (i.e. methyl, fluorine, or chlorine substituent) yielding the same Tanimoto score even though they may be at different ring positions relative to the AP. Thus, for this group of TAILs the rank-ordering within the group itself is not significant. While this points to a possible criticism of the DOCK6 Tanimoto function employed for ranking small library fragments, and it would be worthwhile to explore other functional forms, the rank ordering between groups of "quantized" fragments is expected to be significant. It is also important to emphasize that the DOCK_SWAP fragment selection protocol employed here is based on trying all of the top 20 fragments and not their actual rank.

Figure 6.

Iso-table comparisons showing the top 9 scoring TAIL fragments (tan) in relationship to the HEAD fragment (green) for the benzene sidechain (sid.7) for the three different iso-libraries: (a) HMS/HMS, (b) VOH/VOH, and (c) VOH/TAN. The numerical scores used for ranking and the frequency of occurrence for each fragment in the ZINC13M library are listed below each entry. HMS scores range from −5 (most similar) to ∞, VOH score range from 0 to 1 (most similar), TAN scores range from 0 to 1 (most similar). Colored squares indicate iso-table entries in common. The overlapped fragments from each example are also shown.

An additional point worth noting is that the frequencies associated with each fragment do not correlate with the rank. This is intentional because the alignment protocols do not consider frequency during iso-table construction. Future work could examine use of frequency for rank-ordering as well as exploring more stringent frequency cutoffs than the value of 130 applied here (Table S1). Fragment selection priority could also be a function of rank (as discussed recently in Bickel et al).²³ In the present study we opted to strike a balance in term of achieving a wider range of functionality while also removing extremely rare fragments (< 130 times in 13M molecules). For the examples in Figure 6, the frequencies for the top 10 ranking fragments vary widely between 1,130 and 1,816,793. For additional context, the 382 fragments employed as the parent HEAD fragments were originally determined using a much higher frequency cutoff of 13,000 which roughly 0.1% of the ca. 13M molecules deconstructed to generate the library.

Figure 6 also visually highlights the fact that for a given HEAD, different iso-tables may or may not contain the same fragments, which is consistent with the discussion above regarding iso-table overlap (Figure 5) and congruence (Figure S4). For example, the only TAIL fragments in common between the different iso-tables examples in Figure 6 are the four entries outlined with colored squares (Figure 6a versus 6c, red, cyan, blue, orange) although, as noted above, the rank position of these specific VOH/TAN entries are not significant given they all have the same Tanimoto value. Interestingly, for these examples, VOH/VOH (Figure 6b) has no TAILs in common with the other two iso-tables. It is important to emphasize that VOH prioritizes only heavy atom overlap between a candidate and reference, unlike the HMS or TAN functions which also include atomic features. Consistent with this rationale, the overlapped group of top ranked TAILs from VOH/VOH is more "geometrically converged" to the benzene HEAD than HMS/HMS or VOH/TAN. Although these top-ranked VOH/VOH results do not have ring substituents, there is considerable variability in terms of the atomic makeup of the ring containing one or two nitrogen atoms in different arrangements. Ultimately, the VOH/VOH TAILs show similar but not identical alignment scores to the benzene HEAD. Interestingly, the entries with one nitrogen (Figure 6b ranks 2-4) or two nitrogen atoms (Figure 6b ranks 5-10) are clustered together, which likely arises from different combinations of N-C and N-N bonds being shorter than C-C. Taken together, Venn diagram analysis (Figure 5), congruence index profiles (Figure S4), and visual examination of fragments (Figure 6) indicate there is considerable variety among top-ranked TAIL entries across different iso-libraries.

(III) Search performance, molecular scores, ligand heterogeneity, and growth path coverage.

Search performance.

