Zefeng Zhu

SteerAF: Distogram-based Steering of AlphaFold2 toward Alternative Conformations 1. SteerAF is an inference-time optimization framework built on OpenFold/AlphaFold2 that steers predictions away from the default (training-biased) conformation and toward alternative states by exploiting multi-peak signals already present in AF2 distograms derived from deep MSAs. 2. Core idea: freeze model weights and optimize the input MSA feature (not the network). At each step, compute a distogram cross-entropy loss between the step-0 (default) structure’s distance bins and the current step’s predicted distogram, then perform gradient ascent on the MSA feature to amplify non-default (alternative) distogram peaks while maintaining foldability via a pLDDT term. 3. A key technical contribution is block gradient ascent (BGA): only a small fraction of MSA-feature “blocks” (defined per residue per MSA row) are updated each step using importance sampling based on gradient norms. This sparsity improves performance slightly and makes the steering signal more site-specific and interpretable. 4. Benchmarks span four widely used two-state datasets (domain motion, open–close oc23, transmembrane tp16, fold-switch). With comparable prediction budgets, SteerAF matches or exceeds most competing approaches on the majority of systems, but is weaker on fold-switch targets where conformations differ drastically and distogram guidance may be insufficient. 5. Efficiency and robustness: SteerAF typically reaches strong alternative-state models within ~10–20 runs, each producing a short trajectory of structures over optimization steps. Using a “near-best run ratio” metric (fraction of runs that contain a model within 97% of the best TM-score), SteerAF shows a low failure rate and often high near-best ratios, supporting practical use when exhaustive sampling is undesirable. 6. Reference-free state selection (no experimental alternative needed): the paper proposes PCA on internal distance matrices restricted to common ordered secondary-structure segments (derived from step-0), followed by HDBSCAN clustering. Across tested systems where both states are predicted, this unsupervised pipeline recovers correct alternative-conformation clusters in 85% (39/46) of cases. 7. Interpretability: because SteerAF optimizes MSA inputs, the resulting sparse MSA-feature modifications can be summarized into PSSM shifts and mapped to residues. In tested proteins, modified sites correlate with experimentally characterized functional/conformational residues, recovering them at ~50% precision with recall >50% (far above random expectation). 8. Orthogonal validation via “site restoration”: reverting SteerAF-modified MSA features at curated functional sites back to step-0 values reduces alternative-state TM-scores more than reverting similarly modified control sites, supporting that the steering signal concentrates on biologically meaningful positions rather than arbitrary perturbations. 9. Downstream MD workflow: SteerAF predictions can seed conformational landscape exploration. For the transporter MdfA, selecting diverse predicted structures as relay targets for targeted MD and then running short production MD expands sampled conformational space >3× and reveals a multi-basin free-energy surface, including basins not cleanly separable by clustering on static predictions alone. 10. Limitations and outlook: SteerAF struggles on fold-switch proteins and systems with 3+ discrete terminal states, likely because multiple alternative peaks are mixed in a single distogram and the loss cannot disentangle them. The authors suggest hybrid strategies and extension to all-atom AF3-like models, noting conceptual alignment with distogram/confidence-driven hallucination in diffusion-era predictors. 📜Paper: biorxiv.org/content/10.648… #AlphaFold2 #ProteinStructure #ConformationalDynamics #MSA #Distogram #ComputationalBiology #StructuralBioinformatics #MDsimulations #ProteinDynamics #bioRxiv

@rishabh16_ Reminds me of the old days—Fig. 17 from Hanson and Thakur vibes :-) (doi.org/10.1016/j.jmgm…)

The diff geom course had a final term project that required us to show what we've learned about Lie groups, dual quaternions, slerp, etc Vibe-coded a protein visualizer three.js demo that converts backbones to SE(3) frames AF2-style with some extra bells and whistles

@aidenosinetrip1 @sokrypton @MazAbulnaga Great work! Technically, BioZernike could serve as a baseline for comparison, which convert point clouds into voxel grids and thus no sentive to wiring, circular permutation, stoichiometry etc. Though the study already stands on its own and such comparison may not be necessary:)

Excited to share my recent work, CIRPIN 🐍 with @sokrypton and @MazAbulnaga on the discovery of thousands of naturally occurring circularly permuted protein domains

@jiarlu @GabriCorso @jeremyWohlwend @meteos_97 @BarzilayRegina @WengongJin @Mattmcpartlon1 @axel_s_feldmann It's indeed a pretty "old" idea, for instance, DMPFold2(doi.org/10.1073/pnas.2…) implement that.

