An Introduction to PROTACs ¶

David Zhang and David Figueroa ¶

PROTACs represent an emerging wave of new therapeutics capable of modulating proteins once thought nearly impossible to target. By eliminating rather than merely inhibiting disease-causing proteins, PROTACs address the limitations of existing drug modalities. As PROTACs progress through clinical trials with promising results, there is growing anticipation surrounding their therapeutic potential.

This DeepChem tutorial serves as a starting point for exploring the world of PROTACs and the exciting field of targeted protein degradation. The tutorial is divided into five partitions:

1. Background literature
2. Data extraction
3. Featurization
4. Model deployment
5. References

With that in mind, let's jump into how we can predict efficacy of PROTAC degraders!

1. Background literature ¶

Traditional drug modalities, such as small-molecule drugs or monoclonal antibodies, are limited to certain modes of action, like targeting specific receptors or blocking particular pathways. Targeted protein degradation (TPD) represents a promising new approach to modulate proteins that have been traditionally difficult to target. TPD has given rise to major classes of molecules that have emerged as promising therapeutic approaches against various disease contexts.

1.1 Targeted protein degradation ¶

Targeted protein degradation represents a way of leveraging the cell's natural degradation mechanisms to target disease causing proteins. Typically, the cell maintains protein homeostasis through clearance by proteasomes or lysosomes. By leveraging these intrinsic cellular mechanisms, TPD methods can target a variety of proteins throughout the cell. This has given rise to a collection of TPD methods aimed at degrading proteins that may play roles across many disease states. One of these major class of molecules is proteolysis-targeting chimera (PROTACs).

1.2 How do PROTACs work? ¶

PROTAC molecules are ternary structures consisting of a linker, a ligand to recruit and bind to the target protein, and a ligand to recruit the E3 ubiquitin ligase. Before we dive into how PROTACs mediate this degradation mechanism, it is crucial to understand the underlying biological pathway that makes this all possible.

Figure 1: Molecular structure of PROTACs molecules designed to inhibit epidermal growth factor receptor (EGFR). The PROTAC linker connects the EGFR ligand and E3 ligase which are highlighted in yellow and gray, respectively [1] .

1.1.1 Ubiquitin Proteasome System ¶

The ubiquitin proteasome system, or UPS for short, is a crucial cellular maintenance mechanism. Ubiquitin-dependent proteolysis is a three-step process which involves ubiquitin-activating enzymes (E1), ubiquitin-conjugate enzymes (E2), and ubiquitin-protein ligases (E3). In general, E1 activates ubiquitin, priming it for transfer to E2 which interacts with E3 at which point E3 ligases are responsible for binding of the target protein substrate for subsequent ubiquitination by E2. Once the protein is tagged with a polyubiquitin chain, it is recognized by the proteasome, a large protease complex that degrades that protein into peptides.

Figure 2: The ubiquitin proteasome system is one of the cell's internal degradation mechanism crucial for targetting dysfunctional proteins. Naturally this opens up opportunities to leverage this in a therapeutic context [2] .

1.1.2 Connection to PROTACs ¶

The realization that the UPS could be leveraged for therapeutic purposes was initially made through early studies of viruses and plants. The underlying idea involves design of small molecules capable of recruiting the E3 ligase and inducing degradation of a protein of interest (POI). This general idea naturally extended itself to the case of PROTACs. Together, the POI ligand, linker, and E3 ligase ligand make up the PROTAC complex responsible for protein degradation. Note that the presence of two ligands enables simultaneous recruitment of the E3 ligase and POI, hence its heterobifunctionality property.

Furthermore, after the POI is degraded by the proteasome, PROTACs can disassociate and continue to induce further degradation, enabling low concentrations to be efficacious. This catalytic mechanism of action and event-drive pharmacology prevents PROTACs from suffering the same limitations as conventional therapeutic strategies such as drug resistance and off-target effects.

Figure 3: The mechanism of action of PROTACs center around the UPS. In a heterobifunctional manner, recruiting both a target protein of interest and an E3 ligase, PROTACs are able to promote protein degradation in diseases [3] .

1.3 Molecular Glues ¶

It is worth noting that PROTACs are not the only TPD method. Another major class which which also leverages the UPS to elicit degradation are molecular glues. As implied by its name, molecular glues stabilize protein-protein interactions between the target protein and E3 ligase. Notice how this is different than PROTACs which consists of two separate binding ligands connected by a linker. Consequently, molecular glues may be less sterically hindered without the need for a linker. However, identifying binding sites to induce new protein-protein interactions is typically harder than accomodating for existing ligands, as in the case for PROTACs, making it harder to design molecular glues.

