Drug Design: Methodology, Concepts, and Mode-of-Action SpringerLink

Table Of Content

• Advancements in small molecule drug design: A structural perspective
• Pharmacophore-Based Drug Design Approach¶
• TransformerCPI2.0 model and training details
• How did your time at the USC Mann School impact you?
• Temple University
• 1. Target Selection and Validation: Possible Expansion of Chemical Space
• Strengths of Receptor Based Drug Design¶

Protein embedding and atom embedding serve as the target sequence and memory sequence of the transformer decoder, respectively. Consistent with the encoder, the decoder consists of 3 decoder layers, 8 attention heads for each layer, 768 dimensions for the hidden state, and 3072 dimensions for feedforward layers. In addition, the original transformer was designed to solve seq2seq tasks and utilize a causal mask operation to cover the downstream context of each word in the decoder. We removed the mask operation of the decoder to ensure that our model accesses the whole target sequence. Since we introduced a new virtual atom, as described above, we used the last layer representation of this virtual atom rather than the weighted sum of the last layer atom representation to predict the compound protein interaction probability. The last layer presentation of virtual atoms was fed into fully connected layers and finally returned the compound protein interaction probability.

Advancements in small molecule drug design: A structural perspective

Recent advancements have focused on studying molecular interaction networks, known as interactomes, which encompass various types of interactions such as protein-protein interactions, drug-target interactions, and drug-drug relationships. Analyzing these interactomes enables the prediction of previously unknown interactions and provides insights into the network topology24,25,26,27. Studying molecular interaction networks as a holistic entity offers a distinct advantage by allowing the analysis of long-range relationships between different nodes that are connected through multiple edges. This approach enables a comprehensive examination of the interconnectedness and dependencies among various components within the network24.

Pharmacophore-Based Drug Design Approach¶

Neural multi-task learning in drug design - Nature.com

Neural multi-task learning in drug design.

Posted: Tue, 20 Feb 2024 08:00:00 GMT [source]

230D7 specifically inhibited the growth of ccRCC cell lines with an IC50 of approximately 20 μM compared with non-ccRCC cell lines. To determine if 230D7 is suitable for in vivo studies, we investigated the pharmacokinetics and acute toxicity profile of 230D7. 230D7 can be efficiently absorbed into the blood circulation after intraperitoneal injection and has low acute toxicity (Supplementary Fig. 2i–l). A dose-dependent reduction in 786-O tumor growth rate could be observed in NSG mice treated with 230D7 (Fig. 5i), revealing a significant anti-ccRCC therapeutic effect of 230D7 in vivo. Statistically, no body weight loss was observed in NSG mice throughout the entire pharmacodynamics study of 230D7 (Supplementary Fig. 2m). Finally, we checked the effect of 230D7 on oncogenic SPOP signaling in ccRCC xenograft tumors.

TransformerCPI2.0 model and training details

The key advantage of database searching is that it saves synthetic effort to obtain new lead compounds. Another category of structure-based drug design methods is about “building” ligands, which is usually referred as receptor-based drug design. In this case, ligand molecules are built up within the constraints of the binding pocket by assembling small pieces in a stepwise manner.

PocketFlow is a data-and-knowledge-driven structure-based molecular generative model

Cells were seeded in 96-well plates and incubated with serially diluted compounds for 72 h. Cell viability was determined using the CellTiter-Glo® Luminescent Cell Viability Assay kit (Promega, G7573) following the manufacturer’s instructions. IC50 values were determined by nonlinear regression (curve fit) using a variable slope (four parameters) in GraphPad Prism (9.0).

The Difference Is the Data: Drug Discovery's AI Revolution - Genetic Engineering & Biotechnology News

The Difference Is the Data: Drug Discovery's AI Revolution.

Posted: Thu, 25 Apr 2024 11:00:20 GMT [source]

How did your time at the USC Mann School impact you?

There are many intermediate stages in the development of the project, which are constantly changing according to newly generated information. For the purpose of finding highly active pyrazole amide compounds, Jin-Xia Mu, Xing-Hai Liu, Bao-Ju Li, and their coworkers designed and synthesized a series of novel pyrazole amide derivatives by multi-step reactions from phenylhydrazine and ethyl 3-oxobutanoate as starting materials. They characterized the structures and antifungal activities of the title compounds and used DFT calculations to study the structure-activity relationships.

Finally, deep learning-based methods for predicting drug-target binding sites follow a similar approach to the feature extraction methods mentioned above for drug-target interaction and affinity prediction. It began by constructing a graph based on the structure of target and extracting the sequence features from target sequence. These features were used as node features in the graph of target, obtained through a pre-trained model. Graph Transformer was then used to extract structural features from the graph of target.

