2025 年 45 巻 3 号 p. 159-175
Primary progressive aphasia (PPA) offers a unique lens into understanding the neural organization of speech and language and the disease mechanisms of neurodegeneration. This article reflects on how the study of PPA―through its three canonical clinical variants：semantic, logopenic, and non-fluent/agrammatic―has transformed our understanding of speech and language networks, clinical-pathological correlations, and disease trajectories and vulnerabilities. Drawing from two decades of multidisciplinary clinical research across Europe, Asia, and the United States, we explored how large-scale PPA studies have revealed the complex brain networks involved in speech and language processing and shown that distinct molecular pathologies exhibit selective affinity towards these networks in variant-specific manners. We also expounded how language typology and neurodevelopmental trajectories critically shape disease manifestations, highlighting the need for linguistically tailored and multidimensional clinical care strategies that incorporate precision diagnosis tools, biomarker-driven diagnostics, speech rehabilitation, and emerging molecular based disease-modifying therapies. This perspective article advocates for ongoing international and cross-linguistic collaborations to refine diagnostic criteria and improve care for individuals with PPA worldwide, regardless of language and neurodevelopmental backgrounds.
Our initial understanding of primary progressive aphasia （PPA） has largely stemmed from the clinical descriptions of patients with neurodegenerative diseases, who exhibit isolated and progressive speech and language deficits in the early stages of their disease （Pick, 1892；Rosenfeld, 1909；Sérieux, 1893）. The seminal work of Marsel Mesulam in the early 1980s was pivotal in identifying PPA as a distinct clinical spectrum （Mesulam, 1982, 1987）. At that time, Mesulam’s work were largely derived from detailed clinical observations, as MRI and other brain imaging technologies were far from routine, and the general understanding of dementia was almost exclusively linked to memory loss. His recognition that progressive speech and language impairment could be the first and predominant manifestation of a neurodegenerative disease marked a significant departure from previous understandings, even though earlier foundational case reports from Pick et al. （Pick, 1892）, Warrington et al. and Imura et al. （Imura, 1943） had already laid important groundwork. Mesulam’s earlier work also noted that brain atrophy in PPA typically affected the perisylvian regions of the left hemisphere, akin to what was seen in stroke aphasia patients, and was likely not associated with Alzheimer’s pathology （Cummings et al., 1985；Heath et al., 1983；Mehler, 1988；Mesulam, 1982, 1987）. Over the past four decades, thanks to Marsel’s and others’ revolutionary work, researchers have been able to establish larger PPA cohorts. Through these cohorts, we now learn that while Mesulam’s initial descriptions of PPA remain largely accurate, PPA can also exhibit atrophy beyond the perisylvian regions （Gorno-Tempini et al., 2004a；Wilson et al., 2009） and can be sometimes associated with Alzheimer’s disease pathology （Rabinovici et al., 2008；Santos-Santos et al., 2018；Spinelli et al., 2017）.
As interest in PPA grew, clinicians and scientists began to observe a variety of clinical presentations in PPA patients that did not entirely conform to a single clinical phenotype. John Hodges, Julie Snowden and colleagues described semantic dementia, in which fluency of speech remained intact, but word comprehension and naming were severely compromised due to semantic memory impairment and were associated with left anterior temporal lobe atrophy （Hodges et al., 1992；Mummery et al., 2000；Snowden et al., 1989）. The clinical phenotype resembled what Mesulam had termed fluent PPA；however, the pattern of anatomical damage did not consistently align, leaving the relationship between the two entities unclear. Around the same time, Murray Grossman and coworkers introduced the concept of progressive non-fluent aphasia, which emphasized grammatical processing deficits and was associated with left frontal atrophy （Grossman et al., 1996；Turner et al., 1996）. Further complicating matters, some of these clinically distinct syndromes showed unexpectedly high frequencies of Alzheimer’s disease pathology upon postmortem analysis （Caso et al., 2013；Spinelli et al., 2017；Turner et al., 1996）. This finding suggested that primary progressive aphasia could be associated with both frontotemporal lobar degeneration （FTLD） and Alzheimer’s disease pathologies.