By design, iso-libraries will increase chemical searching, however, their larger size is also expected to lead to longer simulation times compared to using the smaller generic library. Figures S5-S7 examine run-time performance in terms of simulation time, and fragment attachment metrics, across a large group of 3,420 independent protein simulations (DOCK_DN = 855, DOCK_SWAP x 3 iso-libraries= 2,565). As anticipated, DOCK_SWAP runs are considerably slower than DOCK_DN (Figure S5 top), and the number of attachments attempted is highly correlated with simulation time (Figure S5 bottom). However, a much greater number of molecules are generated and classified as "Retained" (Table 3, index 1) using DOCK_SWAP (12-14M, Figure 7a red, green, blue) compared DOCK_DN (1.5M, Figure 7a black) which is desirable. Somewhat surprisingly, DOCK_SWAP simulations yielded slightly fewer Retained molecules generated per second (Figure 7b) even though >95% of the simulations yielded higher attachment velocity ratios (Figure S7). For example, 55 of the 855 DOCK_DN simulations (Figure 7b black) generated > 6 Retained molecules per second compared to the different DOCK_SWAP simulations which yielded between 13, 34, or 11 out of 855 (Figure 7b red, green, blue). We hypothesize this discrepancy likely stems from an algorithmic difference between the two methods regarding the number of conformers saved at each stage of growth. As noted in Methods, DOCK_DN can save multiple torsions for each successful attachment while DOCK_SWAP currently saves only a single (best-scoring) torsion. Thus, should a sidechain be successfully attached to an anchor using the DOCK_DN method, and should the group of conformers pass all pruning steps, the entire group could be classified as Retained and thus written to file.

Figure 7.

(a) Total number of Retained molecules for 855 processes (57 protein systems seeded with 15 anchors each) using DOCK_DN (black) or DOCK_SWAP methods with HMS/HMS (red), VOH/VOH (green), or VOH/TAN (blue) iso-libraries. (b) Number of Retained molecules generated per second for each procedure (ranked high to low) and the overall number of processes (N proc) leading to > 6 molecules per second.

To help gauge if retaining only a single torsion might limit conformational sampling, we compared DOCK_SWAP and DOCK_DN ligand poses in terms of their similarity to crystallographic poses across 57 systems. Similar outcomes would support that the torsion sampling protocol used in DOCK_SWAP was reasonable. It is important however to emphasize that growth was performed without requiring fragment libraries to contain all possible building blocks in the control group of 57 ligands, and similarity-based scoring was not enforced, thus it is unlikely the exact same chemotypes could be rebuilt. The primary focus here was comparing the methods themselves. For analysis, we employed our HMS²⁹ scoring function because it provides an easy-to-understand metric through the incorporation of both geometric and chemical overlap into a single term which facilitates rank-ordering and examination at different cutoffs. As a reminder, an HMS score of −5.0 represents perfect overlap between any two species. From the group of Retained molecules generated in each system with DOCK_DN (black) or DOCK_SWAP (red, green, blue) methods, Figure 8 plots the best HMS score obtained (8a) along with the number of molecules generated within a specific similarity cutoff range (8b): HMS < 0 (top), HMS < −1 (middle), HMS < −2 (bottom). To help discern general trends, the results in each panel are independently rank-ordered from best to worst (8a) or largest to smallest (8b). Reassuringly, all four methods appear to be comparable in terms of their ability to generate at least one molecule with HMS scores in the range −2 to 0. Among the four methods, the overall best HMS scoring molecule generated was −3.20 using DOCK_SWAP with the HMS/HMS iso-library (Figure 8a). Further, and for HMS < 0 and HMS < −1 in particular, the DOCK_SWAP protocols yield many more molecules within the specified cutoffs (N=48,739 to 53,748 and N=2,096 to 2,598, respectively) compared to DOCK_DN (N=8,690 and N= 493, respectively). From another perspective, Figure S8 compares top-scored HMS poses generated using DOCK_DN with the generic library, and DOCK_SWAP with the HMS/HMS iso-library, to cognate crystallographic poses for 3 protein systems. Taken together, the analysis in Figure 8 and Figure S8 suggests that the DOCK_SWAP torsion protocol does not appear to hamper limit conformational sampling in a practical sense when compared to DOCK_DN. For additional context, the de novo design program OpenGrowth,¹⁴ also saves only the single best scoring rotamer (orientation) for each new fragment that is added to a growing molecules.

