DNA Barcoding a Complete Matrix of Stereoisomeric Small Molecules

Abstract

It is challenging to incorporate stereochemical diversity and topographic complexity into DNA-encoded libraries (DELs) because DEL syntheses cannot fully exploit the capabilities of modern synthetic organic chemistry. Here, we describe the design, construction, and validation of DOS-DEL-1, a library of 107 616 DNA-barcoded chiral 2,3-disubsituted azetidines and pyrrolidines. We used stereospecific C-H arylation chemistry to furnish complex scaffolds primed for DEL synthesis, and we developed an improved on-DNA Suzuki reaction to maximize library quality. We then studied both the structural diversity of the library and the physicochemical properties of individual compounds using Tanimoto multifusion similarity analysis, among other techniques. These analyses revealed not only that most DOS-DEL-1 members have "drug-like" properties, but also that the library more closely resembles compound collections derived from diversity synthesis than those from other sources (e.g., commercial vendors). Finally, we performed validation screens against horseradish peroxidase and carbonic anhydrase IX, and we developed a novel, Poisson-based statistical framework to analyze the results. A set of assay positives were successfully translated into potent carbonic anhydrase inhibitors (IC₅₀ = 20.1-68.7 nM), which confirmed the success of the synthesis and screening procedures. These results establish a strategy to synthesize DELs with scaffold-based stereochemical diversity and complexity that does not require the development of novel DNA-compatible chemistry.

Fig. 1 |. Design and synthesis of DOS-DEL-1.

a, The eight azetidine- and pyrrolidine-based scaffolds that constitute the basis of DOS-DEL-1 can be accessed from commercially available chiral α-amino acids. b, Synthesis of NHS esters 4a-4d for scaffold attachment to DNA. Azetidine 2a and pyrrolidine 2b were synthesized from D-azetidine-2-carboxylic acid (1a) and D-proline (1b), respectively according to the procedures described by Maetani et al. Reagents and conditions can be found in the Supporting Information. c, Synthesis of DOS-DEL-1. After scaffold attachment (via NHS esters 4a-4d and ent-4a-4d) to a double-stranded DNA “headpiece” (DNA-HP; Fig. S1), the azetidines and pyrrolidines undergo Boc deprotection separately. Then, all eight scaffolds are pooled for two rounds of DNA-encoded split–pool synthesis: 114 N-capping reactions and 118 Suzuki reactions. The resulting library has three sources of diversity, comprises 107,616 unique compounds, and includes all possible stereoisomers of the azetidine and pyrrolidine scaffolds (as indicated by the red balls)

Fig. 2 |. DOS-DEL-1 synthetic decisions influence small-molecule physicochemical properties.

a, DOS-DEL-1 sub-libraries corresponding to divergent synthetic decisions. Subsets of DOS-DEL-1 may be defined according to the size of the heterocyclic scaffold (azetidine=“az” or pyrrolidine=“pyr”), the identity of the N-capping reagent (aldehyde=“ald” or sulfonyl chloride=“sc”), or the functionalization of the aryl iodide (Suzuki=“Suz” or no Suzuki=“no Suz”). Red balls indicate both possible configurations at that position. b, Distribution statistics for physicochemical properties corresponding to Lipinski’s “rule of 5” (ref. 47) or Veber’s bioavailability guidelines (ref. 48). Values depict mean ± SD, median, or % compliance with “rules” across all members of the library or indicated sub-library. c, Physicochemical property distributions of DOS-DEL-1 sub-libraries. Histogram bins are colored according to adherence to (blue) or violation of (red) Lipinski’s (MW < 500 Da; LogP ≤ 5) or Veber’s (TPSA < 140 Å) guidelines. Bin height or depth represents the fraction of the sub-library’s compounds that fall within the bin’s boundaries. Sub-libraries that represent divergent synthetic decisions are juxtaposed to highlight the consequences of those decisions on physicochemical properties.