Further research undertaken across multiple centers, including groups in the United Kingdom （UK） and the United States （US）, has offered critical insights into this issue. Notably, at the University of California, San Francisco （UCSF） Memory and Aging Center, systematic cognitive and neuroimaging assessments of a large PPA patient cohort led to the identification of the logopenic variant. This variant is marked by evident phonological and lexical retrieval deficits and is typically associated with left posterior temporal-parietal atrophy （Gorno-Tempini et al., 2004a, 2008；Henry & Gorno-Tempini, 2010；Rohrer et al., 2013）. Consequently, three canonical PPA variants were recognized：semantic, non-fluent/agrammatic, and logopenic variant PPAs （Gorno-Tempini et al., 2004a, 2011；Mesulam et al., 2009b；Wilson et al., 2009）. To enhance uniformity in research reporting and foster collaboration, an international group of PPA researchers published the consensus criteria for classifying the three canonical variants of PPA in 2011 （Gorno-Tempini et al., 2011）. This tripartite classification is not entirely rigid, and healthy scientific debate continues, but it has proved valuable in linking clinical features, neuroanatomical patterns, and underlying molecular pathologies, all of which bear relevance for guiding diagnosis, prognosis, and research into future treatments. Of note is that the proposed classification acknowledges the occurrence of mixed cases , often termed as “PPA unclassifiable”, especially in individuals carrying pathogenic FTD mutations （Saracino et al., 2021）and at later stages of the disease.
One of the most transformative shifts in our understanding of neurodegenerative diseases has been the recognition that these are network-based disorders （Agosta et al., 2010；Seeley et al., 2009）. Each variant of PPA appears linked to a specific large-scale brain network with a particular “epicenter” of vulnerability and pattern of pathology spreading through network connections （Guo et al., 2013；Mandelli et al., 2018；Ranasinghe et al., 2017）. When neurodegenerative disease strikes that epicenter, it gradually spreads along established anatomical or functional connections. The network-based approach to neurodegenerative diseases hypothesizes that disease progression follows connectivity pathways, such that pathology spreads from the epicenter to connected nodes, generating predictable symptom trajectories. Seminal work with resting-state functional MRI reveals distinct intrinsic connectivity networks in healthy adults, each corresponding to specialized functions such as language, episodic memory, or socioemotional processing （Battistella et al., 2020）.
This network-based view has revolutionized our understanding of disease progression and enhanced our ability to anticipate the evolution of clinical symptoms. It has also prompted intriguing questions about why specific neural networks exhibit increased susceptibility to particular pathologies in certain individuals. Within the spectrum of PPA, studies have indicated that distinct speech and language networks exhibit selective vulnerability to specific proteinopathies, shaped in part by developmental factors （Spinelli et al., 2017）. As a result, this model―where proteinopathies preferentially target specific connectivity networks with varying degrees of individual vulnerability―partially accounts for the clinical heterogeneity that underlies distinct PPA syndromes and the diverse clinical manifestations observed across individuals. For instance, a left frontal cortical-subcortical network―involving the premotor areas （comprising the supplementary motor area（SMA） and pre-SMA）, frontal cortex, insula, basal ganglia, and associated white matter tracks （i.e. aslant track）―is particularly vulnerable to corticobasal degeneration 4R tauopathy. However, the precise location of the initial epicenter within this network likely determines the early clinical presentation：patients may first manifest with a nonfluent aphasia or motor speech syndrome, or a limb-predominant motor syndrome characteristic of corticobasal syndrome （Gorno-Tempini et al., 2004b；Josephs et al., 2010；Lee et al., 2011；Mandelli et al., 2018）. Similarly, within a posterior-temporal and inferior parietal network ―regions more vulnerable to AD pathology, patients may initially present with distinct phenotypic profiles：either a lexical-semantic phenotype, characterized by naming and auditory comprehension difficulties, or a phonological short-term memory phenotype, marked by deficits in repetition and working memory （Mandelli et al., 2023；Ramanan et al., 2022）.
