|
|
|
In this report
October's results put more weight behind function-directed protein design and experimentally tested materials discovery, while autonomous science moved into facility-scale construction. Funding and model work also focused on reusable infrastructure, with evidence ranging from device operation and synthesis to computational transfer tests.
- IPD and Nvidia released RFD4-Proteina, an open-source protein-design model reporting five- to 10-fold gains in generation speed and design success across four computational tasks:
- Pinal linked natural-language functional descriptions to protein sequences and structures, with experimental validation across four tasks and a 1.7-fold titre increase for one engineered H-protein:
- An AI-guided perovskite interface reached 27.2% efficiency and retained more than 95% after 2,000 hours at 65°C, while a separate data-driven study synthesised 18 multilayer MXene phases:
|
|
|
|
Protein design
IPD and Nvidia release RFD4-ProteinaAI drug discovery & protein design
The Institute for Protein Design and Nvidia released RFD4-Proteina, reporting five- to 10-fold faster generation and five- to 10-fold higher computational design success across four protein-design tasks. The open model changes the practical scale of programmable candidate generation, but the reported gains are benchmark results rather than experimental validation.
- RFD4-Proteina combines molecular inputs, constraints and scoring objectives with search, covering protein binding, small-molecule binding, enzyme scaffolding and DNA binding.
- The model is available through Amazon Bio Discovery, Anthropic's Claude Science and Google Cloud, linking it to agent and scientific-computing workflows.
RFD4-Proteina extends generative protein design from candidate production toward programmable search over functional objectives.
|
|
|
Protein design
Pinal validates natural-language protein designAI drug discovery & protein design
Pinal, a 16-billion-parameter model trained on 1.7 billion synthetic protein-text pairs, generated proteins from natural-language functional descriptions and validated the intended function in four experimental tasks. One designed H-protein raised product titre 1.7-fold against the native E. coli protein in a multi-enzyme CO2-fixation pathway, although the result is a single engineered system rather than evidence of broad performance.
- The four tasks covered a fluorescent protein, a polyethylene terephthalate hydrolase, an alcohol dehydrogenase and a metabolic H-protein.
- Generated candidates combined predicted foldability, functional-description alignment and sequence diversity before experimental testing.
Pinal tests whether language-level functional intent can survive the transition from model output to measured protein behaviour.
|
|
|
Materials discovery
AI-guided SAM reaches durable perovskite performanceAI for materials discovery & chemistry
Researchers used graph-neural-network retrosynthesis and high-throughput screening to identify a self-assembled monolayer that delivered 27.2% perovskite-solar-cell efficiency and retained more than 95% of its initial efficiency after 2,000 hours at 65°C. The result is experimentally demonstrated device performance, with durability measured in this device configuration rather than a general guarantee across perovskite systems.
- The framework reduced more than 90,000 structures to 300 candidates using a composite score built from eight descriptors.
- The selected AI-1 molecule formed S-Pb coordination bonds at the buried interface, improving energetic alignment and suppressing non-radiative recombination.
The study connects AI-guided molecular selection to both efficiency and operational stability, rather than to a screening score alone.
|
|
|
Materials discovery
Researchers synthesise 18 newly identified MXene phasesAI for materials discovery & chemistry
A Science study combined machine learning, database mining and historical materials knowledge to identify and synthesise 18 multilayer M2CT2 phases, including 11 previously unexplored rare-earth MXenes. The work supplies experimental evidence for a broader materials family, while the reported properties remain a platform for subsequent characterisation rather than an application result.
- The phases were made using a self-propagating high-temperature synthesis method, providing a direct route to the predicted materials.
- The study expands the MXene family by 33 historically unrecognised candidates and reports diverse semiconducting and magnetic properties.
This work links data-driven discovery to synthesis and enlarges the experimentally accessible design space for MXene chemistry.
|
|
|
Autonomous science
PNNL begins construction of M2PC microbial labAutonomous laboratories
Pacific Northwest National Laboratory has begun construction of the 30,800-square-foot Microbial Molecular Phenotyping Capability, planned to open in 2029 with more than 100 connected automated instruments. It is a facility-scale implementation of integrated microbial experimentation, not yet an operating demonstration.
- Custom software is intended to coordinate microbial growth, sample preparation, experimental conditions and phenotyping, connecting observable traits to molecular mechanisms.
- The facility is planned to be available to the wider research community through the Environmental Molecular Sciences Laboratory's competitive proposal process at no cost.
M2PC moves integrated automation from the AMP2 prototype into a planned, continuously connected national-laboratory capability.
|
|
|
Autonomous science
DOE funds reusable infrastructure for autonomous scienceAutonomous laboratories
The Department of Energy selected four National Laboratory-led projects totalling $30 million to develop reusable robotics software, digital twins, benchmarks, robot skills and safety practices. The awards establish an implementation programme for autonomous scientific systems, but they are funding commitments rather than evidence that the proposed capabilities already work across facilities.
- MAESTRO, DART, TRACE and SPIRE address different stages of the autonomous-science lifecycle, including learning through experience, reliability testing and facility operations.
- The programme emphasises connections across laboratories and stress-testing, alongside the reusable technical components.
The DOE programme supplies a policy and funding layer for turning autonomous-science prototypes into shared operational capabilities.
|
|
|
Scientific models
EVENET transfers across collider analysesScientific foundation models & benchmarks
EVENET, pretrained on 500 million simulated collision events, outperformed established baselines in heavy-resonance and exotic-Higgs searches and rediscovered the ϒ meson in CMS Open Data. The evidence supports reusable representation learning for collider analysis, while its pretraining depends on a large simulated dataset and the result is not a new physics discovery.
- The event-level model represents variable-length collections of physics objects as a shared particle-cloud representation.
- Performance was especially strong in low-statistics regimes, and quantum-correlation observables remained stable under systematic variations.
EVENET tests whether expensive domain-specific pretraining can become a reusable backbone across downstream collider analyses.
|
|
|
Materials discovery
NSF funds a five-lab soft-materials networkAI for materials discovery & chemistry
The National Science Foundation awarded $20 million to the University of Chicago to build PoLARIS, a remote network linking automated systems across five UChicago and Argonne laboratories. The planned network targets 100 to 300 experiments per day in polymers, gels and flexible electronics, but those throughput figures describe the project design rather than demonstrated operation.
- PoLARIS will combine AI-guided planning, active learning, robotics and digital twins across the connected laboratories.
- The project is intended to make soft-materials experiments more reproducible and accessible across sites, rather than concentrating automation in one lab.
PoLARIS extends autonomous discovery from individual instruments to a multi-laboratory operating network for soft materials.
|
|
