Skip to main content
NCBI home page
Bioinformatics Advances logoLink to Bioinformatics Advances
. 2025 Mar 14;5(1):vbaf056. doi: 10.1093/bioadv/vbaf056

Challenges in predicting PROTAC-mediated protein–protein interfaces with AlphaFold reveal a general limitation on small interfaces

Gilberto P Pereira 1,2,2, Corentin Gouzien 3,2, Paulo C T Souza 4,5, Juliette Martin 6,✉
Editor: Michael Gromiha
PMCID: PMC11938821  PMID: 40144455

Abstract

Motivation

Proteolysis Targeting Chimeras (PROTACs) are heterobifunctional molecules composed by ligands binding to a target protein and a E3-ligase complex, connected by a linker, that induce proximity-based target protein degradation. PROTACs are promising alternatives to conventional drugs against cancer. Predicting PROTAC-mediated complexes is often the first step for in silico PROTAC design pipelines. We previously noted that AlphaFold2 (AF2) fails to predict PROTAC-mediated complexes.

Results

Here, we investigate the potential causes of this limitation. We consider a set of 326 protein heterodimers orthogonal to the AF2 training set, and evaluate AF2 models focusing on the interface size and presence of interface ligand. Our results show that AF2-multimer predictions are sensitive to the size of the interface to predict even in the absence of ligands, with the majority of models being incorrect for the smallest interfaces. We also benchmark both AF2 and AF3 on a set of 28 PROTAC-mediated dimers and show that AF3 does not significantly improve upon the accuracy of AF2. The low accuracy of AF2 on complexes with small interfaces has strong implications for computational pipelines for PROTAC design, as these stabilize typically small interfaces, and more generally on any prediction task that involves small interfaces.

Availability and implementation

All the models analyzed in this article are available in the Zenodo archive https://zenodo.org/records/14810843.

1 Introduction

In the quest to develop innovative therapies, pharmaceutical companies and research laboratories invest billions of dollars every year into the different stages of drug design pipelines (Anon 2023). Recently, many of them have invested into the development of Proteolysis TArgeting Chimeras (PROTACS). PROTACs are heterobifunctional ligands composed of two small molecules connected by a linker region. One of the small molecules binds to a protein target and the other binds to a protein called E3 ligase, which is attached to the ubiquitination machinery (Cyrus et al. 2011, Burslem and Crews 2020). This machinery is responsible for ubiquitination of proteins, which tags them for proteasomal degradation. The formation of a complex between the E3 ligase and the protein target via the PROTAC thus induces the degradation of the target. PROTAC-based approaches have been explored for a variety of therapeutic targets, including proteins intimately connected to cancer (Crew et al. 2018, Jaime-Figueroa et al. 2020, Pettersson and Crews 2019). These molecules have several advantages over small molecule-based drugs due to their catalytic mode of action. Due to it, the concentration at which a therapeutic effect is achieved tends to be lower than standard small molecule inhibitors, leading to a lower risk of side-effects (Cromm 2022). However, up until a few years ago, these molecules were designed serendipitously due to both a lack of structural information and limited successful applications of in silico tools.

Thanks to the effort undertaken the last few years, there now exists a wealth of structural, biochemical and activity data regarding several different PROTAC families which can be exploited toward the construction of computational pipelines for PROTAC design. The first step in these pipelines typically pertains to the prediction of the E3-ligase/target complexes, which can be achieved by a myriad of methods. While some groups have developed tools for this purpose (Weng et al. 2021, Li et al. 2022, Zheng et al. 2022, Almodóvar-Rivera et al. 2023, Duran-Frigola et al. 2023), they are typically not general because they require that the PROTAC molecule be known a priori. Within our group we developed a pipeline called PROTACability (Pereira et al. 2023), which bypasses this constraint and achieves satisfactory accuracy with minimal prior information. PROTACability was based on predictions of the E3 ligase/target interfaces produced by LightDock, a classical docking software (Roel-Touris et al. 2020).

The field of protein structure prediction has been deeply impacted by the AlphaFold method (Senior et al. 2020), as well as protein–protein docking. Indeed, although not initially developed for protein–protein complex prediction, AlphaFold has been rapidly exploited by the community toward this goal, using hacks such as concatenating the sequence of several proteins using multi-glycine spacers (Ko and Lee 2021) or modifying the input residue index (Bryant et al. 2022, Gao et al. 2022, Mirdita et al. 2022). AlphaFold-multimer was then specifically developed for protein assemblies (Evans et al. 2022). These developments offer a new alternative for protein–protein docking. Recently, the new iteration of AlphaFold culminated into AlphaFold 3, the newest version which is able to take into consideration the information of ligands and model multimeric proteins accurately (Evans et al. 2022, Roy and Al-Hashimi 2024). AlphaFold3 also differs from AlphaFold2 (AF2) in several other aspects, including a smaller and simpler embedding step for the multiple sequence alignment (MSA), a new diffusion model in place of the AF2 structure module, early stopping during training and random noise sampling during inference time (Abramson et al. 2024). This results in an improved prediction of protein–protein complexes, particularly for antibody-antigen complexes, suggesting a lower dependency on co-evolutionary signal between entities.

