OpenFold3-preview: Modular Protein Prediction

• OpenFold3-preview is a protein structure prediction system that employs modular input embedding, iterative pairwise reasoning via the Pairformer, and a diffusion-based denoising head.
• It integrates ConforNets for channel-wise latent modulation, enabling efficient, reusable control over protein conformations and alternate-state prediction.
• The system supports supervised conformational transfer across protein families, significantly improving prediction success rates for challenging benchmarks.

OpenFold3-preview (OF3p) is a protein structure prediction system based on the AlphaFold3 (AF3) architecture, characterized by modular embedding, explicit pairwise neural reasoning, and a denoising diffusion head. OF3p forms a chassis for recent innovations in structural variability modeling, notably the integration of channel-wise latent transforms ("ConforNets") to enable efficient, reusable control over protein conformation. This approach delivers state-of-the-art alternate-state prediction and introduces a new paradigm for supervised conformational transfer across protein families (Lee et al., 20 Apr 2026).

1. Core Architecture of OpenFold3-preview

The OF3p model consists of three principal modules: input embedding, iterative pairwise reasoning via the Pairformer, and a diffusion-based structure generation head. For a protein sequence $x$ of length $L$ and a multiple sequence alignment (MSA) $\mathcal{M}$, OF3p computes initial embeddings $s^{\mathrm{pre}} \in \mathbb{R}^{L \times c_s}$ and $$z^{\mathrm{pre}} \in \mathbb{R}^{L \times L \times c_z}$$ using an embedding network:

$$(s^{\mathrm{pre}}, z^{\mathrm{pre}}) = \mathrm{Embedder}(x, \mathrm{Subsample}(\mathcal{M}))$$

where $c_s$ and $c_z$ are the dimensions of single- and pairwise channels (typically 256 and 128).

The Pairformer trunk recursively refines these representations through a combination of triangular updates and self-attention, usually over $R$ recycles:

$$(s^{\mathrm{post}}, z^{\mathrm{post}}) = \mathrm{Pairformer}(s^{\mathrm{pre}}, z^{\mathrm{pre}})$$

The diffusion head is a variance-exploding (EDM) denoiser that iteratively predicts 3D coordinates or distance maps, consuming both the initial and refined embeddings at each generation step. Importantly, $L$0 encodes residue–residue couplings and is the principal target for representation-level interventions in subsequent conformational control.

2. ConforNets: Channel-wise Latent Modulation

ConforNets introduce a global, channel-wise affine transform $L$1 on the pair latent $L$2. The general form is:

$L$3

Here, $L$4 is the latent (typically $L$5), $L$6 is trainable and initialized as the identity, and $L$7 is initialized as zero. A diagonal constraint on $L$8 results in a "scaling and shift" transform:

$L$9

or in vector form, $\mathcal{M}$0 with $\mathcal{M}$1.

The key operational point is immediately prior to the last Pairformer pass: $\mathcal{M}$2 is replaced with $\mathcal{M}$3 globally for all sequence positions. This global, channel-only modulation distinguishes itself from residue-specific or edge-specific latent interventions.

3. Protein-Agnostic and Length-Independent Control

ConforNets' transformation acts solely on the latent channel axis, ensuring the following attributes:

• Global Application: The same $\mathcal{M}$4 parameters are applied uniformly across all residue pairs.
• Length Independence: A $\mathcal{M}$5 trained on a protein of length $\mathcal{M}$6 can be directly reused on a protein of length $\mathcal{M}$7.
• Homolog Generalization: ConforNets enable a form of "universal bias" towards specific conformational states that generalize across related protein families.
• Efficiency: No per-protein latent optimization is required at inference, facilitating rapid deployment.

This protein-agnostic mechanism supports transfer learning scenarios and circumvents the limitations associated with inference-time latent modification.

4. Benchmarks for Unsupervised Alternate-State Prediction

ConforNets, when trained for diversity, exhibit competitive performance on the generation of alternate biologically relevant conformational states across a panel of established benchmarks:

• Dataset: 104 proteins, each with two conformational states, across five benchmark categories: cryptic pockets (apo/holo), domain motions, OOD60 (out-of-distribution), membrane transporters, and fold switchers.
• Metric (success@B): The probability that at least one out of $\mathcal{M}$8 samples achieves backbone RMSD below a state-specific threshold ($\mathcal{M}$9; 1 Å–3 Å depending on benchmark category) to the reference conformation.
• Competing Methods: Default OF3p with MSA subsampling, shallow MSA sampling, AFsample3 (random MSA column masking), ConforMix (diffusion guidance), and BioEmu (separately trained model).

