We curated a dataset for conformational flexibility from the Docking Benchmark Set 5.5 (DB5.5) (Vreven et al., 2015), which comprises experimentally characterized (X-ray or cryo-EM) structures of bound protein complexes and their corresponding unbound protein subunits. Each protein target (with unbound and bound structures) is classified based on their unbound-to-bound root-mean-square-deviation (RMSD_UB) as rigid (RMSD_UB ≤ 1.2 Å), medium (1.2 Å < RMSD_UB ≤ 2.2 Å), or difficult (RMSD_UB ≥ 2.2 Å). Furthermore, due to the poor performance of AF and other predictor groups in predicting antibody–antigen targets in the recent CASP15-CAPRI round CASP15, 2022, we identified a subset comprising only antibody–antigen complexes (including single-domain antibodies or nanobodies) extracting all 67 antibody–antigen structures from the DB5.5 (Guest et al., 2020, Vreven et al., 2015) set. The comprehensive dataset includes 254 protein targets exhibiting binding-induced conformational changes.

For each protein target, we extracted the amino acid sequences from the bound structure and predicted a corresponding three-dimensional complex structure with the ColabFold implementation (https://github.com/YoshitakaMo/localcolabfold; Moriwaki, 2023) of the AlphaFold multimer v2.3.0 (released in March 2023) for the 254 benchmark targets from DB5.5. Being trained on experimentally characterized structures deposited in the PDB, AF is expected to produce models analogous to the PDB structures. Since most of the benchmark targets in DB5.5 were included in AF training, there would be training bias associated with their predictions (i.e., our measured success rates are an upper bound). However, since both unbound and bound structures exist for the benchmark targets in the PDB, we first investigated whether AFm exhibits any bias toward either unbound or bound forms for the same protein sequence. Appendix 1—figure 1 compares the Cα-RMSD of all protein partners (calculated on a per-chain basis) of the AFm predicted complex structures from the bound (B) and unbound (U) crystal structures on a log-log scale (a few AFm predicted models were 20 Å apart from both bound and unbound structures). As evident from Supplementary Fig.S1A, the protein partners from the AFm top-ranked model deviate from both unbound and bound forms and skew more often toward the bound state. Antibody–antigen targets further demonstrate a similar trend, however with fewer targets predicted within sub-angstrom accuracy to the bound form (29.7% for Ab-Ag targets as opposed to 41% for DB5.5). We also calculated the TM-scores (Zhang and Skolnick, 2004) of the AFm predicted complex structures with respect to the bound and the unbound crystal structures (Appendix 1—figure 2). As TM-scores reflect a global comparison between structures and are less sensitive to local structural deviations, no strong conclusions could be derived. This is in agreement with our intuition that since both unbound and bound states of proteins will share a similar fold, and AF can predict structures with high TM-scores in most cases, gauging the conformational deviations with TM-scores would be inconclusive.

AF employs multiple sequence alignments with a multi-track attention-based architecture to predict three-dimensional structures of proteins and complexes. Further, for each structural prediction, it provides a residue-level confidence measure: the predicted local-distance difference test (pLDDT), estimating the agreement between predicted model to an experimental structure based on the Cα LDDT test (‘Methods’). Tunyasuvunakool et al. analyzed pLDDT confidence measures for the human proteome demonstrating the correlation between lower pLDDT scores with higher disordered regions in protein structures (Tunyasuvunakool et al., 2021). Building on this observation, we evaluated whether there is a correlation between AF pLDDT confidence metric and the experimental metrics of conformational change between unbound and bound structures. In this regard, we compared the computational (AF-pLDDT) and experimental (per-residue RMSD and LDDT) metrics against each other.

As a reference, we first superimposed the unbound partners over the bound structures and calculated residue-wise Cα deviations to determine the per-residue RMSD_BU values. LDDT_BU was measured by calculating the local distance differences in the unbound structure relative to the bound form. These metrics capture the extent of motion in the unbound–bound transitions for each of the protein targets. Next, we compared the per-residue pLDDT score from AFm predicted monomer models with the experimental metrics. Figure 1A and B shows the results for two representative protein targets: kinase-associated phosphatase in complex with phospho-CDK2 (1FQ1; Song et al., 2001) and TGF-β receptor with FKBP12 domain (1B6C; Huse et al., 1999). In both cases, pLDDT confidence scores correlate with the experimental measurements of binding: pLDDT decreases as LDDT_BU decreases and RMSD_BU increases. This is further illustrated with the AF2 predicted structures of the two targets superimposed over the bound structures (Figure 1C). In regions of low confidence/pLDDT (highlighted in red), the prediction is inaccurate, but higher confidence/pLDDT regions (highlighted in blue) have high accuracy of prediction with the bound form. The results for the benchmark set (Figure 1—figure supplements 1 and 2) show similar trends for most targets. The pLDDT, thus, can suggest protein residues that move upon binding.

