AlphaFold 3: Protein Complex Prediction
- AlphaFold 3 (AF3) is a next-generation biomolecular structure prediction system that extends modeling from individual protein folding to direct simulation of protein complexes and molecular interactions.
- AF3 achieves an average RMSD of 1.61 Å and an ipTM of 0.803 on protein–protein benchmarks, underscoring competitive accuracy while revealing limitations compared to experimental structures.
- Its role as a structural prior supports downstream tasks like mutation-effect prediction and ligand-complex modeling, although challenges persist for flexible and disordered regions.
AlphaFold 3 (AF3) is a biomolecular structure prediction system that extends the AlphaFold lineage from monomer-centric protein folding toward direct modeling of molecular complexes. Independent studies describe AF3 as the latest AlphaFold generation for protein structure prediction, with practical use across protein–protein complexes and, more broadly, interactions involving nucleic acids, small molecules, and ions. Its importance derives not only from structural accuracy itself, but from its use as a structural prior or front end for downstream tasks such as mutation-effect prediction, ligand-complex modeling, and large-scale screening. At the same time, independent benchmarks consistently show that AF3-generated structures are useful but not interchangeable with experimental structures, particularly when conformational heterogeneity, intrinsically flexible regions, or interface-specific failure modes dominate the problem (Wee et al., 2024, Gopalan et al., 8 Oct 2025).
1. Scope, access model, and confidence outputs
AF3 is used in the literature primarily through the public AlphaFold Server, and this deployment model has shaped how it is benchmarked. Independent studies typically report server-generated predictions rather than locally controlled runs, and many of them note that low-level generation details such as seeds, recycles, template handling, or sampling settings are only sparsely documented. A related practical constraint is that AF3 itself was not open-sourced in the period covered by these studies, and only partially accessible through a limited online server; this is one reason why open reproductions and AF3-style systems emerged rapidly afterward (Liu et al., 2024).
In these studies, AF3 is treated as a stochastic model-selection problem rather than a single deterministic predictor. AF3 generates multiple configurations with different random seeds in some evaluation settings, making ranking and external quality assessment important for practical use (Han et al., 2024). The confidence outputs most frequently discussed are pTM and ipTM for global-complex and interface confidence, respectively, and pLDDT at residue level. In the protein–protein benchmarking literature, ipTM is operationally interpreted as follows: values indicate high-confidence interfaces, values suggest likely incorrect predictions, and values between $0.6$ and $0.8$ are ambiguous; pTM can be misleading when a correctly predicted larger partner dominates the score despite an incorrect interface or smaller chain (Wee et al., 2024). In disordered-protein work, residue-level is treated as a high-confidence threshold, but that threshold proves insufficient as a disorder reliability criterion in its own right (Gopalan et al., 8 Oct 2025).
2. Protein complexes and downstream energetic prediction
One of the clearest ways to assess AF3 is to ask whether its predicted complexes preserve enough structural signal for demanding downstream tasks. Two independent lines of work make this question concrete: mutation-induced binding free-energy prediction on protein–protein interfaces, and mutational scanning on viral receptor-binding complexes.
For protein–protein complexes, AF3 was benchmarked on SKEMPI 2.0 using 317 predicted complexes and 8,330 single-mutation samples. In that setting, the downstream target was mutation-induced binding free-energy change, , and AF3-derived structures retained most—but not all—of the predictive utility of experimental PDB structures. After structural alignment against experiment, the AF3 complexes had an average RMSD of $1.61$ Å, an average ipTM of $0.803$, and an average pTM of $0.847$. When MT-TopLap was retrained on AF3-derived features, it achieved and RMSE 0 kcal/mol, compared with 1 and RMSE 2 kcal/mol on the original PDB structures. The same study concludes that AF3 complex predictions are not reliable for intrinsically flexible regions or domains, and that the penalty from replacing experimental structures appears more clearly in absolute error than in correlation (Wee et al., 2024).
For viral protein–receptor interfaces, AF3 was used as the structure generator in an AF3-assisted MT-TopLap pipeline for SARS-CoV-2 RBD–ACE2 mutation-effect prediction. Across four experimental deep mutational scanning datasets, replacing experimental structures with AF3-predicted complexes caused only an average 3 decrease in PCC and an average 4 increase in RMSE. In the transfer-learning demonstration on the HK.3 variant, where no high-quality experimental HK.3–ACE2 complex was available in the PDB at the time, the AF3-assisted model achieved PCC 5 and RMSE 6. This suggests that AF3 can function as an effective structural surrogate when rapid-response modeling is more important than full equivalence to experiment (Wee et al., 2024).