Figure 8.

Similarity to 57 crystallographic ligands from DOCK_DN (black) and DOCK_SWAP (red, green blue) simulations in terms of best scoring (a) HMS scoring molecule per system and (b) for three different HMS score cutoff ranges: HMS < 0 (top), HMS < −1 (middle), HMS −2 (bottom). Each group rank-ordered individually to observe the overall trends.

Molecular score analysis (energetic, cheminformatic).

As shown in Figure 9, additional analysis included an examination of ligand-protein interaction energies (DOCK SGE scores),⁴⁵ quantitative estimate of drug-likeness (QED),⁴⁶ and synthetic accessibility (SynthA).⁴⁷ SGE scores provide a measure of protein-ligand compatibility in terms of favorable intermolecular VDW and ES interactions. QED estimates "drug-likeness" for a given molecule and is derived from a scoring function that include key drug-based properties such as MW, LogP, #Hydrogen Bond Donors, #Hydrogen Bond Acceptors, Polar Surface Area, #Rotatable Bonds, #Aromatic Rings, and #ALERTS. SynthA estimates "synthetic accessibility" in terms of organic synthesis for a given molecule and is based on consideration of features such as substructures of previously-known synthesizable structures, structural size, the number of large rings, the presence of fused rings, and stereochemical features. Depending on size and composition, SGE scores for favorably docked molecules typically range from about −100 to −50 kcal/mol. The QED function ranges from 0.0 (worst) to 1.0 (best) and the SynthA function ranges from 1.0 (best) to 10.0 (worst). For each method, all simulations were pooled together (15 anchors x 57 protein-ligand systems = 855 simulations each) and both normalized and non-normalized distributions (shown as insets) are shown.

Figure 9.

Normalized and non-normalized (shown as insets) distributions for (a) SGE, (b) QED, and (c) SynthA using DOCK_DN (black), HMS/HMS (red), VOH/VOH (green) and VOH/TAN (blue) simulation methods. The number of molecules with SGE scores < −50 kcal/mol, QED scores > 0.80, and SynthA scores < 3 are shown below each inset. Normalized distributions generated using the kernel density estimation method.

Normalized SGE distributions (Figure 9a) shows all methods yield a wide range of intermolecular energy scores from negative to positive, and the plots in general look similar with the exception of the DOCK_DN distribution which has a higher population at about −25 kcal/mol (Figure 9a black line). In contrast, the raw non-normalized distributions (Figure 9a inset) indicate all three DOCK_SWAP methods yield dramatically more molecules with more negative SGE scores. For example, at a cutoff of < −50 kcal/mol (Figure 9a dashed vertical line) DOCK_SWAP yields 6.46 (HMS/HMS), 9.75 (VOH/VOH), and 6.14 (VOH/TAN) times more molecules, respectively, than DOCK_DN (raw number of molecules at this cutoff indicate below the inset). Readers should note that the present work employed an SGE pruning score cutoff of 100.0 kcal/mol during on-the-fly construction (DOCK6 default) however a more stringent cutoff could be specified if desired or tailored to the specific scoring function(s) employed. In general, molecules with SGE scores > 0 would most likely not be considered further but in the present analysis they were retained to observe algorithmic behavior.

For QED (Figure 9b), and to a somewhat lesser extent SynthA (Figure 9c), normalized distributions were similarly shaped. As for SGE, non-normalized distributions highlight the dramatic difference in terms of the raw numbers of molecules generated with favorable properties. For example, using a favorable QED score cutoff of > 0.80 DOCK_SWAP generated ~2.7 to 3.1M molecules compared to DOCK_DN which generated 249,416 molecules (Figure 9b insets). The observation here that all four methods trend towards more favorable QED scores is attributed to several factors including use of fragments derived from drug-like molecules combined with enforcement of only previously seen bond types (allowable torsion check),¹⁶ and reasonable MW, #ROT, and FC limits (see Table 2). Further, as recently reported in Duarte and Pak et al,¹⁹ QED scores are in general more favorable when DOCK6 is employed for de novo growth in protein binding sites compared to a vacuum. Similarly, all four methods trend towards molecules having more favorable SynthA scores (Figure 9c, smaller values), although the normalized distributions show DOCK_DN is slightly left-shifted. More importantly, in terms of raw numbers, DOCK_SWAP yields dramatically more molecules with favorable SynthA scores < 3 (N=~3.4 to 4.1M) compared to DOCK_DN (N=662,899 molecules, Figure 9c inset).