When the prediction accuracy is lower, it is often evident from lower confidence metrics (such as average pLDDT or PAE). However, for AFm complex predictions, the confidence metrics of the overall prediction do not correlate with the accuracy of the docked prediction, that is, even if the complex exhibits higher confidence, the docking interfaces could be incorrect. Figure 2 shows a few examples of failed AFm predictions including rigid (2FJU [Jezyk et al., 2006]), medium (5VNW [McMahon et al., 2018]), and flexible targets (1IB1 [Vetter et al., 1999], 2FJG [Fuh et al., 2006]). In all the examples, the AFm model (highlighted in red to blue based on residue-wise pLDDT) is superimposed over an individual binding partner, and the bound structure is highlighted in pale green. AFm models predict the individual subunits (protein partners) accurately in almost all scenarios; however, the docking orientation is incorrect.

We investigated whether any of the AF predictive metrics could be repurposed for distinguishing native-like binding sites from non-native ones. That is, can one utilize pLDDT or PAE from AFm models to determine whether the predicted docked complex has the accurate binding orientation? Thus, we evaluated accuracy with the DockQ score, the standard metric for docking model quality (Basu and Wallner, 2016). DockQ ∈[0,1] combines interface RMSD (Irms), fraction of native-like contacts (fnat), and ligand-RMSD (Lrms). DockQ scores above 0.23 correspond to models with a CAPRI quality of ‘acceptable’ or higher. As an acceptable quality target implies docked decoys are in the near-native binding region, we chose a binary classification of success with a threshold of DockQ = 0.23. We then tested how well DockQ correlated with several AFm-derived metrics: (1) interface residues: the number of interface residues (atoms of residues on one partner within 8 Å from an atom on another partner); (2) interface contacts: the number of interface contacts between the residues on the interface (Cβ atoms within 5 Å); (3) average pLDDT, determined by averaging over the per-residue LDDT score of the entire protein complex; and (4) interface-pLDDT, determined by averaging the per-residue LDDT score only over the predicted interfacial residues (as identified in case a).

Figure 3A shows the classification accuracy of each of these metrics with a receiver-operating characteristics curve. The interface-pLDDT metric stands out with a higher true positive rate (TPR) with an area under curve (AUC) of 0.86. With interface-pLDDT as a discriminating metric, we tested multiple thresholds to estimate the optimum cutoff for distinguishing near-native structures (defined as an interface-RMSD <4 Å) from the predictions. Figure 3B summarizes the performance with a confusion matrix for the chosen interface-pLDDT cutoff of 85. 79% of the targets are classified accurately with a precision of 75%, thereby validating the utility of interface-pLDDT as a discriminating metric to rank the docking quality of the AFm complex structure predictions. With newer structure prediction tools such as AlphaFold3 (Abramson et al., 2024) and ESM3 (Hayes et al., 2024) being released, investigating features that could predict flexible residues or interface site would be valuable as this information may guide local docking. This discrimination is also evident in the highlighted interface residues in Figure 2, where the AFm predicted models have lower confidence at predicted interfaces (red). Finally, we show the trend between DockQ scores and interface-pLDDT for each target in Figure 3C. The interface-pLDDT threshold of 85 (dashed line) thus can serve as the AF-derived metric to distinguish acceptable quality docked predictions from incorrect models.

With metrics to identify the flexible regions in the protein and the docking accuracy of generated docked models, we next fused AFm with our docking protocol, ReplicaDock 2.0 (Harmalkar et al., 2022), to build a protocol for (1) improving on incorrect AF docking predictions and producing alternate, near-native binding models and (2) capturing backbone conformational changes with our induced-fit protocol ReplicaDock2.0 (Harmalkar et al., 2022). We named the protocol AlphaRED. AlphaRED uses AFm predicted structures as the primary template, estimates docking accuracy metrics, and initiates global docking or refinement protocols as required.