During the development of PROTACability, we also considered using AlphaFold-multimer, with little success (Pereira et al. 2023). A recent benchmark of docking tools for PROTAC-mediated complexes confirm these poor quality results (Pereira et al. 2023, Rovers and Schapira 2024). In this article, we explore further the reasons for this failure. We first use a test set of protein–protein heterodimers never seen by AlphaFold during the training phase to investigate the effect of the interface size and presence of ligand at the interfaces on the quality of the generated models. We then test the capacity of AlphaFold3 on the PROTAC-mediated interfaces on the relevant dimers, as well as the dimers in presence of the accessory proteins to better mimic the biological context of the interactions and reduce the search space. The benchmark on protein–protein heterodimers should help us to identify whether the difficulty with PROTAC-mediated complexes lies in (i) the interface size, (ii) the presence of a ligand, or (iii) the stabilization of a non-natural complex resulting in the absence of any co-evolutionary signal. Our results point at an inherent limitation of AlphaFold-multimer on small interfaces in the general case, which compromises the prediction of PROTAC-mediated interfaces which are generally small. AlphaFold3 does not remedy this situation, even when the prediction context is provided. Nonetheless, as AF3 predictions were carried out without the presence of PROTACs in the available web server, it could be the case that AF3 full capabilities have not been explored.

2 Methods

2.1 Dataset

A dataset of 28 PROTAC-mediated complexes was extracted from the PDB. These 28 cases represent the cases available at the time of the study and were chosen based on the following criteria: (i) they should contain at least the ternary complex (E3-ligase, protein target and PROTAC), (ii) resolution better than <4 Å, (iii) no missing residues at the protein–protein interface. In 26 cases, accessory proteins bound to the E3 ligase receptor are present. These accessory proteins are ignored in the prediction, unless otherwise stated.

A test set was extracted from the PDB as follows. All protein heterodimers (two different peptidic chains in the assembly) deposited after 1 October 2021 (training date of AF2 v3) were retrieved from the RCSB PDB website. The redundancy against the AF2 training set was removed using Foldseek (Kim et al. 2025) (software version fc40ba1afdcb20602116907fdc73a47fd4c78615) against PDB sequences with a 30% sequence identity cutoff and an 80% sequence coverage cutoff (easy-multimersearch—min-seq-id 0.3—cov-mode 2 -c 0.8). Dimers with detectable sequence similarity to complexes acquired before 30 September 2021 were removed. The intraset redundancy was then reduced using cd-hit (Li and Godzik 2006, Fu et al. 2012) (software version 4.8.1), with a 40% cutoff (step 1: redundancy reduction on a per-chain basis; step 2: addition of all chains of complexes selected in step 1, step 3: removal of remaining redundancy). Sequences shorter than 30 residues were removed. The resulting test set is composed of 340 heterodimers. Among these 340 dimers, 5 cases only were mediated by a ligand (see below).

Structures anterior to the AlphaFold training date were added to increase the number of ligand-mediated interfaces. We used the Dockground resource (https://dockground.compbio.ku.edu/bound/index.php) (Collins et al. 2022) to build a set of non-redundant (<30% sequence identity) heterodimers X-ray structures with a resolution better than 3 Å deposited earlier than 30 September 2021. A list of 20 ligand-mediated complexes was extracted from this set. The list of the complexes used throughout this study is given in Table S1.

2.2 Interface size

The interface size was defined as the accessible surface area (ASA) of separated chains minus the ASA of the complex:

ΔASA=ASAchain1+ASA(chain2)-ASA(complex).

ASA is computed using naccess (Hubbard and Thornton n.d.) with default parameters (probe size = 1.4 Å, Z slice = 0.05 Å). Unless otherwise specified, heteroatoms (i.e. non-standard residues, solvent and ligands) are excluded from the computation.

2.3 Ligand-mediated interfaces

Ligand-mediated interfaces were identified by comparing the ASA computed with and without heteroatoms, thus excluding peptidic ligands. Cases where the difference between ASA with and without heteroatoms was greater than 10% were manually verified to exclude complexes with cross-links, modified residues at the interface, or non-specific ligands at the interface (such as sulfate ions, glycerol, or PEG groups).