ConforNets are optimized by training $s^{\mathrm{pre}} \in \mathbb{R}^{L \times c_s}$0 independent transforms with random noise initialization, maximizing the pairwise discrepancy in structural metrics after a single deterministic denoise step. Empirical results (success@100) against competing approaches are as follows:

Benchmark	ConforNets (dist)	AFsample3	ConforMix
Cryptic–apo	48.8%	44.7%	—
Cryptic–holo	78.9%	73.6%	—
Domain motions	81.9%	80.6%	—
OOD60	60.7%	57.7%	—
Membrane transporters	51.1%	46.9%	—
Fold switchers	54.4%	48.7%	—

ConforNets consistently outperform AFsample3 and ConforMix by 5–15 percentage points, establishing a new state-of-the-art for structural diversity under the OF3p system (Lee et al., 20 Apr 2026).

5. Supervised Conformational Transfer Across Families

The conformational transfer task leverages supervised training of $s^{\mathrm{pre}} \in \mathbb{R}^{L \times c_s}$1 to align generated structures from a source protein to a desired reference state, then applies the same $s^{\mathrm{pre}} \in \mathbb{R}^{L \times c_s}$2 to other family members:

• Training: For a source protein $s^{\mathrm{pre}} \in \mathbb{R}^{L \times c_s}$3 and target structure $s^{\mathrm{pre}} \in \mathbb{R}^{L \times c_s}$4, $s^{\mathrm{pre}} \in \mathbb{R}^{L \times c_s}$5 is optimized to minimize $s^{\mathrm{pre}} \in \mathbb{R}^{L \times c_s}$6 over one deterministic denoise, with robustifying MSA resampling.
• Inference: For any homolog $s^{\mathrm{pre}} \in \mathbb{R}^{L \times c_s}$7, the pipeline applies $s^{\mathrm{pre}} \in \mathbb{R}^{L \times c_s}$8 to $s^{\mathrm{pre}} \in \mathbb{R}^{L \times c_s}$9 just before the Pairformer, followed by standard OF3p diffusion.
• Benchmarks: Tasks defined for GPCR activation (TM6 helix), kinase DFG-out/A-loop, and outward-open membrane transporters.

Results (success@5):

• GPCR: default OF3p 24.3%, ConforNets 79.1%
• Kinase: default 5.9%, ConforNets 22.8%
• Transporter: default 16.1%, ConforNets 56.7%

Success@100 reachability jumps from 37.3% → 86.0% (GPCR), 10% → 26.3% (kinase), and 33.3% → 73.3% (transporter) with ConforNets; template- or AFsample3-based induction offers only marginal benefits, highlighting the specific utility of $$z^{\mathrm{pre}} \in \mathbb{R}^{L \times L \times c_z}$$0 for learned conformational control.

6. Implementation Protocols and Integration

Integration into OF3p is realized via concise pipelined modifications:

Unsupervised Diversity Training

$$z^{\mathrm{pre}} \in \mathbb{R}^{L \times L \times c_z}$$8

Supervised Transfer Training and Inference

$$z^{\mathrm{pre}} \in \mathbb{R}^{L \times L \times c_z}$$9 Compute overhead is modest: diversity training is $$z^{\mathrm{pre}} \in \mathbb{R}^{L \times L \times c_z}$$1–$$z^{\mathrm{pre}} \in \mathbb{R}^{L \times L \times c_z}$$2 the default OF3p run for a 200-residue protein (approximately 40 seconds on an A100 GPU). $$z^{\mathrm{pre}} \in \mathbb{R}^{L \times L \times c_z}$$3 application at inference is negligible ($$z^{\mathrm{pre}} \in \mathbb{R}^{L \times L \times c_z}$$4 extra, due to a single matrix multiply and bias addition).

7. Practical Constraints and Projected Impact

ConforNets as implemented in OF3p display robustness to overfitting via aggressive MSA subsampling and maintain physical plausibility of predicted structures. Key practical notes include:

• Recycle Settings: $$z^{\mathrm{pre}} \in \mathbb{R}^{L \times L \times c_z}$$5 or $$z^{\mathrm{pre}} \in \mathbb{R}^{L \times L \times c_z}$$6 can be used for diversity, $$z^{\mathrm{pre}} \in \mathbb{R}^{L \times L \times c_z}$$7 is preferred for transfer to avoid bias to the initial state.
• Structural Robustness: No increase in steric clashes is observed relative to baseline OF3p outputs.
• Limitations: While large-scale loop motions (e.g., kinase activation loop flips) exhibit some improvement, fine-grained sidechain sampling is not explicitly targeted. Constraining loss functions with experimental biophysical data (SAXS, NMR) is suggested as a possible extension.
• Intended Utility: Single-step modulation for cryptic pocket exploration, “at-will” conformational induction to support docking and molecular design, improved initializations for molecular dynamics or ensemble-based study, and compatibility as a plugin for large-scale inference.

ConforNets furnish a lightweight and reusable "affine knob" for steering AF3-based generative protein models, enabling broad, practical conformational experimentation with minimal computational overhead and without retraining the underlying model (Lee et al., 20 Apr 2026).