Figure 4 illustrates this docking pipeline. After AFm predicts a model from the protein sequences, we calculate the interface-pLDDT to determine the docking scheme to follow. If the AFm model is likely to be inaccurate (interface pLDDT < 85), we initiate a global replica exchange docking simulation to explore the protein conformational landscape and identify putative binding sites. On the other hand, if the interface-pLDDT > 85 for the AFm predicted model, the docked complex is likely in the correct binding orientation. This implies the global docking stage of the protocol can be skipped and local docking simulations can be directly initiated from the complex coordinates. Global docking follows an exhaustive, rigid-body search (no backbone moves) between the protein partners to sample putative landscapes in the energy landscape. An unbiased global docking simulation is initiated by randomizing the spatial orientation of protein partners from the input structure. The replica exchange MC routine ReplicaDock 2.0 performs rigid-body rotations (8°) and translations (4 Å). Sampled decoys are clustered from all replicas (based on energies and structural similarity) and the five top clusters are passed along for flexible local docking.

For flexible local docking, we perform aggressive backbone moves (backrub + kinematic closure, ‘Methods’) on candidate encounter complexes (clustered decoys), with fine rigid-body rotations and translations. To narrow conformational sampling, backbone moves are explicitly performed over residues identified as ‘mobile’ based on the per-residue pLDDT metric (residue pLDDT < 80). Unlike ReplicaDock 2.0 that performs induced-fit over putative interfaces, this approach targets backbone motions over these predicted mobile residues, reducing the sampling space. Local docking decoys are further refined for side-chain packing and minimization to obtain the final docked structures (details in ‘Methods’). The methodological advancements and Rosetta movers in AlphaRED are further detailed in the ‘Methods’ section.

We investigated AlphaRED’s performance on all 254 benchmark targets (Figure 5). Ninety-seven targets under the threshold of interface-pLDDT (≤85) were passed to the global docking branch. Targets with interface-pLDDT over 85 proceeded directly to local docking refinement. For all benchmark targets, we compared AlphaRED performance of the top-scoring decoys against initial AFm-predicted complex structures. Figure 5A. shows the interface-RMSD (Irms) of the AFm and AlphaRED predictions from the bound structure, respectively. The lower Irms values indicate that AlphaRED improves on existing predictions for almost all targets. For targets where AFm prediction is determined to be a failure (interface-pLDDT ≤ 85, red), AlphaRED demonstrates a vast improvement in Irms for 93 out of 97 targets. Additionally, for targets where AFm prediction is considered acceptable (interface-pLDDT > 85), local docking slightly improves performance. AlphaRED captures lower interface-RMSDs (under 10 Å) for targets where AFm models dock at binding sites ∼40 Å away. Figure 5B demonstrates the improvement in recapitulating native-like contacts (fnat) with AlphaRED.

Figure 5C shows the performance of the subset of antibody–antigen targets in the benchmark. Antibody targets are critical for understanding adaptive immune responses and for the design and engineering of antibody therapeutics (Chungyoun and Gray, 2023). However, antibodies have proven challenging for deep learning methods, especially those reliant on multiple sequence alignments, as each antibody evolves in a different organism, and their antigens evolve on a different timescale altogether. Finally, to evaluate docking success rates, we calculate DockQ for top predictions from AFm and AlphaRED, respectively (Figure 5D). AlphaRED demonstrates a success rate (DockQ > 0.23) of 63% for the benchmark targets. Particularly for Ab-Ag complexes, AFm predicted acceptable or better quality docked structures in only 20% of the 67 targets. In contrast, the AlphaRED pipeline succeeds in 43% of the targets, a significant improvement. Most of the improvements in the success rates are for cases where AFm predictions are worse. For targets with good AFm predictions, AlphaRED refinement results in minimal improvements in docking accuracy.

Figure 6 highlights a global docking (a) and local docking (b) example for targets 2FJU and 5C7X, respectively. Starting from the incorrect AFm prediction (orange), AlphaRED samples over the conformational landscape to identify a top-scoring decoy (blue) with 2.6 Å Irms from the native (gray). Figure 6b shows the extent of backbone sampling with ReplicaDock 2.0 local docking. The top-scoring decoy (blue) samples backbone closer to the bound form, improving model quality and docking accuracy.