2.4 AlphaFold 2

AlphaFold-Multimer calculations were carried out using ColabFold version 1.5.1(Evans et al. 2022, Mirdita et al. 2022, Brahma and Raghuraman 2024) using the alphafold2_multimer_v3 model with default parameters (ie paired+unpaired MSA, with model recycling). The best model was selected using the composite AF2 predicted quality score defined by 0.8 ipTM + 0.2pTM (Evans et al. 2022). We will refer to this score as the AF2 confidence score. Among the 340 cases of the test set, AF2 was able to produce a model for 326 cases, including the 5 cases with ligand-mediated interface.

2.5 AlphaFold 3

AlphaFold 3 predictions were run on the web server (https://alphafoldserver.com/) on the 6th of June 2024, solely for PROTAC cases. At that time, the webserver could only model common biological ligands, i.e. not PROTACs. Two experiments were carried out: in the first one, AF3 was provided with both sequences of the ligase and target (like AF2); in the second experiment we added the sequence of the accessory protein which is bound to the ligase in physiological context and present in the experimental structures. We refer to this experiment as “prediction with context.” Model choice was based on the ranking score provided by AF3.

2.6 Model evaluation

To evaluate the similarity between the experimental structures and the predicted structures from AF-Multimer and AF3, the DockQ criteria were employed(Basu and Wallner 2016). In short, DockQ provides a continuous score from 0 to 1 (with 1 being perfect similarity) which takes into account the fraction of conserved native contacts, RMSD of the target protein and the interface RMSD between a reference structure and a predicted structure (Pereira et al. 2023). We used the cutoff of 0.23 for acceptable models, in agreement with CAPRI criteria (Mirabello and Wallner 2024). The choice of a relaxed cutoff is motivated by the fact that PROTAC systems are highly plastic and mobile (Dixon et al. 2022).

When predicting PROTAC-mediated interfaces with context (i.e. in presence of accessory proteins), only the DockQ score of the PROTAC-mediated interface was reported.

2.7 Data availability

All the models analyzed in this article and the data used in Fig. 2 are available in the Zenodo archive https://zenodo.org/records/14810843.

Figure 2.

Figure 2.

Open in a new tab

Interface size in the models as a function of the native interface size, for wrong models (DockQ < 0.23). Black points: dimers from the test set with no ligand at the interface (108 cases). Filled Red triangle: dimers from the test set with ligand at the interface (1 case). Open red triangles: dimers from the training set with ligand at the interface (4 cases). Open green squares: PROTAC-mediated dimers (23 cases).

3 Results

3.1 Predicting protein–protein interfaces using AlphaFold-Multimer

We collected 340 protein dimers not included in and not similar to the AF2 training set to evaluate the performance of AF2 on an unbiased set. Since PROTAC-mediated interfaces are both small and mediated by a ligand, we paid special attention to the interface size and the presence of ligand at the interface. Since very few interfaces are mediated by ligands in the test set, we also considered 20 ligand-mediated interfaces included in the AF2 training set. The predictions were evaluated by comparison with the experimental structures via the DockQ score, a continuous score in the range [0,1], that integrates the RMSD of the backbone of the shortest chain after superimposition of the longer chains, the RMSD of backbone interface atoms and the fraction of native interfacial contacts correctly predicted. A cutoff of 0.23 was used to discriminate correct from incorrect models, in agreement with CAPRI criteria (Basu and Wallner 2016, Mirabello and Wallner 2024). In Fig. 1, we show the model quality, assessed by the DockQ score, as a function of the size of the native interface, measured by ΔASA. Globally, AF2 produces correct models for 68% of the test cases that do not involve interface ligand (219 out of 316 back circles in Fig. 1). In the presence of ligands (both open and filled red triangles in Fig. 1), the proportion of correct models is equal to 85% (17 out of 20 open rend triangles and 4 out of 5 filled triangles in Fig. 1). For the PROTAC-mediated complexes, the proportion of correct models drops to 18% (5 out of 28 green squares in Fig. 1).

Figure 1.

Figure 1.

Open in a new tab

DockQ scores as a function of interface size. Here, the interface size is computed on the native complex. Black points: dimers from the test set with no ligand at the interface (316 cases). Filled Red triangles: dimers from the test set with ligand at the interface (5 cases). Open red triangles: dimers from the training set with ligand at the interface (20 cases). Open green squares: PROTAC-mediated dimers (28 cases). The horizontal dashed line indicates the threshold for acceptable models (DockQ = 0.23).