| Benchmark setting | AF3-based result | Experimental/PDB comparator |
|---|---|---|
| SKEMPI 2.0 protein–protein interfaces | 7, RMSE 8 kcal/mol; average complex RMSD 9 Å | $0.6$0, RMSE $0.6$1 kcal/mol |
| SARS-CoV-2 RBD–ACE2 DMS | Average $0.6$2 PCC decrease and $0.6$3 RMSE increase; HK.3 PCC $0.6$4 | Experimental-structure MT-TopLap |
The broader implication is that AF3 is already good enough to support competitive downstream inference in many rigid or moderately deformable systems, but the substitution of AF3 for experiment is not lossless. A plausible implication is that AF3 is strongest when the relevant downstream quantity is primarily rank-order-sensitive and the interface geometry is close to a dominant rigid state.
3. Conformational state dependence and dynamic-state limitations
A recurring theme in the AF3 literature is that AF3 is substantially more reliable for dominant or low-activity states than for rare, active, or strongly rearranged states. GPCR benchmarks make this especially explicit. On a 75-structure GPCR set spanning classes A, B1, C, and F, AF3 did not outperform AF2 overall for conformational-state recovery. AF2 deformation ranged from $0.6$5 Å to $0.6$6 Å, whereas AF3 ranged from approximately $0.6$7 Å to $0.6$8 Å. Both models performed better on lower-activity conformations than on higher-activity conformations, and both showed smaller $0.6$9H3–H6 errors for inactive states than for active ones. The study’s conclusion is not that AF3 solves active-state GPCR modeling, but that AF2 remains more consistent on this benchmark and AF3 still struggles with activation-linked rearrangements (Chib et al., 24 Feb 2025).
A more constructive line of work shows that these failures can sometimes be corrected without changing AF3 weights, by changing the MSA prior. In state-aware protein–ligand prediction, AF-ClaSeq was used to derive purified sequence subsets associated with a desired structural state, and those custom MSAs were then supplied to AF3 together with ligand SMILES and no templates. For EGFR allosteric inhibitors, default AF3 failed with ligand RMSDs of $0.8$0 Å, $0.8$1 Å, and $0.8$2 Å on 8A2A, 8A2B, and 8A2D, respectively. With purified inactive-state sequence subsets, those RMSDs were reduced to $0.8$3 Å, $0.8$4 Å, and $0.8$5 Å. For IL-1β, the top-20 purified sequences yielded 40 AF3 structures that aligned perfectly with the experimental cryptic-pocket complex. This suggests that a substantial subset of AF3’s apparent ligand-prediction failures are really state-selection failures induced by heterogeneous MSAs, not irreducible ligand-placement failures (Xing et al., 30 May 2025).
A separate AF3-like line of work, ConforNets in OpenFold3-preview, further suggests that alternate conformations are encoded in AF3-style pair latents and can be steered by small channel-wise affine transforms of pre-Pairformer pair representations. In that framework, pre-Pairformer pair latents outperformed other intervention sites for alternate-state generation and family-level conformational transfer, supporting the view that AF3-class models contain latent conformational modes that are accessible but not normally expressed under default inference (Lee et al., 20 Apr 2026).
4. Flexibility, disorder, and hallucination
AF3’s limitations become sharper in flexible complexes and intrinsically disordered systems. In the protein–protein benchmark on SKEMPI, the strongest biological failure mode was linked to intrinsic flexibility: using PDB B-factors as a proxy for experimental mobility, residues with high B-factors tended also to have large per-residue RMSD between AF3 and the corresponding experimental structure. The authors therefore conclude that AF3 complex predictions are not reliable for intrinsically flexible protein regions or domains, particularly where disorder-to-order transitions, mobile domains, or induced fit are mechanistically important (Wee et al., 2024).
The most direct disorder benchmark used 72 highly disordered proteins from DisProt and defined “hallucinations” as disagreement between AF3-derived order/disorder calls and experimental disorder annotations. Across the analyzed residues, $0.8$6 aligned with DisProt, while $0.8$7 did not; that disagreement was decomposed into $0.8$8 hallucinations and $0.8$9 possible context-driven misalignment. More than 0 of proteins had less than 1 alignment with DisProt, and 2 of residues involved in biological processes showed hallucinations. The study also found that varying seeds did not produce significant variance in AF3 predictions: the model appeared relatively stable across tested seeds and ensembles, but not necessarily correct. This is why the paper argues that pLDDT alone is insufficient as a hallucination detector for IDRs (Gopalan et al., 8 Oct 2025).
A plausible implication is that AF3’s confidence stability in disordered regions may reflect internal consistency rather than calibrated uncertainty. In practice, this makes AF3 more dangerous in disorder-rich workflows than a purely noisy model would be, because apparently reproducible outputs can still be wrong in functionally consequential ways.