All results presented thus far may be biased by the fact that these benchmark target structures were used in the AFm training. The ultimate challenge for protein structure prediction protocols is to perform successfully over blind targets such as those in CASP (Critical Assessment of protein Structure Prediction) or CAPRI (Critical Assessment of PRotein Interactions) competitions (Kryshtafovych et al., 2021, Lensink et al., 2021). CASP15 (Summer 2022) provided multiple protein docking targets (CASP15, 2022) that were not included in AFm training, allowing an unbiased evaluation of our AlphaRED pipeline (Lensink et al., 2023). Thus, we tested the protocol on the five heterodimeric nanobody–antigen complexes where most of the groups performed poorly (Figure 7).

Since the nanobody–antigen complexes were CASP targets, we did not have unbound structures, rather only the sequences of individual chains. Therefore, for each target, we employed the AlphaRED strategy as described in Figure 4. All targets predicted with AFm had low interface-pLDDT, thereby demanding global docking. This is unsurprising since the targets were nanobody–antigen targets and their CDRs, particularly CDR H3, are not conserved with a scarcity of co-evolution data with the antigen (Adolf-Bryfogle et al., 2015). For representative target T205, our docking strategy improves the performance drastically (interface RMSD 11.4 Å for AFm model to 2.84 Å for AlphaRED) and binds in the correct site. The interface scores versus interface-RMSD plot shows a distinct funnel with low-energy medium-quality structures (Figure 6, top). Since the crystal structures are not yet released, the reference structure here is the top model predicted for each category in CASP15 (Lensink et al., 2023). For all the targets, Figure 6, bottom shows similar improvements for other nanobody–antigen complexes. These cases validate our strategy for blind targets and demonstrate the ability of AlphaRED to serve as a robust pipeline, integrating AF with biophysical attributes to better predict protein complex structures.

For each target in the DB5.5 dataset, we first extracted the corresponding FASTA sequence for the bound complex and then obtained AF predicted models with the ColabFold v1.5.2 (Ovchinnikov, 2021) implementation of AF (Jumper et al., 2021) and AlphaFold-multimer (v.2.3.0) (Evans et al., 2021). Each prediction run was performed without templates, with automatic alignments and the default number of recycles to generate five relaxed predictions. Each AF prediction includes a per-residue pLDDT (predicted LDDT) measurement (Mariani et al., 2013), a confidence measure in prediction accuracy, and predicted template alignment (pTM) score (Zhang and Skolnick, 2004). The models were structurally compared with the unbound and bound structures (deposited in the PDB) for measuring flexibility, similarity, and accuracy of docking prediction.

Structures of proteins deposited in the PDB (Berman et al., 2000) provide a static representation of the native state of the protein. However, structural diversity has been captured by experimental techniques to identify different states of a protein in diverse physiological or chemical states, for example, catalysis (Kingsley and Lill, 2015), transport (Gora et al., 2013), and ligand binding (Gunasekaran and Nussinov, 2007). For protein docking challenges in particular, conformational changes are binding-induced, leading to structural differences between unbound and bound structures of protein targets.

To measure the conformational change in protein structures, we calculated two metrics: Cα RMSD and LDDT (Mariani et al., 2013). To get a detailed representation of the intrinsic motion of a protein, we calculated RMSDs at a residue level, that is, per-residue Cα RMSD for each residue of a protein target. The sequences + structures of unbound and bound proteins were aligned and the RMSDs were calculated for the aligned residues. The total sequence lengths were also matched and lingering end-termini residues were trimmed to ensure structural and sequential similarity.

LDDT is a superimposition-free score that estimates local distance differences in a model relative to a reference structure (Mariani et al., 2013). Unlike the global distance test (Zemla, 2003) score based on rigid-body superimposition, the LDDT score measures the conserved local interactions in the protein model to the reference. For every residue, it computes the distance between all pair of atoms D(i,j) in both the model and the reference structure (bound) within a threshold (defined as the inclusion radius, generally set to 10 Å). For each pairwise distance in both distance vectors, if the distance is within the threshold, the distance is considered conserved and the fraction of conserved distances is calculated. The final LDDT score is the average of this fraction for the tolerances of 0.5, 1, 2, and 4 Å.