Among the three PPA variants, the semantic variant―also called semantic dementia―was the first to be formally characterized, with seminal work published by the UK group in the early 1990s （Hodges et al., 1992；Snowden et al., 1989）. Clinically, patients present with fluent and grammatically intact speech but have significant impairments with confrontation naming and single-word comprehension. These deficits arise from a degradation of semantic memory―the store of conceptual knowledge that enables us to recognize objects, people, and words independent of context （Gorno-Tempini et al., 2011）. In everyday life, individuals living with semantic variant PPA often struggle to name common objects and comprehend word meaning, despite retaining normal articulation and syntax structure until later stages of the disease. Surface dyslexia in English is a particularly variant-specific symptom （Gorno-Tempini et al., 2011；Hodges et al., 1992）. Patients with semantic variant PPA often misread irregular words because they lose the semantic knowledge of the specific orthography-phonology mappings, defaulting to a purely phonological decoding approach （Binney et al., 2016）. Although surface dyslexia is most well demonstrated in English, analogous phenomena manifest in other languages in distinct manners―for example, selective difficulties with reading and writing kanji script, but not kana script, in Japanese. This is largely due to the nature of kanji, which requires the memorization of whole-word forms, in contrast to kana, which relies on phoneme-based decoding （Imura, 1943）. The cardinal neuroanatomical finding is bilateral atrophy of the anterior temporal lobes, typically more pronounced on the left side （Gorno-Tempini et al., 2004a；Mesulam et al., 2009b；Mummery et al., 2000）. This corresponds well with the hypothesis that conceptual knowledge depends on the anterior temporal lobe as a crucial hub, integrating modality-specific representations of semantic concepts from high-level association areas （Rice et al., 2015；Zhao et al., 2017）.
Individuals living with semantic variant PPA exhibit characteristic familiarity and prototypicality effects, leading to category-level naming errors shaped by the progressive degradation of conceptual features （Mesulam et al., 2009a）. As finer-grained semantic distinctions are lost, only the most general attributes of a concept may be retained―leading, for example, to all animals being labeled as “dog,” the most familiar exemplar within that category. As disease progresses, patients show more profound semantic knowledge deficits extending beyond words and objects. These manifest as difficulties in identifying famous people （Borghesani et al., 2019）, recognizing environmental sounds （Goll et al., 2010）, and interpreting emotional cues （Multani et al., 2017）. When disease progress and involve the right anterior temporal lobe and orbitofrontal regions, patients frequently manifests as behavioral changes, including mental rigidity, reduced empathy, or compulsive behaviors （Kumfor et al., 2016；Ulugut Erkoyun et al., 2020；Younes et al., 2022）. Individuals who present with predominant right anterior temporal atrophy may initially exhibit prominent socio-emotional semantic deficits and thus do not meet criteria for PPA and instead match the description of the semantic-behavioral variant FTD or right temporal variant FTD （Ulugut Erkoyun et al., 2020；Younes et al., 2022）. These patients often first seek psychiatric care and only later in the disease course, when left temporal involvement and word comprehension deficits arise, do they seek the attention of neurologists （Binney et al., 2016；Thompson et al., 2003）.
Pathologically, semantic variant PPA is most frequently associated with frontotemporal lobar degeneration with TDP-43 inclusions （FTLD-TDP, often type C）（Rohrer et al., 2010；Spinelli et al., 2017）. Approximately 80％ or more of semantic variant PPA patients exhibit underlying TDP-43 pathology, although a small subset of patients may be associated with Pick’s disease （a form of FTLD-tau）. In rare instances, semantic variant PPA may be associated with Alzheimer’s disease pathology, however, such cases are exceptional, usually occur in older patients and often warrant consideration of co-existing neuropathologies （Hodges et al., 2010）. This relatively consistent pathology underpins the clinical-anatomical correlation seen in the semantic variant and stands in clear contrast to the logopenic variant, which is more typically associated with Alzheimer’s disease pathology.
