Abstract: The Pharma Proteomics Project is a precompetitive biopharmaceutical consortium characterizing the plasma proteomic profiles of 54,219 UK Biobank participants. Here we provide a detailed summary of this initiative, including technical and biological validations, insights into proteomic disease signatures, and prediction modelling for various demographic and health indicators. We present comprehensive protein quantitative trait locus (pQTL) mapping of 2,923 proteins that identifies 14,287 primary genetic associations, of which 81% are previously undescribed, alongside ancestry-specific pQTL mapping in non-European individuals. The study provides an updated characterization of the genetic architecture of the plasma proteome, contextualized with projected pQTL discovery rates as sample sizes and proteomic assay coverages increase over time. We offer extensive insights into trans pQTLs across multiple biological domains, highlight genetic influences on ligand-receptor interactions and pathway perturbations across a diverse collection of cytokines and complement networks, and illustrate long-range epistatic effects of ABO blood group and FUT2 secretor status on proteins with gastrointestinal tissue-enriched expression. We demonstrate the utility of these data for drug discovery by extending the genetic proxied effects of protein targets, such as PCSK9, on additional endpoints, and disentangle specific genes and proteins perturbed at loci associated with COVID-19 susceptibility. This public-private partnership provides the scientific community with an open-access proteomics resource of considerable breadth and depth to help to elucidate the biological mechanisms underlying proteo-genomic discoveries and accelerate the development of biomarkers, predictive models and therapeutics1.

Abstract: Abstract Novel functional materials enable fundamental breakthroughs across technological applications from clean energy to information processing 1–11 . From microchips to batteries and photovoltaics, discovery of inorganic crystals has been bottlenecked by expensive trial-and-error approaches. Concurrently, deep-learning models for language, vision and biology have showcased emergent predictive capabilities with increasing data and computation 12–14 . Here we show that graph networks trained at scale can reach unprecedented levels of generalization, improving the efficiency of materials discovery by an order of magnitude. Building on 48,000 stable crystals identified in continuing studies 15–17 , improved efficiency enables the discovery of 2.2 million structures below the current convex hull, many of which escaped previous human chemical intuition. Our work represents an order-of-magnitude expansion in stable materials known to humanity. Stable discoveries that are on the final convex hull will be made available to screen for technological applications, as we demonstrate for layered materials and solid-electrolyte candidates. Of the stable structures, 736 have already been independently experimentally realized. The scale and diversity of hundreds of millions of first-principles calculations also unlock modelling capabilities for downstream applications, leading in particular to highly accurate and robust learned interatomic potentials that can be used in condensed-phase molecular-dynamics simulations and high-fidelity zero-shot prediction of ionic conductivity.

Abstract: Abstract Suppressing errors is the central challenge for useful quantum computing 1 , requiring quantum error correction (QEC) 2–6 for large-scale processing. However, the overhead in the realization of error-corrected ‘logical’ qubits, in which information is encoded across many physical qubits for redundancy 2–4 , poses substantial challenges to large-scale logical quantum computing. Here we report the realization of a programmable quantum processor based on encoded logical qubits operating with up to 280 physical qubits. Using logical-level control and a zoned architecture in reconfigurable neutral-atom arrays 7 , our system combines high two-qubit gate fidelities 8 , arbitrary connectivity 7,9 , as well as fully programmable single-qubit rotations and mid-circuit readout 10–15 . Operating this logical processor with various types of encoding, we demonstrate improvement of a two-qubit logic gate by scaling surface-code 6 distance from d = 3 to d = 7, preparation of colour-code qubits with break-even fidelities 5 , fault-tolerant creation of logical Greenberger–Horne–Zeilinger (GHZ) states and feedforward entanglement teleportation, as well as operation of 40 colour-code qubits. Finally, using 3D [[8,3,2]] code blocks 16,17 , we realize computationally complex sampling circuits 18 with up to 48 logical qubits entangled with hypercube connectivity 19 with 228 logical two-qubit gates and 48 logical CCZ gates 20 . We find that this logical encoding substantially improves algorithmic performance with error detection, outperforming physical-qubit fidelities at both cross-entropy benchmarking and quantum simulations of fast scrambling 21,22 . These results herald the advent of early error-corrected quantum computation and chart a path towards large-scale logical processors.