Ligand heterogeneity via best-first clustering.

As additional means to probe simulation outcomes, we examined ligand heterogeneity with regards to the number of clusters (and properties) obtained from grouping geometrically overlapping and chemically similar molecules for each set of 57 results. Specifically, we compared the total number of occupied clusters, average cluster size, average number of unique fragments per cluster, and overall SGE score ranges (see Figure S9 and accompanying text in Supporting Information for definitions and results). For computational tractability, a "best-first clustering" approach was employed (see Supporting Information for clustering protocol) given the large numbers of molecules involved (millions) which makes pair-wise calculations impractical. As a representative example, Figure 10 shows best-first clustering outcomes (first 5 clusters) from simulations initiated in acetylcholinesterase (pdb 1EVE)⁴⁸ in terms of 3D pose, cluster size, # unique fragments per cluster, and SGE score range. As would be expected, given the rather stringent cutoff employed for best-first clustering (HMS cutoff = < −3.0), the ensemble of molecules in each of the clusters shows considerable geometric and chemical overlap. Consistent with the overall trends in Figure 9, DOCK_SWAP clusters in Figure 10 contain considerably more molecules, contain more unique fragments per cluster, and have more favorable SGE scores compared to their corresponding DOCK_DN clusters. For example, cluster 1 from DOCK_SWAP yields 34, 34, 41 vs 10 molecules, 46, 41, 44 vs 16 unique fragments, and left-shifted SGE scores (lower limit). Importantly, in all but one case (cluster 2, # unique fragments = 12), a comparison of clusters 2-5 shows the same trend across all three of these attributes.

Figure 10.

Representative best-first clustering results (HMS cutoff < −3.0) from simulations in acetylcholinesterase (pdb 1EVE) using rank-ordered simulation outcomes (SGE rank). Each group shows the 3-D poses, cluster size, number of unique fragments, and range of SGE scores associated with the first 5 clusters. Protein residues hidden for clarity.

Growth path coverage.

An interesting observation which surfaced during clustering analysis involved the identification of pairs of molecules with high similarity, despite the fact they were constructed via different growth paths which originated from different anchor fragments. We long hypothesized that geometrically similar (or even identical) molecules might be generated from independently seeded simulations given the constraints inherently imposed by binding site environments, however, until this study we never documented such occurrences. To help capture such events of what is likely to be rare phenomena, we modified the best-first clustering protocol used in Figure S9 and Figure 10 to group molecules together only if their anchors used for growth were distinct from that used to seed the cluster head (abbreviated CH). Clustering here employed a relatively stringent HMS cutoff of −3.0. We make the distinction between such clusters as a whole (termed anchor unique clusters), and individual molecules (termed "Anchor Unique Analogs" abbreviated as AUAs).

Table 5 shows the total number of anchor unique clusters and AUAs obtained using DOCK_DN with the generic library, or DOCK_SWAP with the HMS/HMS, VOH/VOH and VOH/TAN iso-libraries. As hypothesized, using an HMS cutoff of −3.0, the number of anchor unique clusters is very small (Table 5, N = 230 to 804) compared to the total number of occupied clusters obtained with the standard best-first clustering method (Figure S9, N = 29,377 to 51,146). Correspondingly, the number of AUAs is also low (370 to 1,654). Interestingly, the VOH/VOH iso-library generated fewer anchor unique clusters (N=230) and fewer AUAs (N=370) than the other methods. This could be related to the previous observation that although the VOH/VOH benzene sidechain iso-table showed high geometric conservation in terms of TAIL overlap relative to the HEAD (tight grouping with benzene), the rings were not necessarily chemically similar given the different arrangements of N atoms (see Figure 6 discussion). Thus, even if a given HEAD was accepted using VOH/VOH, some TAILs might not pass the allowable torsion check relative to using other two iso-libraries. As partial support for this hypothesis, the number of unique fragments recovered (rec-20) from Retained molecules (Table 4), was somewhat smaller for VOH/VOH (N=2,185) compared to HMS/HMS (N=2,320) or VOH/TAN (N=2,675), although additional exploration on this topic is warranted.