The logopenic variant PPA emerged from detailed investigations of PPA patients at UCSF who exhibited neither the classic non-fluent/agrammatic deficits nor the key features of semantic dementia （Gorno-Tempini et al., 2004a, 2008）. Instead, these individuals displayed a unique pattern of hesitant speech marked by frequent word-finding pauses, impaired sentence repetition―particularly for longer, less meaningful sentences―and phonological errors, such as sounds substitutions or transpositions in phonologically complex words （Gorno-Tempini et al., 2004a, 2011；Henry et al., 2016）. Despite these expressive challenges, their motor speech articulation remains relatively preserved compared to the non-fluent variant. The core underlying deficit involves auditory verbal short-term memory in the phonological loop system （Lukic et al., 2019）.
Neuroimaging studies have associated the logopenic variant PPA with atrophy predominantly in the left posterior temporal and inferior parietal regions, an area strongly involved in phonological processing （Mandelli et al., 2023；Tetzloff et al., 2018）. Intriguingly, most of logopenic variant PPA patients exhibit underlying Alzheimer’s disease neuropathology, as confirmed by biomarkers including amyloid positron emission tomography（PET）, cerebrospinal fluid analyses, or postmortem autopsy findings （Rabinovici et al., 2008；Santos-Santos et al., 2018）. This realization was pivotal, shedding lights on earlier reports of “atypical” Alzheimer’s pathology in certain cases of PPA patients （Caso et al., 2013；Spinelli et al., 2017）. Rather than categorizing these patients as “non-fluent” or “unclassifiable”, recognizing the distinct features of logopenic syndrome allowed for a more precise clinical-pathological correlation. Given that the left posterior temporal and inferior parietal lobes lie at the intersection of multiple cognitive domains, individuals with logopenic variant PPA can display other non-language symptoms, such as apraxia or dyscalculia, paralleling a Gerstmann-like syndrome （Funayama et al., 2013；Ramanan et al., 2022）. Nonetheless, the defining clinical hallmark―and the one that most profoundly disrupts daily communication―is the presence of slow, hesitant speech with phonological retrieval failures and impaired sentence repetition.
An especially noteworthy feature of logopenic variant PPA patients is the higher frequency of subtle―or even overt―histories of developmental dyslexia or familial language-based learning difficulties, suggesting an underlying developmental vulnerability within the same posterior temporal-parietal network （Miller et al., 2013；Miller et al., 2018）. While this vulnerability may not directly cause Alzheimer’s disease, it may help explain why Alzheimer’s pathology selectively targets language networks in these individuals. Although the detailed mechanisms remain unclear, the notion that neurodevelopmental factors contribute to selective vulnerability in later-life neurodegenerative processes opens an intriguing line of research―one that underscores the importance of lifespan approaches to understanding brain health.
The third canonical variant, commonly referred to as the nonfluent/agrammatic variant, is characterized by effortful, halting speech with prominent grammatical deficits and/or motor speech challenges, such as apraxia of speech and dysarthria （Gorno-Tempini, et al., 2004a, 2011）. Early descriptions, particularly from Murray Grossman’s group, emphasized a primarily syntactic and morphological deficit （Grossman et al., 1996）, while subsequent work highlighted the key role of motor speech impairment in many individuals with nonfluent/agrammatic variant PPA （Josephs et al., 2006；Ogar et al., 2007）. This variant is now understood as encompassing a clinical spectrum, in which some individuals present with prominent apraxia of speech and or dysarthria alongside relatively preserved grammar, while others exhibit marked grammatical breakdown. Over time, these deficits often co-occur, reflecting a shared vulnerability over the left dorsolateral frontal and insular regions that are critical for speech planning and production （Lorca-Puls et al., 2024；Mandelli et al., 2018, 2023）.
Neurodegeneration in the nonfluent/agrammatic variant typically centers on the inferior frontal gyrus, premotor cortex, and supplementary motor area, often extending to underlying white matter tracts and sometimes the basal ganglia （Gorno-Tempini et al., 2004b；Josephs et al., 2006）. This anatomical localization provides a plausible explanation for why a substantial subset of individuals with the nonfluent/agrammatic variant subsequently develop clinical features characteristic of corticobasal syndrome or progressive supranuclear palsy （Kertesz et al., 2000；Santos-Santos et al., 2016）. The underlying pathology in nonfluent/agrammatic variant PPA is most commonly a tauopathy （approximately 75％ of cases）, specifically corticobasal degeneration, progressive supranuclear gaze palsy, or Pick’s disease （Shir et al., 2024；Spinelli et al., 2017）. The remaining cases （～25％） typically exhibit TDP-43 pathology, often in the context of specific genetic mutations, such as those involving the progranulin gene （Saracino et al., 2021）.