@GabriCorso @jeremyWohlwend @meteos_97 @BarzilayRegina @WengongJin @Mattmcpartlon1 @axel_s_feldmann Decoding coordinates from distogram using MDS is an awesome idea and I was really impressed when I firstly read it

There is lots of chat these days about simple/fast folding models. Few people know that more than a year ago, @jeremyWohlwend and @meteos_97 built MiniFold, a super fast folding method that actually approaches ESMFold accuracy while being >10x faster. Unfortunately, they never did any PR, so they never got enough credit for it. Go check it out: github.com/jwohlwend/mini…

@mgostIH Seems highly related to 2021.acl-long.223 (Guiding Teacher Forcing with Seer Forcing for Neural Machine Translation).

This paper is pretty cool: The Belief State Transformer Very simple technique and fast to train, makes transformers (or other seq models) better at modelling state and can additionally condition on the end! I wonder what this is like for RL, we might condition on high end reward!

Improving diffusion-based protein backbone generation with global-geometry-aware latent encoding 1/ Introducing TopoDiff: a diffusion-based framework that enhances protein structure generation by incorporating global-geometry-aware latent encoding. This approach captures the compact latent representation of protein geometry for more accurate structure prediction. 2/ TopoDiff outperforms existing models by improving the coverage of fold modes in protein structure space, especially for mainly-beta proteins, a group underrepresented in prior design efforts. This is crucial for advancing de novo protein design. 3/ One of TopoDiff’s key innovations is its two-stage sampling method, which separates latent space sampling from structure generation. This enhances control and interpretability, allowing researchers to guide the generation of novel proteins with desirable topologies. 4/ A notable experimental validation: out of 21 in silico mainly-beta designs, 9 were expressed in E. coli, and 4 were successfully purified. One protein, B10, exhibited a novel structural feature confirmed by X-ray crystallography with an RMSD of 1.31 Å. 5/ TopoDiff introduces controllable generation through latent-space manipulation, offering unprecedented control over global and local structure features. This opens new possibilities for generating proteins with specific design goals. 6/ Beyond traditional metrics like designability and novelty, TopoDiff introduces a new coverage metric to quantify the extent to which generated structures span the natural fold space, addressing a key challenge of mode collapse in generative models. 7/ Experimental results show that TopoDiff successfully generates mainly-beta proteins without relying on predefined blueprints, highlighting its potential to explore uncharted territories in protein design. 💻Code: github.com/meneshail/Topo… 📜Paper: biorxiv.org/content/10.110…

@DdelAlamo A similar preprint: Protokens biorxiv.org/content/10.110…

"Learning the Language of Protein Structure" by Gaujac et al tokenizes protein structure using an autoencoder, whose encoder and decoder are repurposed from ProteinMPNN and AlphaFold respectively. arxiv.org/abs/2405.15840…

@DdelAlamo Foldseek alphabet's is a lossy compression of the 3d structure. It does not attempt to convert from the discrete sequence back to the atomic structure.

One thing I'm having trouble understanding is how foldseek, which uses a similar scheme for per-residue structural discretization, works as well as it does with an alphabet of only twenty (!) possibilities, where as here we see 2-3 orders of magnitude more

@BenjMurrell Cool! I have been working on a really similar thing, but not on protein design fortunately. :-) I am also writing the preprint these days. Glad to see how distant groups coming up the same things. lol

The preprint will be up at bioRxiv after screening, and we'll get the code and models up ASAP. We'll also be trying to see if these to do anything, experimentally. And this is our first excursion into this field, so please let us know if we've missed anything in the literature!

How does this all work? The core of the model looks (mostly) like a transformer, with some tweaks. We use Invariant Point Attention (from AlphaFold2), with two modifications: we added a causal mask to make the whole thing autoregressive, and we use Softmax1 (@EvMill).

@sokrypton PDBe's RESTful API (ebi.ac.uk/pdbe/api/doc) and RCSB's RESTful/graphql API(data.rcsb.org) can help.

Anyone know of a simple way to map "polymer entity identifiers" to auth_asym_id or label_asym_id, without the need to download the individual mmcif files? Looks like we have yet another way to label chains: 🤔

@sokrypton It seems we can regard the trained structure generative model as a compressed database and use the super-secondary structure alphabet (as condition, or keywords) to retrieve the most probable results.

Weekend project: Use GPT4 to generate a GUI to control conditional fold generation of RFdiffusion 😎 colab.research.google.com/github/sokrypt…

@MartinPacesa Looks like an autoregressive process.