|
|
|
In this report
The month extended generative protein design beyond static folds into controllable dynamics, functional optimisation and new routes into experimental structure determination. The strongest results were experimentally validated, while the newer simulation, watermarking and therapeutic-development claims remain at earlier stages of validation.
- Guo et al. demonstrated de novo proteins with controllable conformational switching and calcium-regulated allostery,
- DreamFold reported folding-pathway generation up to roughly 30,000 times faster than molecular dynamics,
- BioMarin and Profluent began a collaboration to generate enzyme-replacement-therapy candidates with protein language models,
|
|
Programmable dynamics and faster generative simulation
Guo team designs proteins with controllable switchingDe novo protein design & generative models
Guo and colleagues designed de novo proteins that switch between user-specified conformations, with four NMR structures validating the designs and a calcium-binding switch demonstrating programmable allostery.
- The method uses AlphaFold2-guided sequence search, molecular dynamics and mutational scans to focus sampling on residues that control state preference.
- A single amino-acid change shifted one design from a strongly favoured state toward roughly equal populations; the switches exchanged on a low-microsecond timescale.
What this means for you De novo design is no longer limited to specifying a single stable structure in this result; experimentally validated control over state populations brings designed dynamics closer to the functional logic used in structural biology. |
|
DreamFold accelerates protein folding-pathway generationDe novo protein design & generative models
DreamFold uses a learned latent representation of protein dynamics to generate folding pathways up to roughly 30,000 times faster than molecular dynamics, but its trajectories remain model-generated rather than direct physical simulations.
- The framework combines variational autoencoders with a neural latent-space transition model and an evolutionary policy that drives structures toward target folds.
- For monomeric proteins up to 449 residues, the authors report linear computational scaling and identify folding intermediates containing possible cryptic binding sites against MD benchmarks and experimental data.
What this means for you Folding-landscape exploration can be screened at a different computational scale, while the biological interpretation still depends on how well the learned latent dynamics reproduce physical kinetics and rare states. |
|
Structure prediction gains from targeted inference controls
Researchers improve AlphaFold3 alternative-state samplingProtein structure prediction
A 107-protein benchmark found that perturbing AlphaFold3's multiple-sequence alignments improves recovery of alternative conformations, adding a measurable inference-time control rather than a new trained model.
- Stochastic subsampling, clustering and column masking each improved top-1% TM-score by at least 0.05 in about 20% of cases and rarely degraded performance.
- Unperturbed AF3 already outperformed AF2 with perturbations on the benchmark and performed comparably to BioEmu for alternative-state sampling.
What this means for you Dynamic-ensemble prediction is becoming more sensitive to inference settings, so AF3 output is less usefully treated as a single canonical structure when multiple experimentally observed states are relevant. |
|
HighFold4 closes cyclic-peptide chemistry gapsProtein structure prediction
HighFold4 adapts AlphaFold3 to predict closed cyclic-peptide conformations with user-specified chemical connectivity, addressing a topology failure that otherwise leaves many predicted macrocycles effectively linear.
- The framework enforces both ring closure and bond-specific connectivity without retraining the base model.
- Its evaluation covered 179 structures, with the stated application being macrocycle design where geometry and linkage chemistry affect affinity, selectivity and stability.
What this means for you For noncanonical peptide design, explicit chemical constraints can matter more than another round of generic model scaling; the relevant prediction target is the chemically specified topology, not just a plausible backbone. |
|
Experimental feedback, binders and structural access
RLXF steers protein language models toward functionProtein language models
Reinforcement Learning from eXperimental Feedback aligned protein language models to measured biochemical objectives across five protein families, with CreiLOV variants reaching the highest fluorescence reported for that protein.
- RLXF combines the evolutionary information in pretrained pLMs with experimental observations rather than relying on zero-shot generation alone.
- The reported gains include synergistic mutation combinations that were difficult to identify through the model's pretrained representation or conventional evolutionary search.
What this means for you pLM design is beginning to function as an experimentally steered optimisation loop, but the evidence here supports functional improvement in tested families rather than a general guarantee across protein objectives. |
|
Fold'EM infers structures from cryo-EM particles directlyCryo-EM & experimental structure methods
The Fold'EM preprint proposes inferring atomic structures and particle orientations directly from cryo-EM images, including distinct conformational states in heterogeneous samples, without first reconstructing an electrostatic potential map.
- The inference-time framework combines sequence-derived generative structural priors with particle images in both known-orientation and ab initio settings.
- Its stated target is the low-sample and low-population-state problem, where conventional reconstruction can require many particles or merge rare conformations into dominant ones.
What this means for you Generative priors may move cryo-EM analysis closer to particle-level heterogeneity, although the evidence remains a preprint claim rather than an established replacement for map-based validation. |
|
RFdiffusion creates a selective unanchored-ubiquitin binderEnzyme, binder & antibody engineering
RFdiffusion and affinity maturation produced UBiP10, a selective binder that enriched unanchored ubiquitin from cells and influenza A virions and exposed a previously uncharacterised ubiquitin coat around aggresomes.
- The work began with the HDAC6 zinc-finger domain, improved its affinity and then used RFdiffusion to design 11 new ubiquitin-binding proteins.
- UBiP10 enabled enrichment and imaging, including evidence that unanchored ubiquitin levels changed after proteasome inhibition.
What this means for you Designed binders are serving as experimental instruments as well as affinity reagents, making previously difficult protein states accessible to cell biology. |
|
Biosecurity and translation
DeepMind introduces SynthIDBio for designed proteinsDe novo protein design & generative models
DeepMind described SynthIDBio, a hidden statistical signature for AI-designed protein sequences and structures, with tests showing retained binding affinity but also showing that the mark can be removed by redesign.
- The signature is intended to be added by tools such as AlphaFold and RFdiffusion and detected by holders of a secret key, including potential DNA-synthesis screening partners.
- Stress tests covered proteins involved in viral infection, blood-vessel formation and immune regulation; redesigning a marked protein can preserve function while obscuring its origin.
What this means for you Provenance can become part of the design-tool stack, but SynthIDBio is a layer in biosecurity rather than a tamper-proof record of synthetic origin. |
|
BioMarin and Profluent target enzyme-replacement therapiesTools, datasets & companies
BioMarin and Profluent announced a research collaboration to generate enzyme-replacement-therapy candidates with protein language models, giving pLM design a concrete path into therapeutic screening while leaving advancement prospective.
- Profluent will generate candidates with its proprietary pLM suite; BioMarin will screen and evaluate them for possible preclinical development.
- The partnership pairs pLM-based sequence generation with BioMarin's existing enzyme-replacement-therapy expertise, including Vimizim and Naglazyme.
What this means for you The field is moving from model demonstrations toward company-specific therapeutic candidate generation, but the relevant validation point is still screening and preclinical advancement, not the announcement itself. |
|