As can be seen in Fig. 1, the quality of the models predicted by AF2 is severely impacted by the interface size. Interfaces smaller than 1000 Å2 are very challenging to predict: the majority of the models have a DockQ score lower than 0.23 (average DockQ score < 0.16, see Table S2). There is a clear shift between 1000 and 2000 Å2, where about 60% of the models are of acceptable quality (average DockQ score 0.46, see Table S2), whereas for interfaces larger than 2000 A2, more than 75% of the models are of acceptable quality (average DockQ score >0.55, see Table S2). We observed a high correlation between the DockQ scores and the AF2 confidence scores, see Fig. S1A, with a Pearson correlation coefficient equal to 0.79. However, some models of low quality have good AF2 confidence scores, see Fig. S1B, particularly for small interfaces.

If we now focus on interfaces mediated by ligands (both open and filled red triangles in Fig. 1), they seem to follow the same trend as other interfaces, with a strong influence of the interface size. PROTAC-mediated interfaces (green squares in Fig. 1) are all small in size, and the predicted models are of poor quality. Except for five cases, the DockQ score is lower than 0.23. Increasing the sampling by AF2 with three seeds instead of one did not improve the results (Fig. S2). We also monitored the influence of the ligand size, however, no clear trend could be identified due to the limited sample size (Fig. S3).

To further investigate the bias related to the interface size, we monitored the size of the predicted interfaces with respect to the native ones, in the case of wrong models (DockQ < 0.23). The results are shown in Fig. 2. We observe a general tendency of AF2 to predict models with larger interfaces than the native complex, when the interfaces to predict are small. Of note, interfaces smaller than 1000 A2, which are frequently mispredicted (see Fig. 1), systematically produce models with much larger interfaces.

In conclusion, the prediction of small interfaces with AF2 is still a challenge; and PROTAC-based interfaces typically belong to this category, preventing us from going further into the reasons of failure. On natural dimers, the presence of a ligand in itself does not seem to perturb the prediction.

3.2 Attempted improvements of PROTAC-mediated interface prediction with AlphaFold 3

We tested the accuracy of AF2 and the newest AF3 on the set of challenging PROTAC-mediated dimers. The predictions were evaluated by comparison with the experimental structures via the DockQ score as explained before. The results are shown in Fig. 3.

Figure 3.

Figure 3.

Open in a new tab

Comparison of calculated DockQ values for all PROTAC-mediated systems using either AF2 or AF3 with respect to experimental crystal structures. The black line highlights DockQ’s acceptable criteria threshold at 0.23. For systems 6w7o and 6w8i, no accessory protein was present in the experimental structure.

As can be seen in Fig. 3, AF3 generates acceptable models for 5 cases out of 28 (grey point in Fig. 2). This is similar to AF2 global accuracy, which also generates 5 acceptable models (red points), although AF2 and AF3 do not succeed on the same cases (3 cases are correctly predicted by both). Of note, correct models generated by AF3 have relatively high DockQ scores (>0.5) compared to AF2, reflecting more accurate predictions.

We tested further the capability of AF3, by providing context to the prediction. Indeed, in physiological context the E3 ligase receptor is usually bound to accessory proteins, making a part of the surface inaccessible to an interaction with the target protein. We hypothesized that the addition of context, in the form of the accessory proteins found in experimentally determined structures, would improve the prediction of the receptor/substrate interface by limiting the search space available.

The E3 ligase receptor/accessory protein interfaces are highly conserved and should be easy to predict. However, the comparison between DockQ scores with and without context (orange versus grey bars in Fig. 3) shows that the addition of context does not improve the prediction: AF3 with context (AlphaFold3: Full, orange bars) generates acceptable models for 3 cases only versus 5 without context. Of note, these 3 cases were already well predicted by AF3 without context (Alphafold: Dimers, grey bars). Within these 3 cases with correct prediction with context, 2 cases have an improved DockQ score, although the increase is small. For example, the model for 8BDS has a DockQ score equal to 0.652 with context and 0.588 without context. For the large majority of systems, we find that the inclusion of the context does not significantly improve the prediction accuracy, and in some cases AF3 predictions including accessory proteins appear to be worse, as shown for 5HXB and 7Q2J in Fig. 4. The predictions for all the 28 PROTAC cases are shown in Fig. S5.

Figure 4.

Figure 4.

Open in a new tab

Top ranked predicted protein–protein interfaces for 5HXB (top) and 7Q2J (bottom). (A, D) Crystal structures; (B, E) AF3 models of dimers (DockQ: 0.833 and 0.032); (C, F) AF3 models with inclusion of the accessory protein (DockQ: 0.036 and 0.043).

The inspection of models generated with context reveals that the interfaces between E3 ligase receptors and accessory proteins are generally well predicted, with all DockQ scores greater than 0.6 and a median DockQ score equal to 0.941, see Fig. S4. However, this did not allow to improve the results for the E3 Ligase-Target interface. Given the current limitations of the AF3 web server in terms of available ligands, it was not possible to test whether inclusion of PROTACs would increase prediction accuracy. In conclusion, PROTAC-mediated interfaces remain very challenging to predict, even with AF3 and some context.

