来源：药智网公众号

摘要：应用计算机模拟靶点预测算法，研究骨形态发生蛋白(Bone Morphogenetic Proteins，BMPs)的全新激动剂三唑并[1,5-c]喹唑啉类化合物，预测其作用靶点、信号通路、可治疗疾病，并使用AI分子对接功能预测靶点与配体的结合情况。

最新研究表明，三唑并[1,5-c]喹唑啉类化合物具有双重靶向CSNK1（酪蛋白激酶1，CK1）和PI3K异构体的独特机制，可以有效放大成骨BMP信号。本研究旨在验证NPAI Engine 天然产物AI研发平台的准确性，预测结果可以帮助设计更精确的靶点鉴定实验，并节省研发资金，缩短研发周期。

本文的创作来自NPAI Engine 天然产物AI研发平台团队，创始人何山是英国伯明翰大学计算机学院副教授、网络药学和机器学习方面的全球领军学者，在Google Scholar中相关领域的论文引用数名列世界前十，团队所做研究对天然产物和人工智能制药技术领域有较高研究和参考价值。

如想试用文中使用的天然产物AI研发平台，可点击链接获取：https://db.yaozh.com/zhuanti/zhuanti_20220518/?yaozh/?shouyedaohang

In-silico target deconvolution: First-In-Class Bone Morphogenetic Protein Amplifier Chemotype

1.Introduction

Bone Morphogenetic Proteins (BMPs) are a group of growth factors (also known as cytokines or metabologens), which play multiple roles during embryonic development. Recent studies of BMPs revealed their clinical potential in various diseases such as bone repair, vascular syndromes and cancer[4]. In a recent Journal of Medicinal Chemistry paper (online date 8th Nov. 2022) [6], the authors employed a phenotypic drug discovery approach to discover a new-in-class in-vitro and in-vivo active amplifier of the BMPs: triazolo[1,5-c]quinazolines. To deconvolute the target, the authors adopted a holistic approach which consisted of Cell Painting assay(CPA), kinome and transcriptomic profilings of 408 kinases. These experiments showed that triazolo[1,5-c]quinazolines present a novel chemotype with a unique mechanism of dual targeting of CSNK1 (Casein Kinase 1, CK1) and PI3K isoforms, which efficiently amplified the osteogenic BMP signalling.

In this case study, we will investigate the novel chemotype triazolo[1,5-c] quinazolines using our in-silico target deconvolution algorithms to validate DrugEngine performance. Using the assumption that we only know the phenotypic screening results demonstrating amplification, we can investigate DrugEngine’s in silico predictions and compare them with the targets (CSNK1 and PI3K isoforms) identified by experimental work[4]. Furthermore, can further insights into the mechanism of BMP activation be gained, such as new targets and signalling pathways [4] These predictions could be used to design more precise target deconvolution experiments with less money and time.

To conduct the in-silico target deconvolution in DrugEngine, we only need the SMILES of the most potent compound 1a:

NC1=NC2=CC=C(Cl)C=C2C3=NC(C4=CC=CO4)=NN13 (see fig.1).

We then execute Bioactivity Prediction and AI Blind Docking to deconvolute the targets of 1a, and the results are detailed below:

Figure 1: Two-dimensional structure of compound 1a

2.Bioactivity Prediction

2.1 CSNK1δ is the top-ranked single protein target

Table 1 lists the top-ranked single protein targets predicted of compound 1a by DrugEngine. Since 1a was originally designed as a non-selective adenosine receptor antagonist [6], not surprisingly, 1a was tested against 10 protein targets including 4 four adenosine receptor types, and the results are in ChEMBL database. DrugEngine automatically retrieved their ChEMBL database records as evidence to support the prediction. However, CSNK1δ (casein kinase 1 isoform delta) was not in the ChEMBL database but ranked 8th by our DrugEngine. In [6], CSNK1δ was deconvoluted as the target of 1a (IC50=0.24).

Table 1: Top 10 ranked targets for compound 1a

2.2 Multi-target disease enrichment analysis deconvoluted PIK3CA and PIK3CB

Based on the predicted targets, DrugEngine predicts the diseases modulated by 1a. By matching these predicted disease phenotypes with the phenotypic screening results, we can then identify those targets that lead to the phenotypic changes. To recapitulate the BMP-relevant cellular context, in [6], the phenotypic screening was performed on a well-established Osteosarcoma cell line U2OS. Indeed, Osteosarcoma was ranked by DrugEngine at 35th among the top 100 predicted disease phenotypes. By inspecting the 35 Osteosarcoma-relevant targets predicted by DrugEngine, we found PIK3CA and PIK3CB as identified in [6].

3. AI Docking

We then submitted the top-10 single protein targets (in total 16 proteins) and all 35 targets enriched in Osteosarcoma for AI Blind Docking.

Figure 2: Predicted bound conformation of compound 1a in CSNK1δ (PDB ID: 7NZY).

Figure 3: Visualisation of residues that compound 1a interacts with in CSNK1δ (PDB ID: 7NZY).

Table 2: Predicted contact residues with 1a in the ATP binding site of CSNK1δ (PDB ID: 7NZY).