Abstract: Low-grade inflammation is a hallmark of old age and a central driver of ageing-associated impairment and disease1. Multiple factors can contribute to ageing-associated inflammation2; however, the molecular pathways that transduce aberrant inflammatory signalling and their impact in natural ageing remain unclear. Here we show that the cGAS-STING signalling pathway, which mediates immune sensing of DNA3, is a critical driver of chronic inflammation and functional decline during ageing. Blockade of STING suppresses the inflammatory phenotypes of senescent human cells and tissues, attenuates ageing-related inflammation in multiple peripheral organs and the brain in mice, and leads to an improvement in tissue function. Focusing on the ageing brain, we reveal that activation of STING triggers reactive microglial transcriptional states, neurodegeneration and cognitive decline. Cytosolic DNA released from perturbed mitochondria elicits cGAS activity in old microglia, defining a mechanism by which cGAS-STING signalling is engaged in the ageing brain. Single-nucleus RNA-sequencing analysis of microglia and hippocampi of a cGAS gain-of-function mouse model demonstrates that engagement of cGAS in microglia is sufficient to direct ageing-associated transcriptional microglial states leading to bystander cell inflammation, neurotoxicity and impaired memory capacity. Our findings establish the cGAS-STING pathway as a driver of ageing-related inflammation in peripheral organs and the brain, and reveal blockade of cGAS-STING signalling as a potential strategy to halt neurodegenerative processes during old age.

Abstract: Perovskite solar cells with the formula FA1-xCsxPbI3, where FA is formamidinium, provide an attractive option for integrating high efficiency, durable stability and compatibility with scaled-up fabrication. Despite the incorporation of Cs cations, which could potentially enable a perfect perovskite lattice1,2, the compositional inhomogeneity caused by A-site cation segregation is likely to be detrimental to the photovoltaic performance of the solar cells3,4. Here we visualized the out-of-plane compositional inhomogeneity along the vertical direction across perovskite films and identified the underlying reasons for the inhomogeneity and its potential impact for devices. We devised a strategy using 1-(phenylsulfonyl)pyrrole to homogenize the distribution of cation composition in perovskite films. The resultant p-i-n devices yielded a certified steady-state photon-to-electron conversion efficiency of 25.2% and durable stability.

Abstract: The mammalian brain consists of millions to billions of cells that are organized into many cell types with specific spatial distribution patterns and structural and functional properties1-3. Here we report a comprehensive and high-resolution transcriptomic and spatial cell-type atlas for the whole adult mouse brain. The cell-type atlas was created by combining a single-cell RNA-sequencing (scRNA-seq) dataset of around 7 million cells profiled (approximately 4.0 million cells passing quality control), and a spatial transcriptomic dataset of approximately 4.3 million cells using multiplexed error-robust fluorescence in situ hybridization (MERFISH). The atlas is hierarchically organized into 4 nested levels of classification: 34 classes, 338 subclasses, 1,201 supertypes and 5,322 clusters. We present an online platform, Allen Brain Cell Atlas, to visualize the mouse whole-brain cell-type atlas along with the single-cell RNA-sequencing and MERFISH datasets. We systematically analysed the neuronal and non-neuronal cell types across the brain and identified a high degree of correspondence between transcriptomic identity and spatial specificity for each cell type. The results reveal unique features of cell-type organization in different brain regions-in particular, a dichotomy between the dorsal and ventral parts of the brain. The dorsal part contains relatively fewer yet highly divergent neuronal types, whereas the ventral part contains more numerous neuronal types that are more closely related to each other. Our study also uncovered extraordinary diversity and heterogeneity in neurotransmitter and neuropeptide expression and co-expression patterns in different cell types. Finally, we found that transcription factors are major determinants of cell-type classification and identified a combinatorial transcription factor code that defines cell types across all parts of the brain. The whole mouse brain transcriptomic and spatial cell-type atlas establishes a benchmark reference atlas and a foundational resource for integrative investigations of cellular and circuit function, development and evolution of the mammalian brain.