Although diagnostic criteria tend to emphasize “typical” presentations, real-world patients frequently evolve into mixed or overlapping phenotypes as disease spreads along vulnerable neural networks. For example, an individual who initially meets criteria for semantic variant PPA may later develop profound behavioral changes if right anterior temporal atrophy becomes more pronounced （Kumfor et al., 2016；Multani et al., 2017）. Similarly, a patient with logopenic variant PPA might eventually show signs of visuospatial or executive impairments as the disease extends dorsally （Rohrer et al., 2013；Watson et al., 2018）. Those with nonfluent/agrammatic variant patients can develop widespread involvement of the basal ganglia and supplementary motor area, leading to a clinical picture consistent with corticobasal syndrome （Santos-Santos et al., 2016）.
Nevertheless, a trained eye can usually discern the hallmark features of each PPA variant even years into the disease course ―at least until the most advanced stages, when extensive cortical degeneration obscures the expression of variant-specific clinical signatures. The overlapping patterns of progression observed across PPA variants support the framework that each syndrome arises from a distinct speech and language connectivity network, with one region acting as the initial epicenter （Agosta et al., 2010；Mandelli et al., 2023；Seeley et al., 2009）. As the disease advances, it propagates along the established pathways within these networks, giving rise to characteristics gradients of atrophy and corresponding clinical decline.
The question of why a particular network becomes selectively vulnerable in an individual patient remains a compelling area of investigation. Multiple factors likely contribute to this selective vulnerability, including genetic predisposition （Lee et al., 2019；Ramos et al., 2019）, neurodevelopmental trajectories （Miller et al., 2013）, environmental influences （Asken et al., 2024）, and stochastic processes. Currently researchers are examining the offsprings of individuals with familial frontotemporal dementia to identify subtle neurodevelopmental markers that could reflect future vulnerability. Additional lines of investigation are exploring the potential intersection of typical aging processes, the propagation of neuropathological proteins, and the influence of individual learning histories―factors that may collectively shape cortical organization and, ultimately, its susceptibility to neurodegeneration.
Cross-linguistic studies offer critical and novel insights into the neural and cognitive mechanisms underlying PPA and aphasia in general. Despite the vast linguistic diversity worldwide―with nearly 83％ of the world’s population speaking a language other than English as their first or second language―the majority of existing PPA research remains relies heavily on English-speaking cohorts （Coppieters et al., 2024）. For instance, approximately 72％ of studies focused on logopenic variant PPA （lvPPA） are based on English native speakers, raising important questions about the generalizability of current findings to non-English populations （García et al., 2023）.