4 Discussion

The prediction of E3-ligase/protein target complex can be the first step in computational pipelines for PROTAC design. In this context, the available information are the structures of E3 ligase in complex with a ligand, and of the protein target with a ligand. An important part of the PROTAC design is the optimization of the linker between the two ligands, which is determinant for the molecule's catalytic activity (Cyrus et al. 2011, Békés et al. 2022, Bashore et al. 2023). In our previous work (Pereira et al. 2023), we developed a protocol to predict those complexes using LightDock (Roel-Touris et al. 2020), a macromolecular docking framework that allows the incorporation of residue-based restraints. At that time, we also benchmarked AF2, with limited success, on PROTAC-mediated ternary complexes. A recent benchmark on a larger set of 43 PROTAC cases conducted by others and published during the preparation of this article confirmed the limitation of AF2 on PROTAC-mediated complexes (Rovers and Schapira 2024).

In this study, we further investigate the reasons for this limitation. PROTAC-mediated interfaces have several peculiarities that could impact the prediction: (i) small interface size/area, (ii) presence of ligand at the interface, (iii) included in larger complexes with other proteins. Here, we tried to deconvolve these different factors. First, we show that AF2 suffers from a significant bias against small interfaces. Within the PPIs we studied, those with larger areas were typically better predicted than small ones (Fig. 1). This bias was already present in the former version of AF2 trained on monomers, as observed by Yin et al (Yin et al. 2022) who benchmarked AF v.2.0 with the residue index hack on the 152 complexes of the Protein–Protein Docking Benchmark 5.5 (Vreven et al. 2015). Our results indicate that this bias persists in the version of AF2-multimer that was specifically trained on protein–protein complexes. More sophisticated approaches that build upon AF like AFSample (Wallner 2023) (that uses dropout during inference) and AFProfile(Bryant and Noé 2024a) (that denoises the input MSA) can improve the prediction in some cases. The potentiality of these approaches in the case of small interfaces remains to be explored. Another alternative would be to fine-tune AF for small interfaces.

Second, concerning the presence of a ligand at the interface, we have considered protein–protein interfaces that involve ligands for more than 10% of their size, to see if we could predict them in the absence of ligands—since AF2 does not model ligands. Our results indicate that these interfaces do not pose a specific challenge for the prediction, as many of them were predicted with DockQ scores between moderate (0.49) and high quality (0.80). However, both ligand-mediated and non-ligand-mediated interfaces appear to be ubiquitously affected by the interface size problem. Because all the known PROTAC-mediated interfaces are small, we cannot disentangle the effect of interface size from a hypothetical PROTAC-specific effect, as it would require a comparison with large PROTAC-mediated interfaces.

Third, despite the limitation of AF2 on PROTAC-mediated complexes, we also tested the effect of providing some context to the prediction: by adding the accessory proteins, the interface accessible to the interface with the substrate would be more limited, narrowing down the solution space. This was done with AF3. However, we noted no improvement in the predictions, with or without context.

A known peculiarity of PROTAC-mediated complexes is that they are induced and stabilized by the ligand, and the two proteins would probably not interact strongly in the absence of the PROTAC molecule. The resulting absence of an evolutionary signal could be another source of challenges for the accurate prediction of these complexes. A recent work of Roney and Ovchinnikov (2022) provides evidence that AF2 indeed learned an implicit energy function that encapsulates the physics governing the folded state and that the evolutionary signal serves as a guide in the global search for the optimal structure. Another recent study leads to the same conclusion, observing that AF has a remarkable capability to recover correct structures from certain perturbations without additional information provided by the MSA (Gut and Lemmin 2024). Along the same line, AF2Complex, proposed by Gao et al (2022), successfully predicted heterodimers starting from unpaired MSAs, suggesting again that the AF2 network learned an implicit energy function that is enough to represent the physics of protein–protein interfaces. So, in theory, it should be possible for AF2, and by extension AF3, to predict a non-natural interface in the absence of an evolutionary signal.

The absence of the PROTAC molecule likely poses a significant challenge, as the PROTAC molecule itself has a key role in stabilizing the protein–protein interface of these complexes. Given that these are large and flexible ligands, any algorithm that explicitly models the ligand must be capable of sampling (or receiving as input) multiple ligand configurations and then generate an ensemble of ternary complex configurations. Ensemble-generation methods could then be used to maximize the likelihood of finding physiologically relevant structures (Shehu et al. 2006, Noé et al. 2019, Dixon et al. 2022, Bryant and Noé 2024b; Lewis et al. 2024).