Abstract: This manuscript is important, well-written, and will be of interest to a wide audience. Methods and data quality are appropriate. Conclusions in line with results. Kudos to the authors for a straightforward, well-executed, and significant study. I mention three points on which authors might choose to expand. 1. Implications of foundation models for medicine Authors might consider adding a sentence or two expanding on the implications of foundation models for medicine. The term (while not the idea) is relatively recent and some readers might benefit from an expanded explanation of the concept’s implications. In particular, it’s not just that foundation models can substantially reduce resources required to build usable health care AI (to democratize it, as authors write). It’s that in so doing these models could help direct support away from many single use, costly, initially impressive, AI tech demos, that rarely become integrated into routine care and feed skepticism of health care AI’s benefits. 2. Potential weaknesses associated with study’s data sets Authors acknowledge limitations associated with relying on datasets drawn almost entirely from the UK. However, because historically this literature (broadly, health care AI) has been prone to exaggeration—enabled often by ignoring limits created by homogeneity of people represented in study datasets—authors might consider highlighting this limitation not only at the end of the paper but also in one of the many passages noting the RETFound’s generalizability. 3. What is a clinical task Claims about the utility of RETFound to perform “a broad range of clinical tasks” (line 172) are accurate in their immediate context. But authors might examine what they mean by clinical tasks. When most people read the phrase, they assume a clinician—a person—is doing something, a task. The clinical tasks to which authors refer in this manuscript are tasks that can be achieved by AI. The point here is not lament AI’s potential replacement of people, it is rather to point out that AI is capable at this stageeven informed by foundation models – of performing only a very limited set of clinical tasks. To perform a “broad range” of “diverse” clinical tasks, health care AI needs still to better integrate human-AI collaboration. Authors might consider revising phrases about clinical tasks to more accurately express AI’s limitations in this regard.

Abstract: Transformer-based large language models are making significant strides in various fields, such as natural language processing1-5, biology6,7, chemistry8-10 and computer programming11,12. Here, we show the development and capabilities of Coscientist, an artificial intelligence system driven by GPT-4 that autonomously designs, plans and performs complex experiments by incorporating large language models empowered by tools such as internet and documentation search, code execution and experimental automation. Coscientist showcases its potential for accelerating research across six diverse tasks, including the successful reaction optimization of palladium-catalysed cross-couplings, while exhibiting advanced capabilities for (semi-)autonomous experimental design and execution. Our findings demonstrate the versatility, efficacy and explainability of artificial intelligence systems like Coscientist in advancing research.

Abstract: The authors present the first system that can fly a drone racing track as fast as human drone racing champions. Their approach, Swift, consists of first performing reinforcement learning in a simulated track. Then, test runs are performed on a real drone in a real track, and a residual observation and dynamics model is learned that captures the remaining reality gap. Swift wins the majority of individual races and also produces the fastest lap time.

Abstract: To close the gap between the rates of computational screening and experimental realization of novel materials1,2, we introduce the A-Lab, an autonomous laboratory for the solid-state synthesis of inorganic powders. This platform uses computations, historical data from the literature, machine learning (ML) and active learning to plan and interpret the outcomes of experiments performed using robotics. Over 17 days of continuous operation, the A-Lab realized 41 novel compounds from a set of 58 targets including a variety of oxides and phosphates that were identified using large-scale ab initio phase-stability data from the Materials Project and Google DeepMind. Synthesis recipes were proposed by natural-language models trained on the literature and optimized using an active-learning approach grounded in thermodynamics. Analysis of the failed syntheses provides direct and actionable suggestions to improve current techniques for materials screening and synthesis design. The high success rate demonstrates the effectiveness of artificial-intelligence-driven platforms for autonomous materials discovery and motivates further integration of computations, historical knowledge and robotics.

