Zhaolian Lu, Shanlan Mo, Duo Xie, Xiangwei Zhai, Shanjun Deng, Kantian Zhou, Kun Wang, Xueling Kang, Hao Zhang, Juanzhen Tong, Liangzhen Hou, Huijuan Hu, Xuefei Li, Da Zhou, Leo Tsz On Lee, Li Liu, Yaxi Zhu, Jing Yu, Ping Lan, Jiguang Wang, Zhen He, Xionglei He, Zheng Hu <<<

AbstractOlfaction is a fundamental sensory modality that guides animal and human behaviour1,2. However, the underlying neural processes of human olfaction are still poorly understood at the fundamental—that is, the single-neuron—level. Here we report recordings of single-neuron activity in the piriform cortex and medial temporal lobe in awake humans performing an odour rating and identification task. We identified odour-modulated neurons within the piriform cortex, amygdala, entorhinal cortex and hippocampus. In each of these regions, neuronal firing accurately encodes odour identity. Notably, repeated odour presentations reduce response firing rates, demonstrating central repetition suppression and habituation. Different medial temporal lobe regions have distinct roles in odour processing, with amygdala neurons encoding subjective odour valence, and hippocampal neurons predicting behavioural odour identification performance. Whereas piriform neurons preferably encode chemical odour identity, hippocampal activity reflects subjective odour perception. Critically, we identify that piriform cortex neurons reliably encode odour-related images, supporting a multimodal role of the human piriform cortex. We also observe marked cross-modal coding of both odours and images, especially in the amygdala and piriform cortex. Moreover, we identify neurons that respond to semantically coherent odour and image information, demonstrating conceptual coding schemes in olfaction. Our results bridge the long-standing gap between animal models and non-invasive human studies and advance our understanding of odour processing in the human brain by identifying neuronal odour-coding principles, regional functional differences and cross-modal integration. <<<

Anna Goldie, Azalia Mirhoseini, Mustafa Yazgan, Joe Wenjie Jiang, Ebrahim Songhori, Shen Wang, Young-Joon Lee, Eric Johnson, Omkar Pathak, Azade Nova, Jiwoo Pak, Andy Tong, Kavya Srinivasa, William Hang, Emre Tuncer, Quoc V. Le, James Laudon, Richard Ho, Roger Carpenter, Jeff Dean <<<

AbstractExtracting the underlying temporal structure of experience is a fundamental aspect of learning and memory that allows us to predict what is likely to happen next. Current knowledge about the neural underpinnings of this cognitive process in humans stems from functional neuroimaging research1–5. As these methods lack direct access to the neuronal level, it remains unknown how this process is computed by neurons in the human brain. Here we record from single neurons in individuals who have been implanted with intracranial electrodes for clinical reasons, and show that human hippocampal and entorhinal neurons gradually modify their activity to encode the temporal structure of a complex image presentation sequence. This representation was formed rapidly, without providing specific instructions to the participants, and persisted when the prescribed experience was no longer present. Furthermore, the structure recovered from the population activity of hippocampal–entorhinal neurons closely resembled the structural graph defining the sequence, but at the same time, also reflected the probability of upcoming stimuli. Finally, learning of the sequence graph was related to spontaneous, time-compressed replay of individual neurons’ activity corresponding to previously experienced graph trajectories. These findings demonstrate that neurons in the hippocampus and entorhinal cortex integrate the ‘what’ and ‘when’ information to extract durable and predictive representations of the temporal structure of human experience. <<<

Rachael B. Chanin, Patrick T. West, Jakob Wirbel, Matthew O. Gill, Gabriella Z. M. Green, Ryan M. Park, Nora Enright, Arjun M. Miklos, Angela S. Hickey, Erin F. Brooks, Krystal K. Lum, Ileana M. Cristea, Ami S. Bhatt <<<

Andrew W. Senior, Richard Evans, John Jumper, James Kirkpatrick, Laurent Sifre, Tim Green, Chongli Qin, Augustin Žídek, Alexander W. R. Nelson, Alex Bridgland, Hugo Penedones, Stig Petersen, Karen Simonyan, Steve Crossan, Pushmeet Kohli, David T. Jones, David Silver, Koray Kavukcuoglu, Demis Hassabis <<<

