Human language facilitates the communication of exceptionally nuanced meanings. However, the neural mechanisms underlying the brain's representation of linguistic meaning remain largely unknown. While research has identified brain regions involved in linguistic and semantic processing, the cellular-level mechanisms and temporal dynamics of meaning derivation remain elusive. Previous studies have utilized neuroimaging techniques, providing insights into brain regions associated with semantic processing. These studies suggest that semantic processing may be broadly distributed or concentrated in semantic hubs. However, these methods lack the spatial and temporal resolution to fully elucidate how individual neurons represent and process word meanings during natural language comprehension. The current study addresses this gap by employing single-neuronal recordings, offering a unique opportunity to investigate the real-time dynamics of word and sentence comprehension at a combined spatial and temporal resolution.

Existing literature highlights the importance of understanding the neural substrates of both linguistic and semantic processing. Studies utilizing fMRI have identified brain areas associated with lexical and syntactic processing, including the left-lateralized network in frontal and temporal regions. However, there is ongoing debate regarding the distribution of semantic processing, with some suggesting a broad cortical distribution while others propose the existence of a few concentrated semantic hubs. These studies also lack the detailed resolution to understand the role of single neurons. Distributional models of meaning posit that words sharing similar contexts tend to have similar meanings. This has been supported by data-driven word embedding approaches, which have shown promise in capturing human semantic judgments. Despite progress in understanding semantic processing through brain imaging, how individual neurons process and represent word meanings during language comprehension remains unclear. Moreover, the influence of context on meaning representations and their instantiation at a cellular level remains largely unknown. The highly structured nature of semantic knowledge is another key area requiring further investigation at the neuronal level.

The researchers used a rare opportunity to record single-neuron activity in humans undergoing planned intraoperative neurophysiology. Recordings were obtained from the left prefrontal cortex, a region known to be involved in semantic processing, using tungsten microelectrode arrays and Neuropixels arrays. A total of 287 units were recorded from 13 participants (18 sessions using microarrays and 3 sessions using Neuropixels). Participants listened to semantically diverse naturalistic sentences, word lists, nonwords, and stories while their neural activity was recorded. Custom software aligned action potentials to each word with millisecond resolution. A word embedding approach was used, where each unique word was represented as a 300-dimensional vector from a large English corpus. Spherical clustering and silhouette analysis were used to identify semantic domains, such as actions, states, objects, food, animals, etc. The selectivity of neurons to specific word meanings was assessed using a selectivity index (SI), ranging from 0 (no selectivity) to 1 (complete selectivity to one domain). Decoding analyses were performed to determine the ability of neuronal activity to predict semantic domains. To examine the context-dependence of meaning representations, the researchers used a word-list control, removing sentence context, and analyzed responses to homophone pairs. Finally, a hierarchical clustering procedure was used to explore the organization of semantic representations within the neuronal population.

The study revealed that a significant proportion of neurons (14% from microarray recordings, 19% from Neuropixels recordings) in the left prefrontal cortex exhibited selectivity for specific semantic domains. These neurons showed preferential responses to specific word meanings and reliably distinguished words from nonwords. The selectivity index (SI) indicated the degree to which neurons responded selectively to words within specific semantic domains. The mean SI across all selective neurons was 0.32 (microarray) and 0.42 (Neuropixels). The responses of these neurons were highly dynamic, reflecting the words' meanings based on their specific sentence contexts and independent of their phonetic form. Multi-class decoders trained on neuronal responses accurately predicted the semantic domains of words, even when tested on novel sentences (story narratives). This demonstrated the generalizability and robustness of these meaning representations. The researchers also observed that the neurons distinguished real words from nonwords. Analysis revealed that the population response patterns captured the semantic relationships among words. The activity of semantically selective neurons closely correlated with the cophenetic distances between words in the word embedding space, indicating that the neurons encoded hierarchical semantic relationships. The study's findings were consistent across participants and different recording techniques (microarrays and Neuropixels). The contextual dependence of neuronal activity was demonstrated by decreased SI during word-list presentations compared to sentence presentations, highlighting the dynamic nature of semantic processing.

The findings provide compelling evidence that individual neurons in the human prefrontal cortex encode specific word meanings and that these representations are dynamic and context-dependent. The ability to accurately decode semantic domains from neuronal activity, even when using novel linguistic materials, suggests a robust and generalizable mechanism for semantic processing. The observed correlation between neuronal activity and hierarchical semantic relationships further supports the rich and detailed nature of semantic encoding at the cellular level. These results challenge previous notions of broadly distributed semantic processing and suggest that focal cortical areas can represent complex meanings effectively. The context-dependent nature of neuronal responses underlines the importance of sentence context in shaping semantic interpretation. The study's consistent findings across multiple recording techniques enhance the robustness of its conclusions. These findings contribute significantly to our understanding of the neural basis of human language.

This study provides the first direct evidence at the single-neuron level of how the human brain represents word meanings during language comprehension. The findings reveal that individual neurons in the left prefrontal cortex show remarkable selectivity for specific semantic domains, dynamically adjust their responses based on sentence context, and contribute to a population-level representation of hierarchical semantic relationships. Future research could explore modality independence of semantic representations, cross-linguistic comparisons, and the role of other brain regions in semantic processing. Investigations into finer-grained semantic distinctions and how word-level representations combine to form phrase- and sentence-level meanings are also crucial next steps.

The study's sample size, while substantial for single-neuron recordings in humans, may still be considered limited. The recordings were focused on a specific region of the prefrontal cortex; therefore, the generalizability of these findings to other brain areas involved in semantic processing remains to be explored. The study primarily focused on auditory language processing; thus, future research should investigate the extent to which these findings generalize to other modalities, such as reading. Finally, while the study controls for several factors, potential confounding variables related to individual differences in cognitive abilities or attentional state may exist.