Table 5.

Number of anchor unique clusters and anchor unique analogs generated using DOCK_DN and DOCK_SWAP methods.

	DOCK_DN	DOCK_SWAP
	generic lib	HMS/HMS iso-lib	VOH/VOH iso-lib	VOH/TAN iso-lib
Anchor unique clusters	504	804	230	671
Anchor unique analogs	643	1,654	370	1,540

Figure 11 compares some interesting CH - AUA pairs identified from simulations initiated in seven protein systems using the HMS/HMS iso-library. The results are organized into four growth path "types" based on different combinations of which anchor "class" (sid, lnk, scf) was used to seed growth and the "position" of anchors in fully grown molecules. Type I = same anchor class, similar 3D position; Type II = different anchor class, similar 3D position; Type III = same anchor class, different 3D position; Type IV = different anchor class, different 3D position. For each example, top and middle rows show the CH or AUA pose, respectively, and their corresponding anchor name and anchor position highlighted with see through magenta ovals. The bottom row shows the pose overlap between the two molecules (CH = red, AUA = green), HMS score, pdb code employed for growth, and cluster number. It is important to emphasize that these results were obtained from independently seeded simulations (protein residues hidden for clarity).

Figure 11.

Representative examples of similar molecules constructed via four growth path types (Type I-IV, see legend and text for definitions) identified from best-first clustering from simulations initiated in four protein systems (Type I from 1AJV; Type II from 3BEL; Type III from 3DLG; Type IV from 1Q84). Top, middle, and bottom rows show, respectively, the CH pose, the AUA pose, and the pose overlap between the constructed molecules (CH = red, AUA = green). Also shown are the anchor name and anchor location (magenta shade) in each final pose. Protein residues hidden for clarity, cluster numbers identified after each pdb code.

Here, the Type I example (1AJV cluster 850) shows that structurally similar molecules with tight pose overlap (HMS score = −3.04) can be obtained even when seeded from a different type of sidechain, for example a hydroxyl (CH = sid.8) or a methoxy (AUA = sid.4) group. The Type II example (3BEL cluster 739) similarly shows that related molecules can be constructed even when seeded from different anchor classes, for example a branched trisubstituted methyl scaffold (CH = scf.9) or ether linker (AUA = lnk.5). The Type III (3DLG cluster 391) and Type IV (1Q84 cluster 422) examples highlight that highly similar chemotypes can be constructed from original seed anchors that reside in distinctly different regions of 3D space in the final poses. For example, the methoxy sidechain (CH = sid.4) and benzene sidechain (AUA = sid.7) in the Type III example, and the ether linker (CH = lnk.5) and ethyl sidechain (AUA = sid.3) in the Type IV example, which are positioned respectively, at the middle or left-hand sides, of the molecules. An analysis of VOH/VOH and VOH/TAN results yielded similar examples (data not shown).

While these CH-AUA outcomes are interesting, and show it is possible to arrive at a similar destination starting from different places, additional studies are warranted. For example, additional HMS cutoffs (or other similarity measures) should be examined, as well as seeding growth with more than 15 fragment types. As the currently identified CH-AUA pairs are relatively small in terms of the number of ligand rotatable bonds it would be worthwhile to determine if growth protocols could be devised that would allow larger molecules to be constructed with high similarity from multiple starting points. If possible, this would give some insight into how well the underlying energy landscape was being searched, as well as provide an additional means to help prioritize candidates in terms of desirable protein-ligand complementarity