5. Confidence metrics, reranking, and interface-specific failure modes
Independent assessments converge on the point that AF3’s native confidence metrics are informative but incomplete. In the SKEMPI complex benchmark, 3 of complexes had ipTM 4 and 5 had pTM 6, yet confidence and actual alignment quality were only weakly related. When complexes were ranked by worst RMSD and by lowest ipTM, only six complexes overlapped between the two rankings, leading the authors to conclude that there is little correlation between ipTM and actual RMSD in that benchmark. Their practical recommendation is implicit but clear: ipTM alone is not a sufficient quality filter for downstream use (Wee et al., 2024).
External quality assessment confirms that AF3’s own ranking is strong on average but not uniformly optimal. On the ABAG-AF3 antibody–antigen dataset, AF3’s self-assessment baseline achieved ranking loss 7 and Top-10 mean DockQ-wave 8, outperforming the external QA models on average. However, the topological QA model TopoQA achieved lower ranking loss than AF3 on 17 of 35 targets and better Top-10 mean DockQ-wave on 16 of 35 targets. The paper’s conclusion is therefore nuanced: AF3 remains the best scorer on average in that benchmark, but external reranking remains useful for nearly half the targets (Han et al., 2024).
A more specific caution comes from CD47 antibody–antigen modeling, where AF3 was investigated as a front end for MM/GBSA-based affinity triage. AF3 performed best overall among the tested methods under the authors’ composite scoring scheme and reproduced a known 5TZ2 complex strongly enough that MM/GBSA on the AF3 structure deviated by only 9 kJ/mol from the experimental-structure estimate. Yet the same study identified a recurrent failure mode called “reverse docking,” in which antibody and antigen were docked in the opposite biologically reasonable orientation. For D2510, across 81 AF3 predictions, the authors reported 53 reverse-docked conformations and 23 reasonable conformations, often with relatively high ipTM and pTM; they reported a one-tailed 0 for the group difference. This establishes that high-confidence, biologically inverted interfaces are not merely hypothetical edge cases in antibody modeling (Xu et al., 18 Nov 2025).
6. Open reproductions, systems optimization, and AF3-derived methodology
AF3’s limited server access and computational cost rapidly produced an ecosystem of reproductions, systems optimizations, and downstream methods. HelixFold3 is explicitly presented as an open-source effort to replicate AF3’s capabilities, using PDB targets released before 2021-09-30 plus self-distillation data, diffusion inference, and confidence-guided ranking. In that report, HelixFold3 is said to achieve AF3-comparable performance on conventional ligands and nucleic acids, while still trailing AF3 on protein–protein complexes (Liu et al., 2024).
On the systems side, AF_Cache addresses AF3’s repeated-MSA bottleneck in high-throughput PPI screening by generating monomer MSAs once, reusing them across jobs, and substituting GPU-accelerated MMseqs2 for slower default search. For AF3 specifically, the paper reports a 5-fold MSA-generation speedup under its more realistic comparison, while using the same AF3 inference code and therefore providing no AF3 inference-time speedup (Narrowe et al., 3 Jun 2026). MegaFold targets AF3 training rather than inference, reporting up to 1 peak-memory reduction, up to 2 and 3 per-iteration training-time improvement on NVIDIA H200 and AMD MI250 GPUs, respectively, and training on 4 longer sequence lengths than PyTorch baselines without running out of memory (La et al., 24 Jun 2025). DCFold goes further by compressing AF3-like iterative inference into one Pairformer recycle and one diffusion step; it reports about 5 average inference acceleration while aiming for AF3-level benchmark accuracy (Zhang et al., 18 May 2026).
AF3 has also been repurposed methodologically. In experiment-guided AF3, AF3 is treated as a sequence-conditioned structural prior and combined with likelihoods from crystallographic density and NMR NOEs to sample posterior conformational ensembles rather than single structures. That framework reports ensembles that improve agreement with experimental data and, in some cases, fit density or restraints better than vanilla AF3 and even better than deposited structures on the experiment-based metric at hand (Maddipatla et al., 13 Feb 2025). In inverse design, AF3’s all-atom complex setting has motivated downstream models such as ADFLIP, which perform inverse protein folding conditioned on all-atom structural contexts that include ligands, nucleotides, and metal ions, although AF3 itself is not the predictor in those experiments (Yi et al., 4 Jul 2025).
AF3 is therefore best understood not as a closed endpoint, but as the organizing center of a broader technical stack: server-side prediction, open reproduction, systems acceleration, external QA, state steering, posterior sampling, and design workflows. The accumulated evidence suggests that AF3’s central scientific role is now twofold: as a practical biomolecular structure generator, and as a learned prior around which increasingly specialized structural-biology methods are being built.