Figure 4: Lenalidomide is a molecular glue which mediates the interaction between CRBN, an E3 ligase, and CK1α, resulting in subsequent ubiquitination. [4]

1.4 How can we leverage machine learning? ¶

As a novel and promising technique, PROTACs have demonstrated positive clinical results thus far. However, much of the clinical validation has been against classically drugged targets. In order for PROTACs to reach their full potential, their clinical efficacy against novel or hard to reach targets must be demonstrated. Consequently, there has been growing research in designing PROTAC molecules capable of elicting an effective response. However, much of the current work is empirical and requires extensive trial-and-error processes. Machine learning could potentially revolutionize this. By correlating molecular structure with physiochemical properties and biological activity, we could potentially streamline the discovery process, significantly reducing the time and cost associated with validation. With that in mind, let's jump into this tutorial to predict efficacy of PROTAC degraders!

For a more in-depth dive into PROTACs, ubiquitin proteasome system, and targeted protein degradation, readers are referred to [5] and [6] .

2. Data extraction ¶

Before we proceed, let's install deepchem into our colab environment.

Now let's download this dataset on PROTACs, curated by [7] , which includes 3270 PROTACs.

Note that there exists a many-to-many mapping between PROTAC compounds and target proteins. A single PROTAC compound can be designed to target multiple proteins, and conversely, multiple PROTAC compounds can be developed to target the same protein. This many-to-many relationship allows for greater flexibility and adaptability in the design and application of PROTACs.

Taking a closer look at the dataset, each PROTAC compound has a SMILEs representation along with its target protein of interest and E3 ligase. For reference, here is an example:

In general, the PROTAC-DB dataset contains information for a variety of different physiochemical and biochemical properties of PROTAC structures. Several useful ones to point out are $ΔG$, which describes the spontaneity of a chemical reaction, $K_d$ which measures the concentration of a ligand to achieve 50% occupancy of the protein binding sites, and $XLogP3$ which measures a compound's solubility, an indication of its absorption and distribution characteristics.

Before we proceed, let's plot the distribution of each of these properties to get a better sense of our PROTAC dataset starting with ΔG values.

Let's take a closer look at the distribution of PROTAC molecules around the -10 range of ΔG values.

There does not appear to be a lot of information on the spontaneity of PROTAC reactions but it is worth noting that the ones with recorded ΔGs appear energetically favorable, as expected.

Let's now take a look at the $K_d$ values.

Similar to ΔG values, there does not appear to be a lot of information on the affinity of formed PROTAC complexes. Since the range is so large, let's plot a second histogram focused on the PROTACs with low $K_d$.

The improved resolution of values illustrates a much cleaner distribution of $K_d$ values. We do see that a number of them have low, favorable $K_d$ values indicating that the PROTAC linker can form a strong connection with the E3 ligase and target protein.

Let's now take a look at XLogP3 values. Note that this is slightly different than the typical LogP partition coefficient. Recall that LogP is defined $$LogP = log(\frac{[solute]_{oct}}{[solute]_{H_2O}})$$

In other words, LogP is the measured ratio of the concentration of a compound in the organic phase to the its concentration in the aqueous phase, measuring the compound's solubility. XLogP3 is a knowledge-based method for calculating the partition coefficient by accounting for the molecular structure, presence of functional groups, and bonding [8] . Both properties estimate a compound's liphophilicity, giving insight into how a compound may behave in biological systems.

All PROTAC compounds have a recorded XLogP3 value. The distribution looks normally distributed with few molecules with extreme logP profiles.

Now, let's take a look at the PROTAC degradation properties. "DC50 (nM)" and "Dmax (%)" represent the half maximal degradation concentration and maximal degradation of the target protein of interest, respectively. Let's take a quick look at their distributions.

Let's first do some data cleaning

Notice that the values are all in string format with non-numerical characters such as '<', '/', and '>'. For the time being, let's remove these values.

The distribution is certainly skewed and has a few outliers. Let's log normalize.

Now, let's take a look at Dmax percentage which represents the maximal degradation a PROTAC can elicit relative to the total activity of the target protein of interest [7] .

Notice that Dmax is represented as a percentage. For now, let's continue with regressing on DC50. We are now ready to featurize!

3. Featurization ¶

Let's featurize using CircularFingerprint which is incorporated in DeepChem! CircularFingerprint is a common featurizer for molecules that encode local information about each atom and their neighborhood. For more information, the reader is refered to [9] .