Language typology―including linguistic features such as orthographic transparency, lexical tone, and phonological and morphological complexity―have been shown to influence the phenotypic expression of PPA variant emerge across diverse linguistic contexts （Auclair-Ouellet et al., 2016a, 2016b；Billette et al., 2020；Canu et al., 2020；Hosogi Senaha et al., 2006；Kavé et al., 2012；Ralph et al., 2011；Sasanuma & Monoi, 1975；Tee et al., 2022a, 2022b；Wilson & Martinez-Cuitino, 2012；Yamadori, 2019）. One of the most notable illustrations of language typology effects in PPA is the phenomenon of over-regularization observed in individuals with the semantic variant. Over-regularization tends to arise in languages where input-output mappings follow regular patterns, with a subset of exceptions that rely on intact semantic knowledge to identify. In semantic variant PPA patient, where semantic knowledge is profoundly degraded, patients often fail to recognize exceptions and default to regular patterns. This results in characteristic errors―such as surface dyslexia in English, where exceptional/irregular words are misread using sublexical, phonological decoding rules. In contrast, speakers of languages with more transparent orthographies―such as Italian, German, and Spanish―rarely exhibit surface dyslexic errors, likely due to their consistent grapheme-to-phoneme mappings, which reduce reliance on semantic processing during reading. Instead, over-regularization phenomenon in these languages often emerges through syntactic or morphological errors in other language tasks （Auclair-Ouellet et al., 2016a, 2016b；Billette et al., 2020；Canu et al., 2020；Kavé et al., 2012；Ralph et al., 2011）. Rather than surface dyslexia, over-regularization in Spanish often manifests in noun article usage. Typically, nouns ending in -o are paired with the masculine article el, while those ending in -a take the feminine article la. However, in cases of irregular noun article assignment―such as “el problema”―individuals with semantic variant PPA may over-regularize and produce “la problema” instead （Ralph et al., 2011）. In French, over-regularization in svPPA often appears in noun-to-verb derivation. While verbs are typically formed by replacing noun endings with standard affixes like “-eur” or “-ateur”, svPPA patients may misapply these regular patterns to irregular forms―for instance, producing “correcter” instead of “corriger”（to correct） from “correcteur”（corrector）. A similar pattern is observed in German, where over-regularization often affects past tense verb inflections （Billette et al., 2020）. On the other hand, Italian speakers with svPPA have been reported to exhibit stress assignment errors during oral reading―specifically, overgeneralizing the default pattern of assigning lexical stress to the penultimate syllable, even in exceptional words with antepenultimate stress （Galante et al., 2000）. In Japanese speakers with semantic variant PPA, the phenomenon known as Legitimate Alternative Reading of Components （LARC）was reported to refer to a distinctive reading error in which compound kanji words are misread by applying alternative―though legitimate―pronunciations of their individual characters. This pattern reflects an over-reliance on frequency-related lexical phonological rules in the absence of intact semantic support （Fushimi et al., 1999, 2003）.
The selective impairment of semantic knowledge in semantic variant PPA also manifests in different linguistic components across languages. In Japanese speakers, this is showcased in pronounced difficulty reading kanji―logographic characters that rely heavily on semantic processing―while reading of kana, a phonologically transparent script, remains relatively preserved （Sakurai et al., 2021；Sasanuma & Monoi, 1975；Yamadori, 2019）. This dissociation, often referred to as gogi aphasia （literally, “word meaning aphasia”）, parallels the selective difficulty with reading irregular words observed in English-speaking individuals with semantic variant PPA, and is uniquely apparent in biscriptal languages like Japanese, where scripts differentially rely on semantic versus phonological processing. In classifier languages, semantic impairment has been found to manifest in the production of specific noun classifiers. In Chinese, the use of classifiers is notably semantically arbitrary―for instance, 把（bǎ） is used for a diverse set of nouns such as fire （火）, comb （梳子）, salt （鹽）, knife （刀）, and umbrella （傘）, despite limited shared semantic or physical attributes. Accurate classifier selection thus relies heavily on intact semantic networks. Chinese-speaking individuals living with semantic variant PPA frequently struggle with selecting appropriate classifiers, a deficit shown to correlate with atrophy in the left anterior temporal lobe （Tee et al., 2024）. As this region is critical for semantic knowledge, the finding underscores the essential role of conceptual-semantic integrity in classifier production―a phenomenon particularly salient in classifier languages, where classifier use is obligatory when determiners or numerals precede nouns, and where classifiers are employed with high frequency and complexity.
Similarly, motor speech and grammatical deficits characteristic of the non-fluent/agrammatic variant PPA vary considerably with language typology. English speakers often present with apraxia of speech （AOS）, particularly through speech distortions in multisyllabic words with complex consonant clusters. In contrast, speakers of languages with simpler phonotactic structures, such as Italian, more commonly exhibit phonological paraphasias （Canu et al., 2020）. Despite comparable patterns of cortical atrophy, Italian speakers with nonfluent/agrammatic variant PPA tend to produce syntactically simpler utterances and perform more poorly on syntactic comprehension tasks than their English-speaking nonfluent/agrammatic variant PPA―possibly reflecting the greater syntactic complexity and inflectional demands of the Italian language. These clinical differences underscore how the phonological and grammatical structures of a given language can shape the presentation―and possibly the threshold at which clinical symptoms of underlying network vulnerability emerge.