The release of RoseTTAFold-All-Atom (Dixon et al. 2022, Krishna et al. 2024) in 2024, might offer an alternative to the AF family of methods, as it is able to consider the effects of arbitrary ligands within protein–protein interfaces. It remains to be seen how well this modeling suite achieves in predicting highly plastic, “artificial” and shallow protein–protein interfaces such as those in PROTAC-mediated systems. Methods like AlphaLink, could get around the problem of ligand modeling by the incorporation of restraints in the prediction network (Stahl et al. 2023).

To summarize, because of their small size and peculiar physics, PROTAC-mediated complex prediction remains challenging even with cutting-edge deep-learning methods. Potential solutions to overcome these limitations include the development of models trained on small interfaces, and models that better consider the effects of ligands or incorporate restraints. Additionally, training models on conformational ensembles extracted from molecular dynamic simulations of PROTAC-mediated complexes is likely to yield better predictions for PROTAC-mediated PPIs. Since the field of deep-learning based structure prediction is rapidly evolving, new ligand-aware models have been developed during the course of the present study. Most recent methods like Chai-1 (Chai Discovery et al. 2024) and Boltz-1 remain to be tested; they could change the way PROTAC design is traditionally addressed since they allow the direct inclusion of ligands in the prediction (Stahl et al. 2023, Wohlwend et al. 2024). To put our result in a broader perspective, the limitation of AF for the prediction of complexes with small interfaces calls for caution when working with multimeric systems that likely involve small interfaces.

Supplementary Material

vbaf056_Supplementary_Data
vbaf056_supplementary_data.docx (3.8MB, docx)

Acknowledgements

We acknowledge the support of the Centre Blaise Pascal’s IT test platform at ENS de Lyon (Lyon, France) for the computer facilities. The platform operates the SIDUS solution developed by Emmanuel Quemener (Quemener and Corvellec 2013).

Contributor Information

Gilberto P Pereira, Laboratoire de Biologie et Modelisation de la Cellule, Ecole Normale Superieure de Lyon, CNRS, UMR 5239, Universite Claude Bernard Lyon 1, Inserm, U1293, Lyon F-69364, France; Centre Blaise Pascal de Simulation et de Modelisation Numerique, Ecole Normale Superieure de Lyon, Lyon 69364, France.

Corentin Gouzien, Laboratoire d'Océanographie Microbienne, UMR 7621, CNRS-SU, Observatoire Océanologique de Banyuls, Banyuls-sur-Mer F-66650, France.

Paulo C T Souza, Laboratoire de Biologie et Modelisation de la Cellule, Ecole Normale Superieure de Lyon, CNRS, UMR 5239, Universite Claude Bernard Lyon 1, Inserm, U1293, Lyon F-69364, France; Centre Blaise Pascal de Simulation et de Modelisation Numerique, Ecole Normale Superieure de Lyon, Lyon 69364, France.

Juliette Martin, Laboratoire de Biologie et Modelisation de la Cellule, Ecole Normale Superieure de Lyon, CNRS, UMR 5239, Universite Claude Bernard Lyon 1, Inserm, U1293, Lyon F-69364, France.

Author contributions

Gilberto P. Pereira (Conceptualization [equal], Data curation [equal], Formal analysis [equal], Investigation [equal], Methodology [equal], Supervision [equal], Validation [equal], Visualization [equal], Writing—original draft [equal], Writing—review & editing [equal]), Corentin Gouzien (Investigation [equal], Writing—review & editing [equal]), Paulo C.T. Souza (Conceptualization [equal], Methodology [equal], Writing—original draft [equal], Writing—review & editing [equal]), and Juliette Martin (Conceptualization [equal], Data curation [equal], Investigation [equal], Methodology [equal], Supervision [equal], Validation [equal], Visualization [equal], Writing—original draft [equal], Writing—review & editing [equal])

Supplementary data

Supplementary data are available at Bioinformatics Advances online.

Conflict of interest

None declared.

Funding

This work has been supported by CNRS (G.P.P., C.G., P.C.T.S., and J.M.), and a research collaboration agreement with PharmCADD (G.P.P. and P.C.T.S.)

Data availability

All the models analyzed in this article and the data used in Figs 1 and 2 are available in the Zenodo archive https://zenodo.org/records/14810843