Abstract: Systematic assessments of species extinction risk at regular intervals are necessary for informing conservation action1,2. Ongoing developments in taxonomy, threatening processes and research further underscore the need for reassessment3,4. Here we report the findings of the second Global Amphibian Assessment, evaluating 8,011 species for the International Union for Conservation of Nature Red List of Threatened Species. We find that amphibians are the most threatened vertebrate class (40.7% of species are globally threatened). The updated Red List Index shows that the status of amphibians is deteriorating globally, particularly for salamanders and in the Neotropics. Disease and habitat loss drove 91% of status deteriorations between 1980 and 2004. Ongoing and projected climate change effects are now of increasing concern, driving 39% of status deteriorations since 2004, followed by habitat loss (37%). Although signs of species recoveries incentivize immediate conservation action, scaled-up investment is urgently needed to reverse the current trends.

Abstract: The integer quantum anomalous Hall (QAH) effect is a lattice analogue of the quantum Hall effect at zero magnetic field1-3. This phenomenon occurs in systems with topologically non-trivial bands and spontaneous time-reversal symmetry breaking. Discovery of its fractional counterpart in the presence of strong electron correlations, that is, the fractional QAH effect4-7, would open a new chapter in condensed matter physics. Here we report the direct observation of both integer and fractional QAH effects in electrical measurements on twisted bilayer MoTe2. At zero magnetic field, near filling factor ν = -1 (one hole per moiré unit cell), we see an integer QAH plateau in the Hall resistance Rxy quantized to h/e2 ± 0.1%, whereas the longitudinal resistance Rxx vanishes. Remarkably, at ν = -2/3 and -3/5, we see plateau features in Rxy at [Formula: see text] and [Formula: see text], respectively, whereas Rxx remains small. All features shift linearly versus applied magnetic field with slopes matching the corresponding Chern numbers -1, -2/3 and -3/5, precisely as expected for integer and fractional QAH states. Additionally, at zero magnetic field, Rxy is approximately 2h/e2 near half-filling (ν = -1/2) and varies linearly as ν is tuned. This behaviour resembles that of the composite Fermi liquid in the half-filled lowest Landau level of a two-dimensional electron gas at high magnetic field8-14. Direct observation of the fractional QAH and associated effects enables research in charge fractionalization and anyonic statistics at zero magnetic field.

Abstract: Abstract Twenty-five years since foundational publications on valuing ecosystem services for human well-being 1,2 , addressing the global biodiversity crisis 3 still implies confronting barriers to incorporating nature’s diverse values into decision-making. These barriers include powerful interests supported by current norms and legal rules such as property rights, which determine whose values and which values of nature are acted on. A better understanding of how and why nature is (under)valued is more urgent than ever 4 . Notwithstanding agreements to incorporate nature’s values into actions, including the Kunming-Montreal Global Biodiversity Framework (GBF) 5 and the UN Sustainable Development Goals 6 , predominant environmental and development policies still prioritize a subset of values, particularly those linked to markets, and ignore other ways people relate to and benefit from nature 7 . Arguably, a ‘values crisis’ underpins the intertwined crises of biodiversity loss and climate change 8 , pandemic emergence 9 and socio-environmental injustices 10 . On the basis of more than 50,000 scientific publications, policy documents and Indigenous and local knowledge sources, the Intergovernmental Platform on Biodiversity and Ecosystem Services (IPBES) assessed knowledge on nature’s diverse values and valuation methods to gain insights into their role in policymaking and fuller integration into decisions 7,11 . Applying this evidence, combinations of values-centred approaches are proposed to improve valuation and address barriers to uptake, ultimately leveraging transformative changes towards more just (that is, fair treatment of people and nature, including inter- and intragenerational equity) and sustainable futures.