An essential prerequisite for evolution by natural selection is variation among individuals in traits that affect fitness. The ability of a system to produce selectable variation, known as evolvability, thus markedly affects the rate of evolution. Although the immune system is among the fastest-evolving components in mammals, the sources of variation in immune traits remain largely unknown. Here we show that an important determinant of the immune system's evolvability is its organization into interacting modules represented by different immune cell types. By profiling immune cell variation in bone marrow of 54 genetically diverse mouse strains from the Collaborative Cross, we found that variation in immune cell frequencies is polygenic and that many associated genes are involved in homeostatic balance through cell-intrinsic functions of proliferation, migration and cell death. However, we also found genes associated with the frequency of a particular cell type that are expressed in a different cell type, exerting their effect in what we term cyto-trans. The vertebrate evolutionary record shows that genes associated in cyto-trans have faced weaker negative selection, thus increasing the robustness and hence evolvability of the immune system. This phenomenon is similarly observable in human blood. Our findings suggest that interactions between different components of the immune system provide a phenotypic space in which mutations can produce variation with little detriment, underscoring the role of modularity in the evolution of complex systems. <<<

Michael S Haney, Róbert Pálovics, Christy Nicole Munson, Chris Long, Patrik K Johansson, Oscar Yip, Wentao Dong, Eshaan Rawat, Elizabeth West, Johannes C M Schlachetzki, Andy Tsai, Ian Hunter Guldner, Bhawika S Lamichhane, Amanda Smith, Nicholas Schaum, Kruti Calcuttawala, Andrew Shin, Yung-Hua Wang, Chengzhong Wang, Nicole Koutsodendris, Geidy E Serrano, Thomas G Beach, Eric M Reiman, Christopher K Glass, Monther Abu-Remaileh, Annika Enejder, Yadong Huang, Tony Wyss-Coray <<<

Several genetic risk factors for Alzheimer's disease implicate genes involved in lipid metabolism and many of these lipid genes are highly expressed in glial cells. However, the relationship between lipid metabolism in glia and Alzheimer's disease pathology remains poorly understood. Through single-nucleus RNA sequencing of brain tissue in Alzheimer's disease, we have identified a microglial state defined by the expression of the lipid droplet-associated enzyme ACSL1 with ACSL1-positive microglia being most abundant in patients with Alzheimer's disease having the APOE4/4 genotype. In human induced pluripotent stem cell-derived microglia, fibrillar Aβ induces ACSL1 expression, triglyceride synthesis and lipid droplet accumulation in an APOE-dependent manner. Additionally, conditioned media from lipid droplet-containing microglia lead to Tau phosphorylation and neurotoxicity in an APOE-dependent manner. Our findings suggest a link between genetic risk factors for Alzheimer's disease with microglial lipid droplet accumulation and neurotoxic microglia-derived factors, potentially providing therapeutic strategies for Alzheimer's disease. <<<

AbstractStable diffusion revolutionized image creation from descriptive text. GPT-2 (ref. 1), GPT-3(.5) (ref. 2) and GPT-4 (ref. 3) demonstrated high performance across a variety of language tasks. ChatGPT introduced such language models to the public. It is now clear that generative artificial intelligence (AI) such as large language models (LLMs) is here to stay and will substantially change the ecosystem of online text and images. Here we consider what may happen to GPT-{n} once LLMs contribute much of the text found online. We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear. We refer to this effect as ‘model collapse’ and show that it can occur in LLMs as well as in variational autoencoders (VAEs) and Gaussian mixture models (GMMs). We build theoretical intuition behind the phenomenon and portray its ubiquity among all learned generative models. We demonstrate that it must be taken seriously if we are to sustain the benefits of training from large-scale data scraped from the web. Indeed, the value of data collected about genuine human interactions with systems will be increasingly valuable in the presence of LLM-generated content in data crawled from the Internet. <<<