References

  1. Abramson J, Adler J, Dunger J  et al.  Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature  2024;630:493–500. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Almodóvar-Rivera CM, Zhang Z, Li J  et al.  A modular chemistry platform for the development of a cereblon E3 Ligase-Based partial PROTAC library. Chembiochem  2023;24:e202300482. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Anon. AI’s potential to accelerate drug discovery needs a reality check. Nature  2023;622:217. [DOI] [PubMed] [Google Scholar]
  4. Bashore FM, Foley CA, Ong HW  et al.  PROTAC linkerology leads to an optimized bivalent chemical degrader of polycomb repressive complex 2 (PRC2) components. ACS Chem Biol  2023;18:494–507. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Basu S, Wallner B.  DockQ: a quality measure for Protein-Protein docking models. PLoS One  2016;11:e0161879. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Békés M, Langley DR, Crews CM.  PROTAC targeted protein degraders: the past is prologue. Nat Rev Drug Discov  2022;21:181–200. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Brahma R, Raghuraman H.  Characterization of a novel MgtE homolog and its structural dynamics in membrane mimetics. Biophys J  2024;123:1968–83. 10.1016/j.bpj.2023.11.3402 [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Bryant P, Noé F.  Improved protein complex prediction with AlphaFold-multimer by denoising the MSA profile. PLoS Comput Biol  2024a;20:e1012253. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Bryant P, Noé F.  Structure prediction of alternative protein conformations. Nat Commun  2024b;15:7328. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Bryant P, Pozzati G, Elofsson A.  Improved prediction of protein-protein interactions using AlphaFold2. Nat Commun  2022;13:1265. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Burslem GM, Crews CM.  Proteolysis-Targeting chimeras as therapeutics and tools for biological discovery. Cell  2020;181:102–14. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Collins KW, Copeland MM, Kotthoff I  et al.  Dockground resource for protein recognition studies. Protein Sci  2022;31:e4481. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Crew AP, Raina K, Dong H  et al.  Identification and characterization of von Hippel-Lindau-Recruiting proteolysis targeting chimeras (PROTACs) of TANK-binding kinase 1. J Med Chem  2018;61:583–98. [DOI] [PubMed] [Google Scholar]
  14. Cromm P.  Inducing Targeted Protein Degradation: From Chemical Biology to Drug Discovery and Clinical Applications. John Wiley & Sons, 2022. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Cyrus K, Wehenkel M, Choi E-Y  et al.  Impact of linker length on the activity of PROTACs. Mol Biosyst  2011;7:359–64. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Chai Discovery. Chai-1: decoding the molecular interactions of life. bioRxiv. http://biorxiv.org/lookup/doi/10.1101/2024.10.10.615955, 2024, preprint: not peer reviewed.
  17. Dixon T, MacPherson D, Mostofian B  et al.  Predicting the structural basis of targeted protein degradation by integrating molecular dynamics simulations with structural mass spectrometry. Nat Commun  2022;13:5884. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Duran-Frigola M, Cigler M, Winter GE.  Advancing targeted protein degradation via multiomics profiling and artificial intelligence. J Am Chem Soc  2023;145:2711–32. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Evans R, O’Neill M, Pritzel A  et al. Protein complex prediction with AlphaFold-Multimer. bioRxiv, p.2021.10.04.463034. https://www.biorxiv.org/content/biorxiv/early/2022/03/10/2021.10.04.463034, 2022, preprint: not peer reviewed (26 September 2022, date last accessed).
  20. Fu L, Niu B, Zhu Z  et al.  CD-HIT: accelerated for clustering the next-generation sequencing data. Bioinformatics  2012;28:3150–2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Gao M, Nakajima An D, Parks JM  et al.  AF2Complex predicts direct physical interactions in multimeric proteins with deep learning. Nat Commun  2022;13:1744. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Gut JA, Lemmin T. Dissecting AlphaFold’s capabilities with limited sequence information. bioRxiv. p.2024.03.14.585076. https://www.biorxiv.org/content/10.1101/2024.03.14.585076v2.abstract, 2024, preprint: not peer reviewed (26 August 2024, date last accessed).
  23. Hubbard SJ, Thornton JM. naccess. Department of Biochemistry and Molecular Biology, University College London. http://www.bioinf.manchester.ac.uk/naccess
  24. Jaime-Figueroa S, Buhimschi AD, Toure M  et al.  Design, synthesis and biological evaluation of proteolysis targeting chimeras (PROTACs) as a BTK degraders with improved pharmacokinetic properties. Bioorg Med Chem Lett  2020;30:126877. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Kim W, Mirdita M, Karin EL  et al. Rapid and sensitive protein complex alignment with foldseek-Multimer. Nat Methods 2025;22:469–72. 10.1038/s41592-025-02593-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Ko J, Lee J. Can AlphaFold2 predict protein-peptide complex structures accurately? bioRxiv. http://biorxiv.org/lookup/doi/10.1101/2021.07.27.453972, 2021, preprint: not peer reviewed.
