arxiv: 2604.18559 · v1 · submitted 2026-04-20 · 🧬 q-bio.BM · cs.LG

Recognition: unknown

ConforNets: Latents-Based Conformational Control in OpenFold3

Minji Lee , Colin Kalicki , Minkyu Jeon , Aymen Qabel , Alisia Fadini , Mohammed AlQuraishi

Authors on Pith no claims yet

Pith reviewed 2026-05-10 02:35 UTC · model grok-4.3

classification 🧬 q-bio.BM cs.LG

keywords ConforNetsAlphaFold3conformational ensemblesprotein structure predictionlatent space modulationmulti-state proteinspair latentsaffine transforms

0 comments

The pith

Channel-wise affine transforms on pre-Pairformer pair latents enable reusable conformational control in AlphaFold3 models

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that a targeted set of learned adjustments to internal representations can make AlphaFold3-style models reliably output multiple biologically relevant protein shapes rather than a single dominant conformation. This matters because many proteins switch between states to carry out their functions, yet standard predictors remain limited to one output even when alternate states are known to exist. By pinpointing the pre-Pairformer pair latents as the best location and using channel-wise affine transforms as the operation, the authors create a reusable module that works across proteins instead of requiring protein-specific tweaks. The resulting approach reaches state-of-the-art performance on benchmarks for generating alternate states without supervision and, when trained supervised on one protein, induces the same conformational shift in other members of its family. This replaces ad hoc inference-time perturbations with a general, learnable control mechanism.

Core claim

ConforNets consist of channel-wise affine transforms applied specifically to the pre-Pairformer pair latents inside the AF3 architecture. These transforms globally modulate the model's internal representations in a manner reusable across different proteins. The method attains state-of-the-art success rates for unsupervised generation of alternate conformational states on all existing multi-state benchmarks. When trained in a supervised setting on one source protein, the same transforms induce a conserved conformational change across an entire protein family while preserving overall structure prediction accuracy.

What carries the argument

Channel-wise affine transforms of the pre-Pairformer pair latents, which globally modulate AF3 representations to provide reusable conformational control across proteins.

If this is right

Unsupervised generation of alternate conformational states reaches state-of-the-art success rates on every existing multi-state benchmark.
Supervised training on one protein allows the learned conformational change to be induced across an entire protein family.
The transforms operate globally and remain reusable across different proteins rather than requiring per-protein retuning.
Core structure prediction accuracy is maintained while adding the ability to sample alternate states.
The mechanism replaces inefficient ad hoc perturbations with a consistent, learnable control layer.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same latent adjustment approach could support systematic generation of conformational ensembles for use in drug design or functional annotation.
If the transforms prove sufficiently general, they might be combined with mutation or condition inputs to predict how sequence changes shift conformational preferences.
Applying the method to larger or more flexible protein systems would test whether the pre-Pairformer location remains optimal beyond the current benchmarks.

Load-bearing premise

That channel-wise affine transforms applied to the pre-Pairformer pair latents supply a general mechanism for conformational control that generalizes across proteins without degrading core structure prediction accuracy or introducing systematic artifacts.

What would settle it

A clear failure of ConforNets to recover known alternate states on a new multi-state protein benchmark, or the appearance of systematic distortions in predicted structures on standard single-state test sets, would show the central claim is not correct.

Figures

Figures reproduced from arXiv: 2604.18559 by Alisia Fadini, Aymen Qabel, Colin Kalicki, Minji Lee, Minkyu Jeon, Mohammed AlQuraishi.

**Figure 1.** Figure 1: ConforNets induce conformations by mixing and biasing the channels of AF3’s pair latents to satisfy different objectives. 1. Diverse conformation prediction: k ConforNets are jointly trained to transform z pre such that the diversity of predicted conformations is maximized. Optimization is done via gradient descent using distogram- or coordinates-based objectives. 2. Supervised transfer of conformation: on… view at source ↗

**Figure 2.** Figure 2: State coverage of membrane transporters. (a) Success@100 rate of covering membrane transporter states (see text for definition) as a function of RMSD for ConforNets and other methods. Shaded regions denote bootstrap standard deviation. (b) Empirical energy landscape of ConforNet samples as a function of RMSD to inward-open (6FHZ;A) and outward-open (3VVS;A) states of MATE family transporter. (c) Best pred… view at source ↗

**Figure 4.** Figure 4: Transfer ConforNets shift state distribution. Empirical energy landscapes of two GPCRs under default OF3p sampling (left) and ConforNets (right). For Adenosine receptor A1 (a), OF3p never samples the active state, while for Cannabinoid receptor 1 (b), OF3p samples it with an arbitrary probability of 35%. ConforNets shift the sampling probabilities of the active states of both GPCRs to over 80% and rarely s… view at source ↗