Abstract: Senescent cells drive age-related tissue dysfunction partially through the induction of a chronic senescence-associated secretory phenotype (SASP)1. Mitochondria are major regulators of the SASP; however, the underlying mechanisms have not been elucidated2. Mitochondria are often essential for apoptosis, a cell fate distinct from cellular senescence. During apoptosis, widespread mitochondrial outer membrane permeabilization (MOMP) commits a cell to die3. Here we find that MOMP occurring in a subset of mitochondria is a feature of cellular senescence. This process, called minority MOMP (miMOMP), requires BAX and BAK macropores enabling the release of mitochondrial DNA (mtDNA) into the cytosol. Cytosolic mtDNA in turn activates the cGAS-STING pathway, a major regulator of the SASP. We find that inhibition of MOMP in vivo decreases inflammatory markers and improves healthspan in aged mice. Our results reveal that apoptosis and senescence are regulated by similar mitochondria-dependent mechanisms and that sublethal mitochondrial apoptotic stress is a major driver of the SASP. We provide proof-of-concept that inhibition of miMOMP-induced inflammation may be a therapeutic route to improve healthspan.

Abstract: AlphaFold2 (ref. 1) has revolutionized structural biology by accurately predicting single structures of proteins. However, a protein's biological function often depends on multiple conformational substates2, and disease-causing point mutations often cause population changes within these substates3,4. We demonstrate that clustering a multiple-sequence alignment by sequence similarity enables AlphaFold2 to sample alternative states of known metamorphic proteins with high confidence. Using this method, named AF-Cluster, we investigated the evolutionary distribution of predicted structures for the metamorphic protein KaiB5 and found that predictions of both conformations were distributed in clusters across the KaiB family. We used nuclear magnetic resonance spectroscopy to confirm an AF-Cluster prediction: a cyanobacteria KaiB variant is stabilized in the opposite state compared with the more widely studied variant. To test AF-Cluster's sensitivity to point mutations, we designed and experimentally verified a set of three mutations predicted to flip KaiB from Rhodobacter sphaeroides from the ground to the fold-switched state. Finally, screening for alternative states in protein families without known fold switching identified a putative alternative state for the oxidoreductase Mpt53 in Mycobacterium tuberculosis. Further development of such bioinformatic methods in tandem with experiments will probably have a considerable impact on predicting protein energy landscapes, essential for illuminating biological function.

Abstract: Speech brain-computer interfaces (BCIs) have the potential to restore rapid communication to people with paralysis by decoding neural activity evoked by attempted speech into text1,2 or sound3,4. Early demonstrations, although promising, have not yet achieved accuracies sufficiently high for communication of unconstrained sentences from a large vocabulary1-7. Here we demonstrate a speech-to-text BCI that records spiking activity from intracortical microelectrode arrays. Enabled by these high-resolution recordings, our study participant-who can no longer speak intelligibly owing to amyotrophic lateral sclerosis-achieved a 9.1% word error rate on a 50-word vocabulary (2.7 times fewer errors than the previous state-of-the-art speech BCI2) and a 23.8% word error rate on a 125,000-word vocabulary (the first successful demonstration, to our knowledge, of large-vocabulary decoding). Our participant's attempted speech was decoded at 62 words per minute, which is 3.4 times as fast as the previous record8 and begins to approach the speed of natural conversation (160 words per minute9). Finally, we highlight two aspects of the neural code for speech that are encouraging for speech BCIs: spatially intermixed tuning to speech articulators that makes accurate decoding possible from only a small region of cortex, and a detailed articulatory representation of phonemes that persists years after paralysis. These results show a feasible path forward for restoring rapid communication to people with paralysis who can no longer speak.