(IV) Lead Refinement Example

The DOCK_SWAP infrastructure also provides a straightforward way to quickly construct a congeneric series about a molecular scaffold, in the context of the protein, which can aid drug-lead refinement. As a representative example, we generated a group of analogs based on celecoxib, a nonsteroidal anti-inflammatory drug selective for COX-2 which leads to reduced side effects (e.g. gastrointestinal toxicity).^49,50 Celecoxib contains a pyrazole scaffold decorated with 4-sulfamoylphenyl, trifluoromethyl, and p-tolyl groups at the 1, 3, and 5 positions, respectively. The crystallographic complex of celecoxib with COX-2 (pdb entry 3LN1)³⁴ shows that the inhibitor is in close contact with ~ 14 residues in the binding pocket (Figure 12a). Based on a 3 Å cutoff distance, the larger 4-sulfamoylphenyl group is within H-bonding distance (magenta) of Leu352, Arg513, Phe518 and close to His89, Gln192, Ser353, Ala516, and Val523 (residue numbering follows Wang et al).³⁴ The trifluoromethyl group is surrounded by hydrophobic residues Val349, Leu531, Leu359, as well as Tyr355. In contrast, the p-tolyl group is in close contact only with only two residues (Gly526, Trp387). We therefore elected to refine this position to try and identify analogs that would maintain the existing 4-sulfamoylphenyl and trifluoromethyl interactions but with enhanced DOCK6 scores. For reference, the energy-minimized SGE score using DOCK6 for celecoxib with COX-2 is −76.39 kcal/mol (SGE_VDW = −70.80 + SGE_ES = −5.59).

Figure 12.

(a) Crystallographic pose of celecoxib complexed with COX-2 showing residues within 3 Å of the ligand and H-bonds highlighted in dashed magenta (Chimera⁵¹ definitions). (b) Celecoxib with p-tolyl group removed showing exposed attachment point (AP). (c) Overlay of TAIL fragments in the p-tolyl iso-table used for AP refinement from 20 to 3446 entries.

As shown in Figure 12b, the input file for celecoxib (mol2 format) was modified by deleting the p-tolyl moiety except for the connection atom to the pyrazole which defines the attachment point (AP). Additions to the AP were subsequently explored using DOCK_DN with the standard generic sidechain library (N=217 fragments), or DOCK_SWAP with a customized p-tolyl iso-table sidechain library with the HMS/HMS alignment protocol in which the number of sampled TAIL fragments was significantly expanded from 20 to 3,446 fragments (Figure 12c). Both simulations employed exhaustive fragment selection (i.e. all fragments were attempted). The SGE scoring function (intermolecular VDW plus ES energy) was also augmented with an HMS term (weight = 1) to help enforce that the refined ensemble of analogs remained relatively well overlapped with the celecoxib reference, provided that they were energetically favorable. It is important to note that 10 most favorably ranked entries in the expanded p-tolyl iso-table (Figure 12c) are analogous to the examples discussed previously (Figure 6), in terms of there being very tight spatial overlap with overall similar topology, however as more and more fragments are considered (e.g. 20, 100, 500, 3,446) there is greater and greater chemical and structural variation.

For each simulation, Figure 13 shows the total number of molecules generated, those that scored better than celecoxib (SGE score), and the top five 5 ranked results. As expected, fewer analogs were generated using DOCK_DN (N=74, Figure 13a) with the generic library compared DOCK_SWAP (N= 738, Figure 13b) with the p-tolyl iso-table for which there is likely to be a higher percentage of compatible torsions. From these ensembles, DOCK_DN and DOCK_SWAP yielded 19 and 154 molecules, respectively, with SGE scores more favorable than celecoxib. Importantly, the poses from both methods maintained the overall celecoxib geometry except for the refinement position which was by design. To help enforce such behavior, provided that the analogs were indeed energetically and sterically compatible with the binding site, we employed a scoring function consisting of SGE plus HMS overlap using celecoxib as the reference (weights = 1).

Figure 13.

Refinement results for celecoxib employing (a) DOCK_DN with the generic library and (b) DOCK_SWAP with the p-tolyl iso-table library. Each row shows the total number of analogs constructed, the subset of analogs that scored better than the celecoxib reference, and the top 5 entries. DOCK6 SGE scores and score components in kcal/mol. Protein residues hidden for clarity.