**Figure 5.** Figure 5: Conformation transfer generalizes to structurally dissimilar proteins. ConforNets trained on source proteins with desired conformations (yellow) can be applied to target proteins whose default OF3p prediction is of a different state (green) and whose global structural similarity with respect to the source—in the desired state—is low (blue). Nonetheless, ConforNet-induced predictions (pink) are highly congr… view at source ↗

read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

ConforNets give a reusable channel-wise affine transform on pre-Pairformer pair latents that improves alternate-state generation in OpenFold3 and adds a supervised transfer task.

read the letter

The key takeaway is that this work distills an empirical search over perturbation locations and operations down to a simple, reusable mechanism: channel-wise affine transforms applied to the pair latents just before the Pairformer in the AF3 architecture. Unlike earlier ad-hoc inference tweaks, this produces a global modulation that can be applied across proteins without per-case retraining. On the unsupervised side it reports state-of-the-art success rates on existing multi-state benchmarks; on the supervised side it shows that a transform learned from one source protein can induce a corresponding shift across a family, which is a new task framing.

Referee Report

0 major / 2 minor

Summary. The paper introduces ConforNets, a mechanism consisting of channel-wise affine transforms applied to the pre-Pairformer pair latents within the OpenFold3 (AF3) architecture. Through empirical investigation of perturbation locations and operations, the authors distill this into a reusable global modulation approach. They claim state-of-the-art success rates for unsupervised generation of alternate conformational states on all existing multi-state benchmarks, and introduce a novel supervised conformational-transfer task where a model trained on one source protein induces conserved changes across a protein family.

Significance. If the empirical results and controls hold, this provides a general, reusable latents-based control mechanism that improves upon ad hoc inference-time perturbations for eliciting conformational variability in AF3-based models. The systematic study of optimal perturbation strategies, combined with quantitative benchmarks and the new transfer task, represents a concrete advance for structural biology applications requiring access to alternate states.

minor comments (2)

[§3.2] §3.2 (Methods): The precise parameterization of the channel-wise affine transforms (scale and shift per channel) is described in prose but would benefit from an explicit equation to ensure full reproducibility.
[Figure 3] Figure 3 and associated text: Include a direct comparison panel or table row showing core structure prediction accuracy (e.g., pLDDT or RMSD on the dominant state) before and after ConforNet application to quantify any degradation.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their supportive summary and recommendation of minor revision. We are pleased that the work on ConforNets is viewed as providing a reusable latents-based control mechanism for conformational variability in AF3-based models.

Circularity Check

0 steps flagged

No significant circularity; empirical investigation with external benchmarks

full rationale

The paper frames its contribution as an empirical search over perturbation locations and operations within the AF3/OpenFold3 architecture, distilling the result to channel-wise affine transforms on pre-Pairformer pair latents. Central claims (SOTA unsupervised alternate-state generation on existing multi-state benchmarks and successful supervised conformational transfer across protein families) are evaluated on external, held-out benchmarks rather than being derived from the model's own fitted parameters or self-referential equations. No load-bearing step reduces by construction to a fitted input renamed as a prediction, a self-citation chain, or an ansatz smuggled via prior work by the same authors. The derivation chain is therefore self-contained against external data and does not exhibit the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the empirical effectiveness of the proposed transforms. No new physical entities are postulated. The main assumption is that the chosen latent location encodes controllable conformational information.

axioms (1)

domain assumption Perturbing pre-Pairformer pair latents via channel-wise affine transforms can modulate conformational states in OpenFold3 without compromising overall structure prediction
This is the distilled finding presented as the core of ConforNets.

pith-pipeline@v0.9.0 · 5499 in / 1328 out tokens · 42790 ms · 2026-05-10T02:35:48.252352+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 11 canonical work pages · 4 internal anchors

[1]

Protein complex prediction with alphafold-multimer

Evans, R., O’Neill, M., Pritzel, A., Antropova, N., Senior, A., Green, T., Žídek, A., Bates, R., Blackwell, S., Yim, J., et al. Protein complex prediction with alphafold-multimer. biorxiv, pp. 2021–10,

2021
[2]

Alphafold meets flow matching for generating protein ensembles.arXiv preprint arXiv:2402.04845,

Jing, B., Berger, B., and Jaakkola, T. Alphafold meets flow matching for generating protein ensembles.arXiv preprint arXiv:2402.04845,

work page arXiv
[3]

and Wallner, B

Kalakoti, Y . and Wallner, B. Afsample3: Generating and selecting multiple conformational states with alphafold3. bioRxiv, pp. 2026–01,

2026
[4]