Abstract: Abstract China’s goal to achieve carbon (C) neutrality by 2060 requires scaling up photovoltaic (PV) and wind power from 1 to 10–15 PWh year −1 (refs. 1–5 ). Following the historical rates of renewable installation 1 , a recent high-resolution energy-system model 6 and forecasts based on China’s 14th Five-year Energy Development (CFED) 7 , however, only indicate that the capacity will reach 5–9.5 PWh year −1 by 2060. Here we show that, by individually optimizing the deployment of 3,844 new utility-scale PV and wind power plants coordinated with ultra-high-voltage (UHV) transmission and energy storage and accounting for power-load flexibility and learning dynamics, the capacity of PV and wind power can be increased from 9 PWh year −1 (corresponding to the CFED path) to 15 PWh year −1 , accompanied by a reduction in the average abatement cost from US$97 to US$6 per tonne of carbon dioxide (tCO 2 ). To achieve this, annualized investment in PV and wind power should ramp up from US$77 billion in 2020 (current level) to US$127 billion in the 2020s and further to US$426 billion year −1 in the 2050s. The large-scale deployment of PV and wind power increases income for residents in the poorest regions as co-benefits. Our results highlight the importance of upgrading power systems by building energy storage, expanding transmission capacity and adjusting power load at the demand side to reduce the economic cost of deploying PV and wind power to achieve carbon neutrality in China.

Abstract: Abstract Animal studies show aging varies between individuals as well as between organs within an individual 1–4 , but whether this is true in humans and its effect on age-related diseases is unknown. We utilized levels of human blood plasma proteins originating from specific organs to measure organ-specific aging differences in living individuals. Using machine learning models, we analysed aging in 11 major organs and estimated organ age reproducibly in five independent cohorts encompassing 5,676 adults across the human lifespan. We discovered nearly 20% of the population show strongly accelerated age in one organ and 1.7% are multi-organ agers. Accelerated organ aging confers 20–50% higher mortality risk, and organ-specific diseases relate to faster aging of those organs. We find individuals with accelerated heart aging have a 250% increased heart failure risk and accelerated brain and vascular aging predict Alzheimer’s disease (AD) progression independently from and as strongly as plasma pTau-181 (ref. 5 ), the current best blood-based biomarker for AD. Our models link vascular calcification, extracellular matrix alterations and synaptic protein shedding to early cognitive decline. We introduce a simple and interpretable method to study organ aging using plasma proteomics data, predicting diseases and aging effects.

Abstract: Abstract Plywood is widely used in construction, such as for flooring and interior walls, as well as in the manufacture of household items such as furniture and cabinets. Such items are made of wood veneers that are bonded together with adhesives such as urea–formaldehyde and phenol–formaldehyde resins 1,2 . Researchers in academia and industry have long aimed to synthesize lignin–phenol–formaldehyde resin adhesives using biomass-derived lignin, a phenolic polymer that can be used to substitute the petroleum-derived phenol 3–6 . However, lignin–phenol–formaldehyde resin adhesives are less attractive to plywood manufacturers than urea–formaldehyde and phenol–formaldehyde resins owing to their appearance and cost. Here we report a simple and practical strategy for preparing lignin-based wood adhesives from lignocellulosic biomass. Our strategy involves separation of uncondensed or slightly condensed lignins from biomass followed by direct application of a suspension of the lignin and water as an adhesive on wood veneers. Plywood products with superior performances could be prepared with such lignin adhesives at a wide range of hot-pressing temperatures, enabling the use of these adhesives as promising alternatives to traditional wood adhesives in different market segments. Mechanistic studies indicate that the adhesion mechanism of such lignin adhesives may involve softening of lignin by water, filling of vessels with softened lignin and crosslinking of lignins in adhesives with those in the cell wall.