For both simulation methods, the top 3 of 5 DOCK_DN and top 4 of 5 DOCK_SWAP candidates favored a fused-ring sidechain (Figure 13). In terms of interaction energy, the top 5 DOCK_SWAP analogs yielded SGE scores ranging from −86.22 to −82.97 kcal/mol which was more favorable than DOCK_DN which ranged from −81.75 to −79.29 kcal/mol. In both cases, the average of the electrostatic components (SGE_ES) for the top 5 entries were similar to the parent (DOCK_DN = −5.07; DOCK_SWAP = −5.69; celecoxib = −5.59 kcal/mol). Consequently, increases in favorable van der Waals packing (SGE_VDW) dominate the observed improvement in SGE scores, relative to the parent (DOCK_DN avg = −75.27; DOCK_SWAP avg = −78.85; celecoxib = −70.80 kcal/mol).

We also examined analogs to identify those with better scores than celecoxib and were predicted to make more hydrogen bonds in the binding site. As hypothesized, the simulation protocols led to analog poses which maintained the three celecoxib H-bonds at Leu352, Arg513, and Phe518. Importantly, depending on the analogs, additional H-bond interactions were predicted at Tyr385, Met522, Val523, and Ser530. Figures 14a and 14b shows there were 2 analogs from the DOCK_DN simulations and 26 analogs from DOCK_SWAP simulations which met both criteria. Encouragingly, the DOCK_SWAP ensemble increases the number of potential H-bond interaction sites from 3 to 7. The small DOCK_DN ensemble increases the potential H-bond sites number from 3 to 4. Taken together, these favorable outcomes suggest that both sidechain refinement protocols, especially DOCK_SWAP which employs an expanded fragment pool, can be used to generate a variety of better scoring molecules, with enhanced H-bonding potential, while maintaining experimentally known interactions.

Figure 14.

Analogs with SGE scores more favorable than the celecoxib reference predicted to make additional H-bonds with COX-2 from (a) DOCK_DN and (b) DOCK_SWAP simulations. Only the subset of protein residues within H-bonding distance of the ligands are shown (Chimera⁵¹ definitions).

At the suggestion of a reviewer, we searched publicly available databases to determine if any of our refinement results had been deposited and tested, or would be available for purchase to support future experimental studies. We converted all 738 analogs (Figure 13b) from mol2 to SMILES format, and queried three public databases, which include PubChem,⁵² ZINC15¹⁰ and ZINC20 (for-sale category),¹¹ for exact matches in terms of Tanimoto identity. The searches yielded 37 PubChem matches (9 purchasable), 39 ZINC15 matches (6 purchasable), and 17 ZINC20 purchasable matches resulting in 26 "unique" entries. For the 23 top-scoring matches, Figure 15 shows 2D representations, ID codes, rank order of each added sidechain in the p-tolyl iso-table (Sid-rnk), SGE scores (ranked most favorable to least favorable) and, if available, IC50 values obtained from PubChem CID searches. As a reference, the celecoxib parent is also shown in Figure 15 (solid green rounded rectangle, SGE score = −76.39 kcal/mol). Notably, 17 analogs yielded a more favorable SGE score than the celecoxib parent and one of the top-scoring members of the group was a re-construction of celecoxib itself (dashed green rounded rectangle, SGE score = −77.51 kcal/mol).

Figure 15.

DOCK_SWAP analogs showing 2D structure, PubChem (CID) and ZINCID codes, sidechain rank (Sid_rnk), and SGE scores (kcal/mol). Purchasable compounds denoted green rectangles. Parent celecoxib highlighted by solid green rounded rectangle (x-ray pose) and dashed green rounded rectangle (re-constructed parent). IC50 values in μM taken from PubChem CID searches. IC50_A = reference 49, IC50_B = reference 53, IC50_C= reference 54, and IC50_D = reference 55.