For a protein structure with N number of residues, the overall LDDT score can be given as follows:(1)dists_to_score(i,j)

where norm is the normalization factor(2)norm=1∑i,jdists_to_score(i,j)

and score(i,j) is the LDDT score for the residue i with respect to every other residue jscore(i,j)=0.25⋅{ bool[ΔD(i,j)<0.5] +bool[ΔD(i,j)<1.0] +bool[ΔD(i,j)<2.0] +bool[ΔD(i,j)<4.0] }

ΔD(i,j) denotes the absolute difference between Dtrue(i,j) and Dpredicted(i,j) calculated as follows:(3)ΔD(i,j)=|Dtrue(i,j)−Dpredicted(i,j)|

where Dtrue(i,j) and Dpredicted(i,j) denote the distances between the Cα coordinates of the ith residue and the jth residue for the true (reference) and predicted (model) structures, respectively. Let xik and yik represent the kth coordinate of the Cα atom in the ith residue in the reference (true) structure and predicted structure respectively, such that(4)Dtrue(i,j)=∑k=13(xik−xjk)2 andDpredicted(i,j)=∑k=13(yik−yjk)2

Finally, the distances to score (dists_to_score(i,j)) are computed as those pairwise distances within an inclusion radius (cutoff = 10 Å). mij is the mask value (1 or 0) indicating if the jth coordinate of the Cα atom in the ith residue exists in the true structure.(5)dist_to_score(i,j)={1ifDtrue(i,j)<cutoff⋅mij⋅mji⋅(1−δjN)0otherwise

where δ=KroneckerDelta.

The advantage of the LDDT measurement lies in the estimation of relative domain orientations in multi-domain proteins or concerted motions (e.g., hinge-like moves in closed and apo proteins). In these cases, the RMSDs would be relatively high for all residues in the mobile domain; however, since the inter-residue distances within the domains are conserved, they would provide an inaccurate depiction of flexibility for the protein. Estimating both RMSDs and LDDT scores allows us to obtain a nuanced perspective of flexibility during protein association based on experimental structures.

Using AlphaFold2 as a structural module, we built a pipeline for protein–protein docking to better predict protein complex structures with relatively higher accuracy. As illustrated in Figure 4, given a sequence of a protein complex, we use the ColabFold implementation of AF2-multimer to obtain a predictive template. An interface-pLDDT filter determines the accuracy of the docking prediction of the top-ranked model from AFm. If the interface-pLDDT ≤ 85, the prediction has lower confidence in the docking orientation, and the protocol initiates a rigid, global docking search with ReplicaDock 2.0. Implementation of ReplicaDock 2.0 (global docking) is similar to the version reported in prior work (Harmalkar et al., 2022). Each simulation initiates eight trajectories across three temperature replicas with inverse temperatures set to 1.5^–1 kcal^–1.mol, 3^–1 kcal^-1.mol, and 5^–1 kcal^–1.mol, respectively. Across each replica within each trajectory, rigid-body perturbations (4 Å translations and 8° rotations) are performed for an exhaustive global search. Next, we perform an energy-based clustering of the models to obtain diverse and energetically favorable clusters. Five cluster centers (decoys) are selected and passed to the flexible local docking stage to sample conformational changes.

On the other hand, if the interface-pLDDT > 85, the binding orientation has higher confidence and the protocol directly performs a flexible local docking simulation skipping the rigid, global docking. In this stage, we perform smaller rigid-body perturbations (1 Å translations and 3° rotations) and aggressive backbone moves using a set of backbone and side-chain movers: Rosetta Backrub (Smith and Kortemme, 2008), Balanced Kinematic Closure (BalancedKIC), and Sidechain. The sampling weights are biased such that backbone and side-chain movers are weighted higher than rigid-body moves (3:1 weightage for backbone:rigid-body moves). We perform directed backbone sampling by focusing on predicted mobile residues (per-residue pLDDT < 80). This is automated with the BFactorResidueSelector that selects contiguous sets of residues below the specified pLDDT threshold.

However, unlike the induced-fit strategy in ReplicaDock (Harmalkar et al., 2022), we perform backbone sampling directed only on the mobile residues (with per-residue pLDDT < 80) identified from the AF model. We automate it using the BFactorResidueSelector to select contiguous sets of residues below the specified pLDDT threshold in the prior section. This residue subset is passed along to the backbone movers to sample backbone moves along with small rigid-body moves. Sampled decoyed are then refined, that is, undergo side-chain packing and minimization, to output docked decoys. The best ranked decoys based on interface scores are then identified as the top-scoring structures.