URL https://doi.org/10.1093/nar/gkaa895

doi: 10.1093/nar/gkaa895. URL https://doi.org/10.1093/nar/gkaa895. Karras, T., Aittala, M., Aila, T., and Laine, S. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35: 26565–26577,

work page doi:10.1093/nar/gkaa895
[5]

Kingma, D. P. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980,

work page internal anchor Pith review Pith/arXiv arXiv
[6]

J., Mordalski, S., Pándy-Szekeres, G., Esguerra, M., Mamyrbekov, A., Munk, C., Keser˝ u, G

Kooistra, A. J., Mordalski, S., Pándy-Szekeres, G., Esguerra, M., Mamyrbekov, A., Munk, C., Keser˝ u, G. M., and Glo- riam, D. E. Gpcrdb in 2021: integrating gpcr sequence, structure and function.Nucleic Acids Research, 49(D1): D335–D343,

2021
[7]

The Power of Scale for Parameter-Efficient Prompt Tuning

Lester, B., Al-Rfou, R., and Constant, N. The power of scale for parameter-efficient prompt tuning.arXiv preprint arXiv:2104.08691,

work page internal anchor Pith review arXiv
[8]

Robust inference- time steering of protein diffusion models via embedding optimization.arXiv preprint arXiv:2602.05285,

Li, M., Han, J., Cossio, P., and Wu, L. Robust inference- time steering of protein diffusion models via embedding optimization.arXiv preprint arXiv:2602.05285,

work page internal anchor Pith review arXiv
[9]

Li, X. L. and Liang, P. Prefix-tuning: Optimizing continuous prompts for generation.arXiv preprint arXiv:2101.00190,

work page internal anchor Pith review arXiv
[10]

Z., Shi, C., Guo, H., Bengio, Y ., and Tang, J

Lu, J., Chen, X., Lu, S. Z., Shi, C., Guo, H., Bengio, Y ., and Tang, J. Structure language models for protein con- formation generation.arXiv preprint arXiv:2410.18403,

work page arXiv
[11]

Maddipatla, A., Rzayev, A., Pegoraro, M., Pacesa, M., Schanda, P., Marx, A., Vedula, S., and Bronstein, A. M. Inference-time optimization for experiment- grounded protein ensemble generation.arXiv preprint arXiv:2602.24007,

work page arXiv
[12]

Meller, A., Bhakat, S., Solieva, S., and Bowman, G

doi: 10.1016/j.bpj.2015.08.015. Meller, A., Bhakat, S., Solieva, S., and Bowman, G. R. Ac- celerating cryptic pocket discovery using alphafold.Jour- nal of chemical theory and computation, 19(14):4355– 4363,

work page doi:10.1016/j.bpj.2015.08.015 2015
[13]

D., Karaguesian, J., Suomivuori, C.-M., and Dror, R

Richman, D. D., Karaguesian, J., Suomivuori, C.-M., and Dror, R. O. Unlocking hidden biomolecular conforma- tional landscapes in diffusion models at inference time. arXiv preprint arXiv:2512.03312,

work page arXiv
[14]

P., Ou, C., and Ovchinnikov, S

Roney, J. P., Ou, C., and Ovchinnikov, S. Protein diffusion models as statistical potentials.bioRxiv, pp. 2025–12,

2025
[15]

AlphaFold meets flow matching for generating protein ensembles

Wang, Y ., Wang, L., Shen, Y ., Wang, Y ., Yuan, H., Wu, Y ., and Gu, Q. Protein conformation generation via force-guided se (3) diffusion models.arXiv preprint arXiv:2403.14088,

work page arXiv
[16]

and Feng, L

Wu, D. and Feng, L. Robust prediction of multiple protein conformations with entropy guidance.bioRxiv, pp. 2025– 04,

2025
[17]

as our evaluation pipeline, with some minor modifications to input formatting, RMSD cutoffs, and success rate computation for different numbers of samples. RMSD is computed following their pipeline: generated structures and reference structures are loaded as trajectories, sequences are aligned, backbone atoms are selected, and RMSD is computed usingmdtraj...

2015
[18]

Conformational prediction accuracy was evaluated on the TM6 helix

Disconnected fragments shorter than 20 residues were discarded. Conformational prediction accuracy was evaluated on the TM6 helix. We do not score other canonical conformational changes, such as TM5 and TM7. D.2. Preprocessing DFG in-out pairs from KLIFS The kinase benchmark dataset was constructed from the KLIFS database (Kanev et al., 2021). We selected...

2021
[19]

Extended results G.1

26.1±4.5% 5.9±2.3% 25.6±5.1% 51.0 15.0 33.3 21 ConforNets: Latents-Based Conformational Control in OpenFold3 G. Extended results G.1. Optimal perturbation location This section provides the full results corresponding to Sec. 6.1. In the main text, we report the K= 1 setting as a representative comparison across perturbation locations. Here, we additionall...

2023