PROTACs are emerging as a promising new class of drugs to target ‘undruggable’ proteins. But what are PROTACs? How do they work? How are PROTACs optimised? In this blogpost, we describe what PROTACs are, how they work, and how DeepMirror Chem can accelerate the development of PROTACs during hit-to-lead by predicting their affinity.

Some disease-causing proteins are notoriously hard to target. Drug discovery aims to find small molecules that can fit into ‘binding pockets’ within the protein. Binding of a small molecule to a pocket often causes a large change in the protein’s shape, leading to its inhibition (and therefore stops its disease-causing effects). However, many proteins have broad or shallow pockets that are hard to cover by small molecules or may have smooth surfaces with no pockets at all. More importantly, binding to some pockets might not change the function of the protein. Proteins with broad, shallow or no pockets are considered hard-to-drug or ‘undruggable’ and account for a high percentage of all disease-related proteins: only 20-25% of known protein targets are being ‘drugged’ or explored for drug discovery (Sun et al., 2019).

How can we drug more ‘undruggable’ proteins? How can we develop drugs against ‘smoother’ proteins with shallow or broad pockets? This is one of the promises of PROTACs, PROteolysis-TArgeting Chimeras. PROTACs do not need to bind to an active binding pocket, instead binding to other sites in the protein to have a therapeutic effect and can therefore broaden the spectrum of proteins we can drug.

Targeted Protein Degradation: a novel way of targeting proteins

PROTACs promise to be effective on undruggable targets because they rely on a novel way of targeting proteins through “Targeted Protein Degradation” (TPD), a process by which proteins are removed to stop their activity (as opposed to being inhibited).

This is usually done by ‘tagging’ a protein with a small protein ‘flag’ called ubiquitin, to direct it to a cell’s own protein degradation machinery, the proteasome.

PROTACs are heterobifunctional small molecules, meaning that they have two different functions. One ‘end’ of these small molecules binds to a protein target (Fig 1a, purple), and the other ‘end’ binds to a ubiquitin ligase (Fig 1a, pink). Ubiquitin ligases are enzymes that tag proteins with ubiquitin, a ‘flag’ for degradation (Fig 1a, yellow). These two ‘ends’ are joined by a linker (Fig 1a, orange). In essence, PROTACs are small molecules that bring together a disease-related protein (Fig 1b, dark purple) and a protein degrader (Fig 1b, dark pink). The tagging enzyme ‘flags’ the disease-related protein with several ubiquitin molecules (Fig 1b, yellow), signalling that the disease-related protein must be degraded by a cell’s proteasome (Fig 1b, green).

Why are PROTACs so great?

The fundamental advantage of Targeted Protein Degradation (TPD) - the mode-of-action of PROTACs - over inhibition is that TPD does not require an active binding pocket to exert its therapeutic effects (Békés et al., 2022). The target-binding ‘end’ of a PROTAC can bind somewhere else in the protein and does not need to induce a change of shape to inhibit the protein. Instead, the cell will degrade the protein and thus stop it from exerting its activity. In addition, PROTACs work through catalysis (as opposed to inhibition as is the case for most small molecules): once a PROTAC has facilitated the ubiquitination of the target protein, the PROTAC is released and becomes free to catalyse the same reaction again. This means that a PROTACs with even modest binding affinity to a target protein can achieve great effects (as long as it is highly specific to the target).

PROTACs in clinical trials

No PROTACs are currently approved by the FDA, so, until recently, questions remained on their clinical safety and efficacy. However, in 2019, two different PROTACs entered first-in-human trials: ARV-110, targeting the Androgen Receptor and ARV-471 targeting the Estrogen Receptor (Fig 1c). Both phase II clinical trials showed these PROTACs were tolerable (safe), could induce degradation of their target proteins, and showed clinical benefit for their indications (prostate and breast cancer respectively).

ARV-110 and ARV-471 first-in-class PROTACs are only the first wave. Their targets, the Androgen Receptor (AR) and the Estrogen Receptor (ER), have strongly validated roles in disease and are ‘classically’ drugged targets against which multiple drugs have already been developed. These clinical trials are fundamental proof-of-concept studies about the immense potential of PROTACs, that solidify TPD as a new therapeutic modality. Having explored how AI can accelerate drug discovery in our previous blogpost, we wondered whether DeepMirror Chem could learn PROTAC properties too. Can we use AI to explore a new and exciting small molecule class such as PROTACs where there is scarce data?

DeepMirror Chem for PROTACs

Given the clinical promise of PROTACs, we used a publicly available PROTAC database (Weng et al., 2023) to assess whether we could identify effective protein degraders using DeepMirror Chem.

We selected the Androgen Receptor (AR) as the target protein as most PROTACs in the dataset were tested against it. In Figure 1b, AR would be represented by the target protein in dark purple. We used DeepMirror Chem to predict DC50, the half-maximal degradation concentration. DC50 measures the concentration required for half of the target protein (AR in this case) to be degraded. The lower the DC50, the lower the concentration of PROTAC required to degrade a protein. We used the DC50 of 49 PROTACs designed against AR to learn from using DeepMirror Chem, and then predicted the DC50 against AR of 5 PROTACs not used for learning (Figure 2).

PROTACs not only bind a target protein (Fig 1b, dark green), they also bind a protein degrader enzyme (Fig 1b., dark blue). Therefore, efficacy of PROTACs is determined by their affinity to the target, but also their affinity to E3 ubiquitin ligase. Hence, we also predicted PROTAC binding to the E3 ubiquitin ligase (IC50). We used the affinity of 129 PROTACs against E3 ligases to learn, and predicted the E3 ligase IC50 of 13 PROTACs not used for training (Figure 2).

To compare the values predicted by DeepMirror Chem against real values, we calculated Pearson’s R as a measure of correlation. A Pearson’s R value of 0 would indicate no correlation, while a value of 1 would indicate perfect correlation. A high correlation (close to 1) between real and predicted values would mean DeepMirror Chem is good at predicting properties of PROTACs, and thus could be used to de-risk new candidates for testing and development.

In our qualitative small-scale study, DeepMirror Chem predicted AR degradation and E3 ligase binding of PROTACs well (Pearson’s R of 0.86 for AR, 0.69 for E3) from as little as 49 datapoints (Fig 2).

Outlook

PROTACs are emerging as an exciting new family of small molecules with the potential to target undruggable proteins. First-in-class PROTAC clinical trials are fundamental proof-of-concept studies about the immense potential of PROTACs that solidify TPD as a new therapeutic modality. However, these first PROTACs have focused on well-established and already drugged disease targets and are therefore yet to expanded the repertoire of druggable targets. The next step is where PROTACs’ true promise lies: can PROTACs really target ‘undruggable’ proteins? A coming wave of new PROTACs against hard-to-drug targets will hopefully answer the question.

During PROTAC development, all three structural parts of a PROTAC must be adapted and optimised: the target binding structure (Fig. 1, purple), the E3 binding structure (Fig. 1, pink) and the linker structure that joins them (Fig. 1, orange). Out of these three, identifying (non-active) binders against new targets and designing effective linkers will pose the greatest challenges.

DeepMirror Chem can assist in the process of optimisation by predicting important properties of PROTACs, such as their degradation capabilities, thereby fast-tracking their hit-to-lead development.