Kateryna D Makova, Brandon D Pickett, Robert S Harris, Gabrielle A Hartley, Monika Cechova, Karol Pal, Sergey Nurk, DongAhn Yoo, Qiuhui Li, Prajna Hebbar, Barbara C McGrath, Francesca Antonacci, Margaux Aubel, Arjun Biddanda, Matthew Borchers, Erich Bornberg-Bauer, Gerard G Bouffard, Shelise Y Brooks, Lucia Carbone, Laura Carrel, Andrew Carroll, Pi-Chuan Chang, Chen-Shan Chin, Daniel E Cook, Sarah J C Craig, Luciana de Gennaro, Mark Diekhans, Amalia Dutra, Gage H Garcia, Patrick G S Grady, Richard E Green, Diana Haddad, Pille Hallast, William T Harvey, Glenn Hickey, David A Hillis, Savannah J Hoyt, Hyeonsoo Jeong, Kaivan Kamali, Sergei L Kosakovsky Pond, Troy M LaPolice, Charles Lee, Alexandra P Lewis, Yong-Hwee E Loh, Patrick Masterson, Kelly M McGarvey, Rajiv C McCoy, Paul Medvedev, Karen H Miga, Katherine M Munson, Evgenia Pak, Benedict Paten, Brendan J Pinto, Tamara Potapova, Arang Rhie, Joana L Rocha, Fedor Ryabov, Oliver A Ryder, Samuel Sacco, Kishwar Shafin, Valery A Shepelev, Viviane Slon, Steven J Solar, Jessica M Storer, Peter H Sudmant, Sweetalana, Alex Sweeten, Michael G Tassia, Françoise Thibaud-Nissen, Mario Ventura, Melissa A Wilson, Alice C Young, Huiqing Zeng, Xinru Zhang, Zachary A Szpiech, Christian D Huber, Jennifer L Gerton, Soojin V Yi, Michael C Schatz, Ivan A Alexandrov, Sergey Koren, Rachel J O'Neill, Evan E Eichler, Adam M Phillippy <<<

Apes possess two sex chromosomes-the male-specific Y chromosome and the X chromosome, which is present in both males and females. The Y chromosome is crucial for male reproduction, with deletions being linked to infertility. The X chromosome is vital for reproduction and cognition. Variation in mating patterns and brain function among apes suggests corresponding differences in their sex chromosomes. However, owing to their repetitive nature and incomplete reference assemblies, ape sex chromosomes have been challenging to study. Here, using the methodology developed for the telomere-to-telomere (T2T) human genome, we produced gapless assemblies of the X and Y chromosomes for five great apes (bonobo (Pan paniscus), chimpanzee (Pan troglodytes), western lowland gorilla (Gorilla gorilla gorilla), Bornean orangutan (Pongo pygmaeus) and Sumatran orangutan (Pongo abelii)) and a lesser ape (the siamang gibbon (Symphalangus syndactylus)), and untangled the intricacies of their evolution. Compared with the X chromosomes, the ape Y chromosomes vary greatly in size and have low alignability and high levels of structural rearrangements-owing to the accumulation of lineage-specific ampliconic regions, palindromes, transposable elements and satellites. Many Y chromosome genes expand in multi-copy families and some evolve under purifying selection. Thus, the Y chromosome exhibits dynamic evolution, whereas the X chromosome is more stable. Mapping short-read sequencing data to these assemblies revealed diversity and selection patterns on sex chromosomes of more than 100 individual great apes. These reference assemblies are expected to inform human evolution and conservation genetics of non-human apes, all of which are endangered species. <<<

Sankaraleengam Alagapan, Ki Sueng Choi, Stephen Heisig, Patricio Riva-Posse, Andrea Crowell, Vineet Tiruvadi, Mosadoluwa Obatusin, Ashan Veerakumar, Allison C Waters, Robert E Gross, Sinead Quinn, Lydia Denison, Matthew O'Shaughnessy, Marissa Connor, Gregory Canal, Jungho Cha, Rachel Hershenberg, Tanya Nauvel, Faical Isbaine, Muhammad Furqan Afzal, Martijn Figee, Brian H Kopell, Robert Butera, Helen S Mayberg, Christopher J Rozell <<<

Deep brain stimulation (DBS) of the subcallosal cingulate (SCC) can provide long-term symptom relief for treatment-resistant depression (TRD). However, achieving stable recovery is unpredictable, typically requiring trial-and-error stimulation adjustments due to individual recovery trajectories and subjective symptom reporting. We currently lack objective brain-based biomarkers to guide clinical decisions by distinguishing natural transient mood fluctuations from situations requiring intervention. To address this gap, we used a new device enabling electrophysiology recording to deliver SCC DBS to ten TRD participants (ClinicalTrials.gov identifier NCT01984710). At the study endpoint of 24 weeks, 90% of participants demonstrated robust clinical response, and 70% achieved remission. Using SCC local field potentials available from six participants, we deployed an explainable artificial intelligence approach to identify SCC local field potential changes indicating the patient's current clinical state. This biomarker is distinct from transient stimulation effects, sensitive to therapeutic adjustments and accurate at capturing individual recovery states. Variable recovery trajectories are predicted by the degree of preoperative damage to the structural integrity and functional connectivity within the targeted white matter treatment network, and are matched by objective facial expression changes detected using data-driven video analysis. Our results demonstrate the utility of objective biomarkers in the management of personalized SCC DBS and provide new insight into the relationship between multifaceted (functional, anatomical and behavioural) features of TRD pathology, motivating further research into causes of variability in depression treatment. <<<

