|
|
|
In this report
September added evidence at several layers of AI-enabled science: statistical signals from clinical data, experimentally tested molecular designs, and laboratories that can close the loop between hypothesis and physical measurement. The boundary is moving, but validation remains uneven, especially for fully computational discovery systems.
- Stanford's 37,000-agent virtual biotech found trial-success signals across 50,000 studies and produced a lung-cancer design later validated independently.
- DreamFold reported roughly 30,000-fold faster protein-folding pathways, while AdaptiveFlow opened active-learning screening across 69 billion compounds.
- Closed-loop biology and optics laboratories demonstrated autonomous experimentation, alongside a RET inhibitor workflow that reached in vivo validation.
|
|
Drug discovery and protein design |
Stanford builds a 37,000-agent virtual biotechGeneral
Stanford Medicine's virtual biotech analysed 50,000 clinical trials and found biological features associated with progression and safety, then generated a lung-cancer therapy later built and validated independently by a pharmaceutical company.
- Trials with high cell-type-specificity and bimodality scores were 40% more likely to move from Phase 1 to Phase 2, 48% more likely to reach market, and had 32% fewer adverse events.
- The result spans cancer, neurological, cardiovascular, renal and pulmonary trials; the therapeutic design centred on B7-H3 expressed by tumour-adjacent fibroblasts.
What this means for you AI-for-science evaluation now has a stronger example where computationally extracted principles connect to an independently validated therapeutic design, rather than stopping at a ranking metric. |
|
DreamFold accelerates protein-folding pathwaysGeneral Yale's DreamFold uses a learned latent environment to model protein folding pathways, reporting a roughly 30,000-fold speedup over molecular dynamics for monomeric proteins up to 449 amino acids. - The framework combines variational autoencoders, a next-state predictor and an evolutionary-policy agent, with O(N) scaling against molecular dynamics' O(N log N).
- The authors compared outputs with established simulation benchmarks and experimental data, and identified intermediates containing cryptic binding sites.
What this means for you Protein-design workflows gain a potentially useful pathway model for exploring transient structures, but the evidence still supports acceleration of simulation, not replacement of experimental structural validation. |
|
AdaptiveFlow screens 69 billion compoundsGeneral Nature Biotechnology's AdaptiveFlow combines ultralarge-library access with active learning, making a 69-billion-compound screening space more tractable than exhaustive docking. - The platform uses an 18-dimensional property grid and more than 1,500 docking protocols, scaling to 5.6 million CPUs in AWS cloud tests.
- Demonstrations produced nanomolar inhibitors of FSP1 and PARP1, including co-crystal evidence for the FSP1 interaction.
What this means for you Virtual screening is shifting from library size alone toward adaptive allocation of compute, while the experimentally resolved hits remain the part that establishes practical value. |
|
XtalPi workflow validates a RET inhibitor in vivoGeneral XtalPi and Jinan University combined patent extraction, medicinal-chemistry rules, free-energy calculations and generative design to identify a RET inhibitor with both biochemical potency and an in vivo tumour response. - Across six patents covering four targets, the workflow used patent-derived SAR to focus its search and synthesised three initial compounds in the RET case study.
- Cpd-31 reached a RET IC50 of 0.48 nM and produced 95.88% tumour inhibition at 10 mg/kg in vivo.
What this means for you The strongest drug-discovery evidence in the roundup is not a single generative model, but a data-centric chain linking proprietary text, physics-based computation, synthesis and animal validation. |
|
MolecureAI adds ADMET gates to autonomous designGeneral MolecureAI reports an autonomous design architecture that carries structural memory and ADMET checks across campaigns, with computational convergence on known EGFR, CDK2 and BRD4 chemistry. - A cost-ordered cascade reserves full docking for a filtered survivor set, while reactive substructures are written into target-specific learned state.
- Advancement rates exceeded 79% across four LLM backends and five ablations, but every reported result remained computational.
What this means for you The architecture offers a clearer test of persistent safety constraints in agentic chemistry, while prospective biochemical validation is still the unresolved boundary between a capable search system and a discovery result. |
Autonomous laboratories and materials |
Chalmers AI scientist validates yeast discoveriesGeneral A Chalmers closed-loop laboratory combined language models, automated reasoning and robotics to generate questions, run yeast experiments, assess results and refine its working model with limited human intervention. - The system was seeded with the yeast genome, metabolism and prior studies, extending the robot-scientist line of work associated with Adam and Eve.
- The researchers position the system as an augmenting scientist: humans still set priorities, interpret broader meaning and provide ethical oversight.
What this means for you Autonomous-lab research is gaining evidence at the level of iterative biological knowledge generation, not just automated execution, although the scope of the discoveries remains tied to the seeded system and experimental domain. |
|
MIT robot assembles and tunes laser experimentsGeneral MIT's reconfigurable optics laboratory autonomously assembled standard components, aligned a tabletop laser cavity and recovered from physical disturbances with micron-scale tuning. - QR-encoded mirrors, lenses and custom housings let a robotic arm build different configurations rather than repeat one fixed protocol.
- The team is developing remote access so researchers can submit experimental protocols for execution by the laboratory.
What this means for you The autonomous-lab concept is extending from biological screening into configurable physical assembly, making reproducibility and remote protocol execution central engineering questions for AI research tools. |
|
Deep learning classifies crystal space groupsGeneral Northwestern researchers reported more than 95% accuracy for classifying crystal space groups from Kikuchi diffraction patterns by adapting simulated training data to experimental constraints. - Unsupervised domain adaptation addressed the scarcity of experimental material phases in the training set.
- Physics-aware relabelling predicted the crystallographic information that Kikuchi diffraction can realistically support, rather than an unavailable ground truth.
What this means for you For AI materials characterisation, the result reinforces that benchmark design must reflect measurement physics; a high score is useful here because the target itself was made experimentally attainable. |
|
DOE funds AI search for permanent magnetsGeneral DOE's ARPA-E awarded $2.88 million to a University of Houston-led team to combine AI prediction with physical development of permanent magnets that could reduce dependence on neodymium iron boride. - The project targets sustainable, domestically sourced alternatives to magnets reliant on foreign-controlled critical minerals.
- It sits within a $100 million DOE initiative supporting early-stage research for domestic magnet manufacturing and critical-mineral security.
What this means for you AI materials discovery is being tied more explicitly to synthesis capacity and supply-chain policy, but this award marks a funded research direction rather than a demonstrated replacement material. |
How Iris covered your topics this month Every topic you track was searched over the last month. | Topic | Sources found | In this briefing | | General | 38 | 9 stories |
New stories that we're watching. Keep or drop these on your dashboard. AI Drug DiscoveryAutonomous Scientific Laboratories |
|