Along with a random split, let's also use a scaffold split which ensures that the split contain a structurally diverse array of compounds. Scaffold split groups molecules according to presence of rings, linkers, combinations of rings and linkers, as well as atomic properties. In general, scaffold splits are a good way of ensuring generalizability of our models.

To see the scaffold split in action, let's visualize the chosen compounds across the splits.

There are certainly functional group differences spread throughout the splits. Notice the presence of the nitrile group in the train set, amine group in the validation set, as well as the sulfonamide group in the test set.

Additionally, notice the structural and conformational differences among the various data splits. It will be interesting to see how well our model generalizes.

4. Model deployment ¶

We have successfully generated our train and test datasets. Let's now create a simple MLP model to predict PROTAC degradation properties!

Let's now wrap everything together to instantiate a DeepChem model! Note that due to the small sample size, a smaller batch size actually helps performance.

Let's define the validation function to prevent overfitting.

We can easily look at how the training went through plotting the recorded losses.

We can see that the model performs less well on the scaffold validation set which makes sense as the scaffold splits ensures that more validation molecules are out of distribution relative to the train distribution.

Let's now perform some inference on our test set to evaluate our models!

The random split appears to do fairly well. Let's see how well our model does on the scaffold split.

The model does significantly worse on the held out scaffold test set which was expected given the simplicity of the model. Developing far more complex models which can generalize out of distribution is a key area of focus in many areas of research from molecule property prediction to computer vision to natural language processing. In general, I hope this tutorial was a informative introduction into the world of PROTACs. Follow along as we explore how we can think about PROTAC design in the next tutorial!

5. References ¶

[1] Kelm, J.M., Pandey, D.S., Malin, E. et al. PROTAC’ing oncoproteins: targeted protein degradation for cancer therapy. Mol Cancer. 2023, 22, 62. https://doi.org/10.1186/s12943-022-01707-5

[2] Tu, Y., Chen, C., Pan, J., Xu, J., Zhou, Z. G., & Wang, C. Y. The Ubiquitin Proteasome Pathway (UPP) in the regulation of cell cycle control and DNA damage repair and its implication in tumorigenesis. International journal of clinical and experimental pathology. 2012, 5, 8.

[3] Sun, X., Gao, H., Yang, Y. et al. PROTACs: great opportunities for academia and industry. Sig Transduct Target Ther. 2019, 4, 64. https://doi.org/10.1038/s41392-019-0101-6

[4] Che Y, Gilbert AM, Shanmugasundaram V, Noe MC. Inducing protein-protein interactions with molecular glues. Bioorg Med Chem Lett. 2018, 28, 15. https://doi.org/10.1016/j.bmcl.2018.04.046 .

[5] Békés, M., Langley, D.R. & Crews, C.M. PROTAC targeted protein degraders: the past is prologue. Nat Rev Drug Discov. 2022, 21, 181–200. https://doi.org/10.1038/s41573-021-00371-6

[6] Liu, Z., Hu, M., Yang, Y. et al. An overview of PROTACs: a promising drug discovery paradigm. Mol Biomed. 2022, 3 (46). https://doi.org/10.1186/s43556-022-00112-0

[7] Gaoqi Weng, Xuanyan Cai, Dongsheng Cao, Hongyan Du, Chao Shen, Yafeng Deng, Qiaojun He, Bo Yang, Dan Li, Tingjun Hou, PROTAC-DB 2.0: an updated database of PROTACs, Nucleic Acids Research. 2023, 51 (D1), Pages D1367–D1372, https://doi.org/10.1093/nar/gkac946

[8] Cheng T, Zhao Y, Li X, Lin F, Xu Y, Zhang X, Li Y, Wang R, Lai L. Computation of octanol-water partition coefficients by guiding an additive model with knowledge. J Chem Inf Model. 2007, 47 (6), 2140-8. https://doi.org/10.1021/ci700257y .

[9] Glem RC, Bender A, Arnby CH, Carlsson L, Boyer S, Smith J. Circular fingerprints: flexible molecular descriptors with applications from physical chemistry to ADME. IDrugs. 2006, 9 (3).

Congratulations! Time to join the Community! ¶

Congratulations on completing this tutorial notebook! If you enjoyed working through the tutorial, and want to continue working with DeepChem, we encourage you to finish the rest of the tutorials in this series. You can also help the DeepChem community in the following ways:

Star DeepChem on GitHub ¶

This helps build awareness of the DeepChem project and the tools for open source drug discovery that we're trying to build.

Join the DeepChem Discord ¶

The DeepChem Discord hosts a number of scientists, developers, and enthusiasts interested in deep learning for the life sciences. Join the conversation!