Toward a Neural Code for Language
Théo Desbordes, Yair Lakretz, Valérie Chanoine, Maxime Oquab, Jean-Michel Badier, et al.
(see pages 5350–5364)
“The quick brown fox jumped over the lazy dog.” Those nine words, arranged into a simple sentence, describe a scene that even a young child can picture. How the brain makes the leap from hearing (or reading) a string of words to comprehending their intended meaning remains mysterious, particularly for more complex sentences like this one. This week, Desbordes et al. try to crack the neural code behind semantic composition with a twofold hypothesis: that the neural representation of a sentence would increase in complexity as its meaning is being built, and that neural signatures of word processing, multiword integration, and sentence wrap-up should be detectable in distinct neural assemblies. After validating these ideas using deep Neural Language Models (NLMs), the researchers recruited 11 participants who had intracranial stereotactic electroencephalography (sEEG) implanted for clinical purposes and who could be simultaneously recorded with magnetoencephalography (MEG). The participants read either normal sentences or nonsensical “Jabberwocky” sentences made up of pseudowords. Using a measure of neural complexity called intrinsic dimensionality (ID) from sEEG or MEG data, the authors found higher ID for normal sentences than for Jabberwocky sentences or simple lists of words, and ID increased as a sentence unfolded. The researchers then showed that NLMs demonstrate a similar ramping activity pattern that varies with semantic composition, as well as response indicative of sentence wrap-up. Together the analyses showed three putative processing stages of semantic composition in brains and language models: a phasic effect early in processing, followed by ramping and then end-of-sentence activity at higher processing levels. In terms of activity in specific brain regions, the ventral occipitotemporal visual pathway, including the fusiform gyrus, was strongly activated by the lexical component but not ramping or end-of-sentence components of the signal, indicative of its role in recognizing written words. Regions including the middle and superior temporal gyri and inferior parietal lobule, however, were activated with both lexical and ramping patterns, suggesting a more sophisticated role in comprehension. Finally, frontal brain regions displayed mixed signatures of ramping, sentence-final, and phasic activity. According to the authors, the findings constrain the search for a neural code of linguistic composition.
GFP labeling indicates infiltration of MM cells into the dorsal root ganglion in a mouse model of MM.
Origins of Multiple Myeloma-Induced Bone Pain in Mouse and Human
Marta Diaz-delCastillo, Oana Palasca, Tim T. Nemler, Didde M. Thygesen, Norma A. Chávez-Saldaña, et al.
(see pages 5414–5430)
Multiple myeloma (MM) is a rare blood cancer that originates in bone marrow and causes bone damage and notoriously excruciating pain, but the source of that pain remains unknown. Current pain-management therapies are mostly limited to opioids, which have serious known side effects. As disease-modifying therapies increasingly extend the lives of people with MM, finding mechanism-based treatments for myeloma-induced bone pain (MIBP) is urgent. This week, Diaz-delCastillo et al. make tremendous strides toward a new understanding of the neuronal changes underlying MIBP in mouse and human. In a mouse model of MM, the emergence of nonevoked pain-like behaviors at postsurgical day 24 (D24) corresponded with the development of cortical bone lesions. Their previous work showed bone marrow denervation in end-stage MM mice. To see earlier disease stages, the researchers labeled sensory and sympathetic nerve fibers with calcitonin gene-related peptide (CGRP) and tyrosine hydroxylase, respectively. As early as D17, before the onset of pain behaviors, the researchers saw no evidence of either fiber in bone marrow. In periosteum, the most densely innervated bone compartment, MM cells had infiltrated by D24 but not by D17. And, importantly, CGRP+ nerve fiber density had increased significantly, suggesting that nerve sprouting contributes to pain onset. Bone biopsy samples from 13 newly diagnosed MM patients also showed increased periosteal nerve density compared with cancer-free control subjects, which also corresponded with MM cell infiltration. Mice treated with an antibody directed at netrin-1, an axon-guidance molecule required for nerve sprouting, displayed delayed pain-like behaviors. The researchers next examined the transcriptome of the dorsal root ganglia (DRGs), where sensory nerves reside. Remarkably, the profile suggested MM cell infiltration of the DRGs, a never-before-seen feature of the disease that was confirmed with histology, which also revealed vascular and nerve damage in the DRG. The surprising results were supported by a study of DRGs recovered from one MM patient, which also showed a transcriptomic signature consistent with MM infiltration not seen in DRGs from patients with other cancers or who were cancer free. The findings provide new insights into the mechanisms behind MIBP that could point the way to new pain therapeutics for people with MM.
Footnotes
This Week in The Journal was written by Stephani Sutherland
