ConforNets: Latents-Based Conformational Control in OpenFold3

Abstract: Models from the AlphaFold (AF) family reliably predict one dominant conformation for most well-ordered proteins but struggle to capture biologically relevant alternate states. Several efforts have focused on eliciting greater conformational variability through ad hoc inference-time perturbations of AF models or their inputs. Despite their progress, these approaches remain inefficient and fail to consistently recover major conformational modes. Here, we investigate both the optimal location and manner-of-operation for perturbing latent representations in the AF3 architecture. We distill our findings in ConforNets: channel-wise affine transforms of the pre-Pairformer pair latents. Unlike previous methods, ConforNets globally modulate AF3 representations, making them reusable across proteins. On unsupervised generation of alternate states, ConforNets achieve state-of-the-art success rates on all existing multi-state benchmarks. On the novel supervised task of conformational transfer, ConforNets trained on one source protein can induce a conserved conformational change across a protein family. Collectively, these results introduce a mechanism for conformational control in AF3-based models.

• The paper introduces ConforNets, employing channel-wise affine transforms on pre-Pairformer latents in OpenFold3 to induce controlled conformational diversity.
• It demonstrates superior success@100 rates on benchmarks like GPCR active states and cryptic pockets, outperforming prior methods such as AFsample3.
• The method facilitates practical applications in drug design by enabling at-will conformation transfer with minimal computational overhead.

Latents-Based Conformational Control in OpenFold3: An Analysis of ConforNets

Introduction and Background

The AlphaFold series has set a high standard for protein structure prediction but remains fundamentally limited in its ability to model conformational diversity and functional state heterogeneity. Traditional AF3-like predictors generate a single predominant conformation, inadequately covering critical alternative functional states such as cryptic pockets, domain motions, and ligand-induced transitions. Prior attempts to elicit this variability focus predominantly on empirical perturbations at the input (MSA subsampling/mutation) or output (diffusion-space steering) levels, yielding modest and inconsistent gains.

This study introduces ConforNets—channel-wise affine transformations applied to pre-Pairformer pairwise latent representations within OpenFold3-preview (OF3p)—that enable explicit, reusable conformational state control. ConforNets are optimized to induce either maximum conformational diversity or direct transfer of well-characterized conformations across protein family members. This methodology targets a fundamental limitation of previous approaches: the inability to specify or consistently induce desired functional states, particularly in poorly sampled regions of conformational space.

Figure 1: Schematic of ConforNets modifying AF3 pair latents for either diversity maximization or supervised conformation transfer.

Methodological Framework

ConforNets introduce a global, channel-wise affine transform ($\mathbf{W}, \mathbf{b}$) to the $L \times L \times c_z$ pairwise latents preceding the Pairformer trunk in OF3p. Key design decisions are validated through systematic ablation, highlighting the pre-Pairformer pair representation as the optimal intervention point for stable, high-fidelity conformational control.

There are two operational modalities:

1. Diversity-Driven Sampling:

A set of $k$ ConforNets are jointly trained to maximize pairwise diversity in either distogram (CDF MSE) or coordinate space (MSE), contrasting with prior entropy-maximization or input-perturbation strategies. Each ConforNet generates a structure sample via a deterministic or one-step diffusion rollout.

2. Supervised Conformation Transfer:

Given a single source protein in a desired conformation, a ConforNet is optimized (via MSE loss to reference structure) to encode the corresponding conformational bias in latent space. This trained ConforNet is then applied, without reparameterization, to any target of the same protein family to elicit the homologous conformational state.

Notably, modifications to ConforNet parameterization (removal of $\mathbf{W}$ or constraining it to diagonal form) reduce efficacy, implicating both scaling and mixing across channels as necessary for robust transfer and diversity.

Empirical Evaluation: Unsupervised Conformation Sampling

The efficacy of ConforNets is established through comparisons on established multi-state benchmarks: membrane transporters, cryptic pockets, domain motions, OOD60, and fold switchers. Datasets comprise manually curated state pairs representative of both local and global functional transitions. ConforNets consistently achieve superior success@100 rates across all benchmarks, outperforming prior state-of-the-art methods such as AFsample3, ConforMix, and BioEmu.