Jinge Tian, Chenglong Wang, Fengyi Chen, Wenchao Qin, Hong Yang, Sihang Zhao, Jinliang Xia, Xian Du, Yifan Zhu, Lishuan Wu, Yan Cao, Hong Li, Junhong Zhuang, Shaojiang Chen, Huayuan Zhang, Qiuyue Chen, Mingcai Zhang, Xing Wang Deng, Dezhi Deng, Jigang Li, Feng Tian <<<

Increasing planting density is a key strategy to enhance maize yields. An ideotype for dense planting requires a 'smart canopy' with leaf angles at different canopy layers differentially optimized to maximize light interception and photosynthesis, amongst other features. Here, we identified leaf angle architecture of smart canopy 1 (lac1), a natural mutant possessing upright upper leaves, less erect middle leaves and relatively flat lower leaves. lac1 has improved photosynthetic capacity and weakened shade-avoidance responses under dense planting. lac1 encodes a brassinosteroid C-22 hydroxylase that predominantly regulates upper leaf angle. Phytochrome A photoreceptors accumulate in shade and interact with the transcription factor RAVL1 to promote its degradation via the 26S proteasome, thereby attenuating RAVL1 activation of lac1 and reducing brassinosteroid levels. This ultimately decreases upper leaf angle in dense fields. Large-scale field trials demonstrate lac1 boosts maize yields under high densities. To quickly introduce lac1 into breeding germplasm, we transformed a haploid inducer and recovered homozygous lac1 edits from 20 diverse inbred lines. The tested doubled haploids uniformly acquired smart-canopy-like plant architecture. We provide an important target and an accelerated strategy for developing high-density-tolerant cultivars, with lac1 serving as a genetic chassis for further engineering of a smart canopy in maize. <<<

Kathie Y Sun, Xiaodong Bai, Siying Chen, Suying Bao, Chuanyi Zhang, Manav Kapoor, Joshua Backman, Tyler Joseph, Evan Maxwell, George Mitra, Alexander Gorovits, Adam Mansfield, Boris Boutkov, Sujit Gokhale, Lukas Habegger, Anthony Marcketta, Adam E Locke, Liron Ganel, Alicia Hawes, Michael D Kessler, Deepika Sharma, Jeffrey Staples, Jonas Bovijn, Sahar Gelfman, Alessandro Di Gioia, Veera M Rajagopal, Alexander Lopez, Jennifer Rico Varela, Jesús Alegre-Díaz, Jaime Berumen, Roberto Tapia-Conyer, Pablo Kuri-Morales, Jason Torres, Jonathan Emberson, Rory Collins, Regeneron Genetics Center, RGC-ME Cohort Partners, Michael Cantor, Timothy Thornton, Hyun Min Kang, John D Overton, Alan R Shuldiner, M Laura Cremona, Mona Nafde, Aris Baras, Gonçalo Abecasis, Jonathan Marchini, Jeffrey G Reid, William Salerno, Suganthi Balasubramanian <<<

Rare coding variants that substantially affect function provide insights into the biology of a gene. However, ascertaining the frequency of such variants requires large sample sizes. Here we present a catalogue of human protein-coding variation, derived from exome sequencing of 983,578 individuals across diverse populations. In total, 23% of the Regeneron Genetics Center Million Exome (RGC-ME) data come from individuals of African, East Asian, Indigenous American, Middle Eastern and South Asian ancestry. The catalogue includes more than 10.4 million missense and 1.1 million predicted loss-of-function (pLOF) variants. We identify individuals with rare biallelic pLOF variants in 4,848 genes, 1,751 of which have not been previously reported. From precise quantitative estimates of selection against heterozygous loss of function (LOF), we identify 3,988 LOF-intolerant genes, including 86 that were previously assessed as tolerant and 1,153 that lack established disease annotation. We also define regions of missense depletion at high resolution. Notably, 1,482 genes have regions that are depleted of missense variants despite being tolerant of pLOF variants. Finally, we estimate that 3% of individuals have a clinically actionable genetic variant, and that 11,773 variants reported in ClinVar with unknown significance are likely to be deleterious cryptic splice sites. To facilitate variant interpretation and genetics-informed precision medicine, we make this resource of coding variation from the RGC-ME dataset publicly accessible through a variant allele frequency browser. <<<

Christina V Theodoris, Ling Xiao, Anant Chopra, Mark D Chaffin, Zeina R Al Sayed, Matthew C Hill, Helene Mantineo, Elizabeth M Brydon, Zexian Zeng, X Shirley Liu, Patrick T Ellinor <<<