Abstract: In mammalian brains, millions to billions of cells form complex interaction networks to enable a wide range of functions. The enormous diversity and intricate organization of cells have impeded our understanding of the molecular and cellular basis of brain function. Recent advances in spatially resolved single-cell transcriptomics have enabled systematic mapping of the spatial organization of molecularly defined cell types in complex tissues1-3, including several brain regions (for example, refs. 1-11). However, a comprehensive cell atlas of the whole brain is still missing. Here we imaged a panel of more than 1,100 genes in approximately 10 million cells across the entire adult mouse brains using multiplexed error-robust fluorescence in situ hybridization12 and performed spatially resolved, single-cell expression profiling at the whole-transcriptome scale by integrating multiplexed error-robust fluorescence in situ hybridization and single-cell RNA sequencing data. Using this approach, we generated a comprehensive cell atlas of more than 5,000 transcriptionally distinct cell clusters, belonging to more than 300 major cell types, in the whole mouse brain with high molecular and spatial resolution. Registration of this atlas to the mouse brain common coordinate framework allowed systematic quantifications of the cell-type composition and organization in individual brain regions. We further identified spatial modules characterized by distinct cell-type compositions and spatial gradients featuring gradual changes of cells. Finally, this high-resolution spatial map of cells, each with a transcriptome-wide expression profile, allowed us to infer cell-type-specific interactions between hundreds of cell-type pairs and predict molecular (ligand-receptor) basis and functional implications of these cell-cell interactions. These results provide rich insights into the molecular and cellular architecture of the brain and a foundation for functional investigations of neural circuits and their dysfunction in health and disease.

Abstract: Owing to a long history of anthropogenic pressures, freshwater ecosystems are among the most vulnerable to biodiversity loss1. Mitigation measures, including wastewater treatment and hydromorphological restoration, have aimed to improve environmental quality and foster the recovery of freshwater biodiversity2. Here, using 1,816 time series of freshwater invertebrate communities collected across 22 European countries between 1968 and 2020, we quantified temporal trends in taxonomic and functional diversity and their responses to environmental pressures and gradients. We observed overall increases in taxon richness (0.73% per year), functional richness (2.4% per year) and abundance (1.17% per year). However, these increases primarily occurred before the 2010s, and have since plateaued. Freshwater communities downstream of dams, urban areas and cropland were less likely to experience recovery. Communities at sites with faster rates of warming had fewer gains in taxon richness, functional richness and abundance. Although biodiversity gains in the 1990s and 2000s probably reflect the effectiveness of water-quality improvements and restoration projects, the decelerating trajectory in the 2010s suggests that the current measures offer diminishing returns. Given new and persistent pressures on freshwater ecosystems, including emerging pollutants, climate change and the spread of invasive species, we call for additional mitigation to revive the recovery of freshwater biodiversity.

Abstract: Large Language Models (LLMs) have demonstrated tremendous capabilities in solving complex tasks, from quantitative reasoning to understanding natural language. However, LLMs sometimes suffer from confabulations (or hallucinations) which can result in them making plausible but incorrect statements [1,2]. This hinders the use of current large models in scientific discovery. Here we introduce FunSearch (short for searching in the function space), an evolutionary procedure based on pairing a pre-trained LLM with a systematic evaluator. We demonstrate the effectiveness of this approach to surpass the best known results in important problems, pushing the boundary of existing LLM-based approaches [3]. Applying FunSearch to a central problem in extremal combinatorics — the cap set problem — we discover new constructions of large cap sets going beyond the best known ones, both in finite dimensional and asymptotic cases. This represents the first discoveries made for established open problems using LLMs. We showcase the generality of FunSearch by applying it to an algorithmic problem, online bin packing, finding new heuristics that improve upon widely used baselines. In contrast to most computer search approaches, FunSearch searches for programs that describe how to solve a problem, rather than what the solution is. Beyond being an effective and scalable strategy, discovered programs tend to be more interpretable than raw solutions, enabling feedback loops between domain experts and FunSearch, and the deployment of such programs in real-world applications.