In total, out of the 22 "true" analogs in Figure 15, 11 were identified as having been previously experimentally tested against COX-2,^49,53-55 13 were identified as being purchasable from commercial vendors (solid green rectangles), and 9 appear to have not been tested. Top-scoring p-tolyl substitutions include fused rings, 5-membered heterocyclics, and substituted phenyl groups, among others. The lack of correlation between SGE score and sidechain rank (Sid_rnk) provides additional evidence for using larger libraries in de novo design. Encouragingly, among the 7 purchasable analogs constructed by DOCK_SWAP with more favorable SGE scores than celecoxib, all showed significant experimental activity (Figure 15, IC50 values). Curiously, in terms of all the previously tested compounds, the poor IC50 value for CID 10666704 (64.7 μM) led to the compound being classified as inactive in PubChem yet the SGE score was favorable (−74.87 kcal/mol). Additional studies would be worthwhile to determine why this specific pyridine analog was inactive yet exhibited a favorable SGE score. Despite this one outlier, overall, the results in Figures 12-15 strongly suggest that DOCK_SWAP can be a useful tool for users wanting to refine chemical leads in the context of a protein binding site in a semi-automated manner.

Conclusions

The primary objective of this work was development of a new fragment-based selection and search algorithm, employing principles of isosteric swapping, to further improve upon the established de novo design infrastructure in DOCK6 (DOCK_DN). The new method, termed DOCK_SWAP, leverages an improved fragment library infrastructure which expands the available chemical search space in a rational manner. A key component of the infrastructure was the development of iso-tables, comprised of fragments pre-aligned and rank-ordered to a reference, for which a collection defines a specific iso-library. This infrastructure enables DOCK_SWAP to easily select groups of topologically/chemically related fragments for potential attachment (or replacement) during molecular construction.

A central benefit of the method is significant expansion of the fragment libraries, which in the present work are larger by about an order of magnitude compared to our earlier de novo design studies.^15,16,19,23 This leads to dramatically more molecules being constructed and classified as "Retained" and thus there is a concomitant increase in the number of compounds with more favorable energetic, drug-likeness, and synthetic accessibility scores. Although DOCK_SWAP is considerably slower than DOCK_DN given the larger iso-libraries, the program appears to be more efficient than DOCK_DN in terms of the efficiency of fragment attachment. Interestingly, DOCK_SWAP simultaneously yields enhanced variety in terms of generating more molecular clusters compared to DOCK_DN, while at the same time yielding clusters containing higher numbers of more similar molecules. DOCK_SWAP also appears to be able to construct highly similar molecules with greater frequency than DOCK_DN, starting from different seed anchors (which yields different growth paths) which indicates an enhancement of chemical searching. The DOCK_SWAP infrastructure naturally lends itself as a robust tool for lead-optimization as shown by a proof-of-concept refinement of the drug celecoxib using a p-tolyl substitution iso-library. Notably, multiple DOCK_SWAP analogs were identified in publicly available databases, for which a subset showed experimental activity against COX-2, further highlighting the methods ability to construct sensible molecules.

In conclusion, the new DOCK_SWAP method outlined in this manuscript is part of our group’s long-term goal of developing well-validated tools for small molecule de novo design. The iso-table and iso-library infrastructure presented here provides a rational way to expand the palette of fragments available for chemical searching and is analogous in spirit to how a medicinal chemist might replace chemical moieties in a molecule. Importantly, users can easily generate their own custom iso-libraries. DOCK_SWAP will be made available to the community through an upcoming DOCK release, to coincide as closely as possible with publication of this manuscript and will include all source code. It is also important to note that DOCK_SWAP is fully compatible with our recent integration of the program RDKit with DOCK6,¹⁹ thereby enabling cheminformatics descriptors to be computed during construction to bias growth towards specific properties. We are currently working to adapt DOCK_SWAP (and RDKit) for use with our genetic algorithm.¹⁵ We hypothesize that the ability to perform on-the-fly isosteric swapping while simultaneously using cheminformatics descriptors to steer evolution over multiple generations will be a particularly powerful tool for lead discovery and refinement.

Data Availability Statement.

The data that support the findings of this study are available in the Supporting Information of this article.