  27. Krishna R, Wang J, Ahern W  et al.  Generalized biomolecular modeling and design with RoseTTAFold All-Atom. Science  2024;384:eadl2528. [DOI] [PubMed] [Google Scholar]
  28. Lewis S, Hempel T, Jiménez-Luna J  et al. Scalable emulation of protein equilibrium ensembles with generative deep learning. bioRxiv. http://biorxiv.org/lookup/doi/10.1101/2024.12.05.626885, 2024, preprint: not peer reviewed.
  29. Li F, Hu Q, Zhang X  et al.  DeepPROTACs is a deep learning-based targeted degradation predictor for PROTACs. Nat Commun  2022;13:7133. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Li W, Godzik A.  Cd-hit: a fast program for clustering and comparing large sets of protein or nucleotide sequences. Bioinformatics  2006;22:1658–9. [DOI] [PubMed] [Google Scholar]
  31. Mirabello C, Wallner B.  DockQ v2: improved automatic quality measure for protein multimers, nucleic acids, and small molecules. Bioinformatics  2024;40:btae586. 10.1093/bioinformatics/btae586 [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Mirdita M, Schütze K, Moriwaki Y  et al.  ColabFold: making protein folding accessible to all. Nat Methods  2022;19:679–82. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Noé F, Olsson S, Köhler J  et al.  Boltzmann generators: sampling equilibrium states of many-body systems with deep learning. Science  2019;365:eaaw1147. 10.1126/science.aaw1147 [DOI] [PubMed] [Google Scholar]
  34. Pereira GP, Jiménez-García B, Pellarin R  et al.  Rational prediction of PROTAC-compatible protein-protein interfaces by molecular docking. J Chem Inf Model  2023;63:6823–33. [DOI] [PubMed] [Google Scholar]
  35. Pettersson M, Crews CM. PROteolysis TArgeting Chimeras (PROTACs)—past, present and future. Drug Discov Today Technol 2019;31:15–27. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Quemener E, Corvellec M.  SIDUS—the solution for extreme deduplication of an operating system. Linux J  2013;235:3–3. [Google Scholar]
  37. Roel-Touris J, Bonvin AMJJ, Jiménez-García B.  LightDock goes information-driven. Bioinformatics  2020;36:950–2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Roney JP, Ovchinnikov S.  State-of-the-art estimation of protein model accuracy using AlphaFold. Phys Rev Lett  2022;129:238101. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Rovers E, Schapira M.  Benchmarking methods for PROTAC ternary complex structure prediction. J Chem Inf Model  2024;64:6162–73. [DOI] [PubMed] [Google Scholar]
  40. Roy R, Al-Hashimi HM.  AlphaFold3 takes a step toward decoding molecular behavior and biological computation. Nat Struct Mol Biol  2024;31:997–1000. [DOI] [PubMed] [Google Scholar]
  41. Senior AW, Evans R, Jumper J  et al.  Improved protein structure prediction using potentials from deep learning. Nature  2020;577:706–10. [DOI] [PubMed] [Google Scholar]
  42. Shehu A, Clementi C, Kavraki LE.  Modeling protein conformational ensembles: from missing loops to equilibrium fluctuations. Proteins  2006;65:164–79. [DOI] [PubMed] [Google Scholar]
  43. Stahl K, Graziadei A, Dau T  et al.  Protein structure prediction with in-cell photo-crosslinking mass spectrometry and deep learning. Nat Biotechnol  2023;41:1810–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Vreven T, Moal IH, Vangone A  et al.  Updates to the integrated Protein-Protein interaction benchmarks: docking benchmark version 5 and affinity benchmark version 2. J Mol Biol  2015;427:3031–41. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Wallner B.  AFsample: improving multimer prediction with AlphaFold using massive sampling. Bioinformatics (Oxford, England)  2023;39:btad573. [DOI] [PMC free article] [PubMed] [Google Scholar]
  46. Weng G, Li D, Kang Y  et al.  Integrative modeling of PROTAC-Mediated ternary complexes. J Med Chem  2021;64:16271–81. [DOI] [PubMed] [Google Scholar]
  47. Wohlwend J, Corso G, Passaro S  et al. Boltz-1 Democratizing Biomolecular Interaction Modeling. bioRxiv. 10.1101/2024.11.19.624167, 2024, preprint: not peer reviewed. [DOI]
  48. Yin R, Feng BY, Varshney A  et al.  Benchmarking AlphaFold for protein complex modeling reveals accuracy determinants. Protein Sci  2022;31:e4379. [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Zheng S, Tan Y, Wang Z  et al.  Accelerated rational PROTAC design via deep learning and molecular simulations. Nat Mach Intell  2022;4:739–48. [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

vbaf056_Supplementary_Data
vbaf056_supplementary_data.docx (3.8MB, docx)

Data Availability Statement

All the models analyzed in this article and the data used in Figs 1 and 2 are available in the Zenodo archive https://zenodo.org/records/14810843

Articles from Bioinformatics Advances are provided here courtesy of Oxford University Press

RESOURCES