Fig. 3 |. DOS-DEL-1 members more closely resemble compounds from diversity-oriented synthesis than those from other sources.

Tanimoto-based multi-fusion (average vs. maximum) similarity representations. a, Comparison of DOS-DEL-1 to itself (black) and to each of 5 published small-molecule collections: 6,152 commercial compounds (red), 2,477 natural products (green), and 6,623 diversity-oriented synthesis compounds (blue) from ref. 36.; plus 12,011 known actives and PubChem screening hits (magenta) and 19,149 diversity-oriented synthesis compounds (cyan) from ref. 51. b, Comparison of 8 library subsets defined by azetidine (square) or pyrrolidine (circle) skeleton, aldehyde (open) or sulfonyl chloride (filled) N-capping, and functionalization (large) or not (small) of the aryl iodide using a Suzuki reaction, to the same published collections. For reference, the intra-library multi-fusion similarities among all DOS-DEL-1 members are represented as a mean (black circle) ±3 SD (black bars) of multi-fusion (median vs. maximum) similarities.

Fig. 4 |. The distribution of DOS-DEL-1 barcodes is more accurately modeled by a Poisson-derived frequency distribution than by a negative-binomial distribution.

The empirical read-count distribution (top) for DOS-DEL-1 is compared to a frequency distribution created by sampling from a set of Poisson-distribution models representing each tag, fit using all compounds containing that tag (middle). The corresponding negative-binomial distribution (bottom) is shown for comparison, which fails to account for tags that were inefficiently incorporated (red arrows) and appears to understate the most efficient tagging (red bars).

Fig. 5 |. DOS-DEL-1 screens using test proteins reveal preferential binding as a function of building blocks and building-block combinations.

a, Normalized fold-change plot for DOS-DEL-1 members binding horseradish peroxidase as a function of input representation rank, and depicting compounds containing select significantly enriched electrophilic building blocks (blue: strong; green: intermediate; gold: weak). b, Enrichment plot for each of 114 N-capping building blocks (building block 1), expressed as standardized residuals (χ) relative to expected occurrence among compounds within the top 2,389 normalized fold-change scores. c, Enrichment scores, expressed as standardized residuals (χ) by size, for co-occurrence of pairs of building blocks relative to expected co-occurrence among compounds within the top 2,389 normalized fold-change scores. d, Normalized fold-change plot for DOS-DEL-1 members binding carbonic anhydrase IX as a function of input representation rank, and depicting compounds containing select significantly enriched individual building blocks (teal: primary para sulfonamide; magenta: primary meta sulfonamide). e, Enrichment plot for each of 119 Suzuki adduct building blocks (building block 2), expressed as standardized residuals (χ) relative to expected occurrence among compounds within the top 1,120 normalized fold-change scores. f, Enrichment scores, expressed as standardized residuals (χ) by size, for co-occurrence of pairs of building blocks relative to expected co-occurrence among compounds within the top 1,120 normalized fold-change scores. In panels 5c and 5f, all three combinations of pairs of building blocks (8 × 114, 8 × 119, 114 × 119) are presented on the “faces” of a single 3D plot for compactness.

Fig. 6 |. Biochemical assay data are consistent with CA-IX screening results.

a, Chemical structures of off-DNA compounds synthesized to assess screening results. Azetidines 5a, 5b, 6a, 6b, 7a, and 7b correspond to on-DNA assay positives; the remaining compounds facilitated deeper validation of the screening data and the success of the on-DNA chemistry. Asterisks indicate atoms of variable stereochemical configuration (as defined below each structure). Synthetic details are provided in the Supporting Information (see Fig. S46; Supporting Information: “Synthesis of off-DNA CA-IX hits”). b, In vitro carbonic anhydrase inhibition assay results. Compounds depicted in Fig. 6a were subjected to a colorimetric carbonic anhydrase inhibition assay. IC₅₀ values and 95% confidence intervals are shown. Statistics were calculated using GraphPad Prism software