1. Target Selection and Validation: Possible Expansion of Chemical Space

By enabling the swift translation of advanced research outcomes into practical application tools, we strive to expand the scope of application for these research results. Affiliated research centers include the Center for Biomolecular Structure and Dynamics, the Center for Environmental Health Science, and the Center for Structural for Functional Neuroscience. The Department of Pharmacology and Pharmaceutical Sciences is an international leader in pharmaceutical research—with particular strengths in the areas of molecular mechanisms of disease and drug design, development, targeting and delivery. The department has garnered millions of dollars in support from the National Institutes of Health and other government and private funders.

The signals of all reaction systems were continuously monitored and recorded from 25 °C to 90 °C for approximately 45 min. The Tm values of SPOPMATH and ARF1 were measured using CFX manager software version 3.1. Fluorescence polarization experiments were conducted in a 384-well black plate (Corning, 3575) using a 42 μL reaction system.

The best starting point in receptor-based drug design is to have the X-ray structure of the target protein complexed with a ligand. If this is not available, one can generate such information by modeling techniques from the primary sequence of the target protein and by homology with the X-ray structures of homologous proteins. Despite our ignorance of anchorage points and the exact interactions that occur with the receptor site, pharmacophore-based drug design has the advantage of guiding the discovery process in projects where the receptor is not known. Moreover, the chances of success are greater when we start with something that looks like a solution (active molecule) than by starting from the problem (protein structure).

The chemical formulas of a series of biologically active molecules are known and act with the same mechanism of action. Based on these structures the aim of the project is to create novel proprietary structures with improved profiles. This is a typical starting point of a pharmacophore-based approach aiming at the creation of new lead molecules mimicking the structures of a known reference series. The ideal situation is to have X-ray structures of complexes between the active compounds and the target protein. However even when the X-ray data is not available, one can predict and construct a model of the binding mode of a ligand by using modeling techniques.

The TransformerCPI2.0 model was trained by the RAdam71 optimizer with a learning rate of 1e-5 and a weight decay of 1e-3. The batch size of 1 was selected to ensure that the longest protein sequence fit into the GPU memory. We employed the gradient accumulation technique to expand the actual batch size to 64. The TransformerCPI2.0 model was trained for ~50 epochs or ~1.5 weeks of wall clock time. Many other AI tools and platforms for drug discovery and development are available in the web and new ones are inconstantly appearing.

Furthermore, using the same evaluation criteria, ligand-based design was compared to structure-based design, with ligand-based design applications outperforming structure-based models in all investigated scenarios (Table 2, Tables S4–S6). Several powerful new tools or improvements of already used tools are now available to medicinal chemists to help in the process of drug discovery, from a hit molecule to a clinically used drug. Among the new tools, the possibility of considering folding intermediates or the catalytic process of a protein as a target for discovering new hits has emerged. In addition, machine learning is a new valuable approach helping medicinal chemists to discover new hits.

The correlations between molecular descriptors of a series of compounds and their activities are assigned as Quantitative Structure Activity Relationships (QSAR). The 1D descriptors are derived from the 1D structure such as element composition and molecular weight. The 2D descriptors are generated from the molecular graph such as the number of hydrogen bond acceptors and donors, the partition coefficients logP and logD, water solubility, etc. Carracedo-Reboredo et al. define 4D descriptors, providing information about the interactions between ligands and protein-binding sites [32]. These are the descriptor sets which describe a complex property or a set of properties. The target macromolecule is an internal molecule which is involved in the disease.

Subsequently, a multi-layer perceptron (MLP) was used to predict whether the nodes in the graph corresponded to binding sites. Another method, SiteRadar [29], utilized graph machine learning to accurately predict the binding pockets (sites) between drugs and targets. However, it is worth noting that in this study, QSAR models were employed for scoring, and these models were trained on existing ligand activity data. The successful machine learning from the relevant training data is evident in the discovery that compound 1 interacts with the receptor in the canonical binding mode53, as evidenced in the crystallographic complex.

This interaction can be depicted as a graph, where nodes represent bioactive ligands and their corresponding macromolecular targets (Fig. 1a). Distinct nodes were used to differentiate between orthosteric and allosteric binding sites within the same target. Edges were established between ligands and proteins that have an annotated binding affinity of less than or equal to 200 nM (Fig. 1a) (values extracted from the ChEMBL database28). As a result of this procedure, an interactome was generated that consisted of ~360,000 ligands, 2989 targets, and around 500,000 bioactivities. In the case of structure-based design, only macromolecular targets with known 3D structures were considered, resulting in an interactome containing around 208,000 ligands, 726 targets, and around 263,000 bioactivities. This data structure based on the interactome facilitated the training of two deep learning models, specifically for ligand-based and structure-based de novo design (Fig. 1b).