Mapping gene networks requires large amounts of transcriptomic data to learn the connections between genes, which impedes discoveries in settings with limited data, including rare diseases and diseases affecting clinically inaccessible tissues. Recently, transfer learning has revolutionized fields such as natural language understanding and computer vision by leveraging deep learning models pretrained on large-scale general datasets that can then be fine-tuned towards a vast array of downstream tasks with limited task-specific data. Here, we developed a context-aware, attention-based deep learning model, Geneformer, pretrained on a large-scale corpus of about 30 million single-cell transcriptomes to enable context-specific predictions in settings with limited data in network biology. During pretraining, Geneformer gained a fundamental understanding of network dynamics, encoding network hierarchy in the attention weights of the model in a completely self-supervised manner. Fine-tuning towards a diverse panel of downstream tasks relevant to chromatin and network dynamics using limited task-specific data demonstrated that Geneformer consistently boosted predictive accuracy. Applied to disease modelling with limited patient data, Geneformer identified candidate therapeutic targets for cardiomyopathy. Overall, Geneformer represents a pretrained deep learning model from which fine-tuning towards a broad range of downstream applications can be pursued to accelerate discovery of key network regulators and candidate therapeutic targets. <<<

Mathias Munschauer, Celina T Nguyen, Klara Sirokman, Christina R Hartigan, Larson Hogstrom, Jesse M Engreitz, Jacob C Ulirsch, Charles P Fulco, Vidya Subramanian, Jenny Chen, Monica Schenone, Mitchell Guttman, Steven A Carr, Eric S Lander <<<

The human genome contains thousands of long non-coding RNAs, but specific biological functions and biochemical mechanisms have been discovered for only about a dozen. A specific long non-coding RNA-non-coding RNA activated by DNA damage (NORAD)-has recently been shown to be required for maintaining genomic stability, but its molecular mechanism is unknown. Here we combine RNA antisense purification and quantitative mass spectrometry to identify proteins that directly interact with NORAD in living cells. We show that NORAD interacts with proteins involved in DNA replication and repair in steady-state cells and localizes to the nucleus upon stimulation with replication stress or DNA damage. In particular, NORAD interacts with RBMX, a component of the DNA-damage response, and contains the strongest RBMX-binding site in the transcriptome. We demonstrate that NORAD controls the ability of RBMX to assemble a ribonucleoprotein complex-which we term NORAD-activated ribonucleoprotein complex 1 (NARC1)-that contains the known suppressors of genomic instability topoisomerase I (TOP1), ALYREF and the PRPF19-CDC5L complex. Cells depleted for NORAD or RBMX display an increased frequency of chromosome segregation defects, reduced replication-fork velocity and altered cell-cycle progression-which represent phenotypes that are mechanistically linked to TOP1 and PRPF19-CDC5L function. Expression of NORAD in trans can rescue defects caused by NORAD depletion, but rescue is significantly impaired when the RBMX-binding site in NORAD is deleted. Our results demonstrate that the interaction between NORAD and RBMX is important for NORAD function, and that NORAD is required for the assembly of the previously unknown topoisomerase complex NARC1, which contributes to maintaining genomic stability. In addition, we uncover a previously unknown function for long non-coding RNAs in modulating the ability of an RNA-binding protein to assemble a higher-order ribonucleoprotein complex. <<<

Proving mathematical theorems at the olympiad level represents a notable milestone in human-level automated reasoning, owing to their reputed difficulty among the world's best talents in pre-university mathematics. Current machine-learning approaches, however, are not applicable to most mathematical domains owing to the high cost of translating human proofs into machine-verifiable format. The problem is even worse for geometry because of its unique translation challenges, resulting in severe scarcity of training data. We propose AlphaGeometry, a theorem prover for Euclidean plane geometry that sidesteps the need for human demonstrations by synthesizing millions of theorems and proofs across different levels of complexity. AlphaGeometry is a neuro-symbolic system that uses a neural language model, trained from scratch on our large-scale synthetic data, to guide a symbolic deduction engine through infinite branching points in challenging problems. On a test set of 30 latest olympiad-level problems, AlphaGeometry solves 25, outperforming the previous best method that only solves ten problems and approaching the performance of an average International Mathematical Olympiad (IMO) gold medallist. Notably, AlphaGeometry produces human-readable proofs, solves all geometry problems in the IMO 2000 and 2015 under human expert evaluation and discovers a generalized version of a translated IMO theorem in 2004. <<<