3.1 CSNK1δ is the top-ranked target among the single protein targets of 1a

Based on our proprietary reverse docking algorithm, Casein kinase I isoform delta (CSNK1δ) was ranked 8th. It is reassuring to see 5 out of the 7 ranked higher than CSNK1δ are known targets of 1a such as adenosine receptors [6]. The predicted binding mode of 1a to CSNK1δ is illustrated in fig. 2, the predicted residue interactions are given in table 2, and visualised in fig. 3. As shown in table 2, the predicted binding mode is similar to the actual binding mode revealed by the crystal structure of 1a-bound CSNK1δ at 1.85 Å resolving in [6] (PDB: 7NZY), i.e., 1a binds to the ATP pocket of CSNK1δ.

3.2 PIK3CB was ranked 2nd among the 35 targets of Osteosarcoma

The AI Reverse Docking results show that PIK3CA was ranked 2nd. The other PI3K isoform, PIK3CB is PIK3CB was ranked 16th.

AI Blind Docking predicts that compound 1a is bound to the ATP-binding pocket of the p110α kinase catalytic domain of PI3Kα, which is a highly conserved lipid kinase ATP binding site. In [6], instead of using a co-crystal structure, the binding mode of 1a was investigated using the molecular docking software Glide [3]. The docking identified the key interactions such as π -stack with T780 and Hydrogen bonds with V851 (see table 3), which confirm our AI Blind Docking results. To further validate our docking results, we compared the binding mode of a PI3Kα inhibitor YXY-4F determined by the crystal structure [7] (PDB ID: 5XGH). Similar to YXY-4F, 1a forms Hydrogen bonds with V851, hydrophobic interactions with I932, π-stack with T780, and π-stack and π-cation interaction with T836.

Table 3: Predicted contact residues with 1a in the ATP binding site of PI3Kα (PDB ID: 5XGH).

3.3 PTH1R is likely modulated by 1a for BMPs activate amplification

Apart from the aforementioned CK1 and PI3K isoforms, DrugEngine also predicts PTH1R (Parathyroid hormone-related peptide receptor) as a top target (ranked 7th) which was not reported in [6] but worth investigating. PTH1R is a class-B GPCR (G protein-coupled receptor) that regulates skeletal development and bone turnover by enhancing BMPs’ activity [5, 8]. The first crystal structure of human PTH1R in complex with a peptide agonist at 2.5-Å resolution was published in 2018 [1, 2]. However, there is no crystal structure of human PTH1R in complex small-molecule agonists. To predict whether 1a modulates PTH1R unbiasedly, we used our AI Blind Docking to dock 1a to Alphafold prediction of crystal structure (PDB ID: alphafold-Q03431) without any binding pocket information.

Table 4: Predicted contact residues with 1a in the ATP

binding site of PTH1R (PDB ID: alphafold-Q03431).

Our AI Blind Docking results suggested that 1a binds to the orthosteric pocket of PTH1R in the TMD (transmembrane domain), where residues R233, F288, H420 and Q451 form a central polar network as shown in [1, 2]. This previously defined central polar network is critical for PTH1R activation. Our results show that 1a forms π -cation interactions with residue R233, which is in direct contact with Q451 via a hydrogen bond (refer to table 4 for a full list of predicted residues and fig. 4for a three dimensional visualisation of 1a in the predicted binding site for PTH1R). This key interaction stabilises the orientation of Q451, a conserved residue within the central polar network that appears to act as a molecular switch between receptor activation states [1, 2].

Although this novel hypothesis, i.e. 1a modulates PTH1R is worth investigation, caution should be taken due to [1, 2]:

Figure 4: Visualisation of residues that compound 1a interacts with in PTH1R (PDB ID: alphafold-Q03431).

Reference List

[1] Janosch Ehrenmann et al. “High-resolution crystal structure of parathyroid hormone 1 receptor in complex with a peptide agonist” . In: Nature Structural & Molecular Biology 25.12 (2018), pp. 1086–1092.

[2] Janosch Ehrenmann et al. “New views into class B GPCRs from the crystal structure of PTH1R” . In: The FEBS Journal 286.24 (2019), pp. 4852–4860.

[3] Richard A Friesner et al. “Glide: a new approach for rapid, accurate docking and scoring.1. Method and assessment of docking accuracy” . In: Journal of medicinal chemistry 47.7 (2004), pp. 1739–1749.

[4] T Katagiri and T Watabe. Bone morphogenetic proteins. Cold Spring Harb Perspect Biol. 2016; 8 (6): a021899.

[5] Yoshihiro Nakao et al. “Parathyroid hormone enhances bone morphogenetic protein activity by increasing intracellular 3’, 5’-cyclic adenosine monophosphate accumulation in osteoblastic MC3T3-E1 cells” . In: Bone 44.5 (2009), pp. 872–877.

[6] Fabian Wesseler et al. “Phenotypic Discovery of Triazolo [1, 5-c] quinazolines as a First-In-Class Bone Morphogenetic Protein Amplifier Chemotype” . In: Journal of Medicinal Chemistry (2022).

声明：本文仅作信息传递之目的，并不代表本平台观点。本文也不是治疗方案推荐，如需获得治疗方案指导，请前往正规医院就诊。如涉及作品内容、版权和其它问题，请在本平台留言，我们将在第一时间核实并处理。