Figure 2: (a) Success@100 for membrane transporter state coverage as function of RMSD; (b) empirical energy landscapes sampled by ConforNets; (c) overlay of best predictions and ground truth conformations.

On membrane transporters, ConforNets outperform all baselines, efficiently sampling both inward- and outward-open states. Unlike incremental improvements seen with MSA-based or diffusion-guidance methods, ConforNets display clear bimodal sampling aligned with functional end states, as visualized by empirical free energy landscapes.

On cryptic pocket benchmarks, they induce local conformational changes corresponding to ligand-induced site opening, despite globally defined optimization objectives.

Figure 3: (a) Success@100 for cryptic pocket state coverage as function of RMSD; (b) empirical energy landscape and visual assessment of pocket fidelity.

Conformation Transfer: At-Will Control Across Protein Families

ConforNets demonstrate a substantial increase in induction frequency for functionally critical, undersampled states (e.g., GPCR active, kinase DFG-out, membrane transporter outward-open). The capacity to transfer conformational bias from a single source protein to structurally and sequentially divergent family members is unprecedented.

Figure 4: Empirical energy landscape shifts—default OF3p is unable or inconsistent in sampling states; ConforNets robustly bias toward targeted active conformations of GPCRs.

The most pronounced gains are observed in at-will induction (success@5): from 24% to 79% for GPCR active, 6% to 23% for kinase DFG-out, and 16% to 57% for transporter outward-open states. Template-based inference, prior diversity-maximizing approaches, and random sampling do not realize commensurate gains, emphasizing the specificity and reusability of learned ConforNets.

Figure 5: Transfer generalization—ConforNets trained on a single source induce correct states in targets with low structural similarity; overlays show predicted state congruence with ground truth.

Source-target similarity exerts only a minor effect on transfer success; selection of a family “centroid” as the source enables robust generalization. On kinases, moderately lower absolute rates reflect fundamental challenges in loop modeling, but even here, ConforNets enhance sampling of the correct state with observable improvements according to both strict and relaxed evaluation criteria.

Figure 6: Success and failure in DFG-out transfer; bar plots and example overlays detail ConforNet performance and limitations, specifically for challenging loop transitions.

Mechanistic Insights and Scalability

Systematic perturbation analyses conclusively demonstrate the pre-Pairformer latent space as the critical locus for conformational control—downstream (post-Pairformer) interventions produce unstable or short-lived modifications under full rollout. Channel mixing is necessary: even subtle off-diagonal weights are essential for success.

Training and inference with ConforNets introduce modest computational overhead, particularly when amortized across many samples or targets. Training cost scales with sequence length and remains practical under current GPU memory constraints.

Implications, Limitations, and Future Prospects

Practical Implications:

The ability to induce desired states at will is poised to accelerate downstream applications such as structure-based drug design, ligand docking, and the study of allosteric regulation. For GPCRs and transporters, in particular, this approach provides structural hypotheses previously unavailable through standard AF3-based pipelines.

Theoretical Impact:

These findings reinforce the hypothesis that deep structure predictors encode rich, implicit information about underlying conformational landscapes, even when trained with single-state supervision. Lightweight, parameter-efficient adaptations can therefore “unlock” latent capabilities, a concept parallel to prompt tuning, LoRA, and ControlNet in NLP and vision.

Limitations and Future Work:

The demonstrated transfer is primarily at the level of domain and large-scale conformational modes; detailed residue-level transfer and energetically calibrated sampling remain out of scope. There are promising directions in conditioning ConforNets on experimental constraints, ligand presence, or population-based ensemble characteristics. Expansion to complex, multi-subunit or ligand-bound states, as well as tighter integration with physically grounded methods (e.g., molecular dynamics), is also warranted.

Conclusion

ConforNets provide a robust, computationally tractable framework for modulating and transferring protein conformational states using OpenFold3. The demonstrated gains in both conformational diversity and at-will functional state induction suggest profound utility across computational structural biology. This paradigm paves the way for finer-grained, controllable protein modeling workflows and identifies a new class of architecture-agnostic, latent-space adaptation strategies relevant to broad generative modeling in molecular sciences.