Furthermore, language typology may also play a critical role in shaping the design and interpretation of assessment tools for specific speech and language deficits. In Chinese, although the language is predominantly monosyllabic and lacks complex consonant clusters, lexical tone has emerged as a valuable and sensitive element for detecting motor speech impairments in individuals with non-fluent/agrammatic variant PPA （Tee et al., 2022a）. Similarly, in agglutinative languages such as Turkish, words often comprise multiple morphemes and syllables. Turkish-speaking individuals with logopenic variant PPA and amnestic Alzheimer’s dementia exhibit increasing difficulty not only with word count but also with increasing syllable counts, highlighting underscoring the impact of morphological and phonological complexity on auditory verbal short-term memory span （Öz et al., 2023）. In addition, Chinese language users display distinct patterns of writing errors across PPA variants compared to English speakers, largely attributed to the structural differences between Chinese logographic script and the alphabetic system of English （Tee et al., 2022b）. These examples highlight that even when the same linguistic tasks are administered, scoring and evaluation methods must be adapted to account for typological differences across languages.
All these findings underscore the vital importance of cross-linguistic collaboration and linguistic diversity in PPA research. However, compared to English-speaking populations, research cohorts of PPA patients who speak non-Indo-European languages remain relatively small with limited typologically specific characterization, revealing a significant gap in our understanding of how the disease manifests across diverse linguistic systems. To promote more inclusive linguistic representation in brain health research, our group launched the Chinese Language Assessment in PPA （CLAP） project in 2019. In collaboration with researchers across Taiwan, Hong Kong, and the United States, CLAP focuses on characterizing Chinese-speaking PPA patients using linguistically and culturally tailored approaches （Tee et al., 2021）. In parallel, we co-founded the International Network for Cross-Linguistic Research on Brain Health （INCLUDE）, a global consortium dedicated to advancing cross-linguistic research and promoting language diversity in neurodegenerative and neurodevelopment conditions. INCLUDE provides a collaborative platform for researchers worldwide, fostering nuanced understanding of language-specific clinical phenotypes and typological features―recognizing that no single investigator can possess expertise across all linguistic systems （García et al., 2023；Tee et al., 2022b）. Such global efforts hold immense potential for the development of refined diagnostic criteria that apply to different languages and improve therapeutic interventions worldwide （Mazzeo et al., 2024）. Beyond clinical implications, cross-linguistic research also enriches theoretical models of brain―language relationships by testing them across diverse linguistic frameworks, ultimately deepening our understanding of how the brain supports speech and language functions, and advancing care for individuals with PPA worldwide.
Historically, clinicians had limited treatment options to offer patients with PPA. Speech-language interventions were often viewed with skepticism among insurance providers or funding agencies, who presumed that, unlike stroke, neurodegenerative diseases inevitably lead to progressive and irreversible decline. However, an accumulating body of evidence from multiple research centers has demonstrated that targeted speech therapy can preserve―and in some cases, enhance―speech and language functions, especially when interventions are tailored to address variant-specific impairments and strategically leverage preserved cognitive and neural networks （Grasso et al., 2021；Henry et al., 2019；Volkmer et al., 2020）. For instance, in individuals with semantic variant PPA, phonological-based interventions may be particularly effective by leveraging on the relatively preserved dorsal language stream. In contrast, patients with logopenic or nonfluent/agrammatic variant variants may benefit more from interventions targeting lexical retrieval and the strategies that draw upon preserved semantic knowledge （Henry et al., 2019；Volkmer et al., 2020）. Emerging evidence also supports the usage of teletherapy, which expands access to such specialized care for a broader and more diverse population （Dial et al., 2019）. When coupling non-invasive brain stimulation with speech therapy, studies have shown promise in enhancing adaptive reorganization within partially intact language networks （LoBue et al., 2024；Nissim et al., 2020）.
Pharmacological trials targeting PPA patients are also becoming increasingly common. For logopenic variant PPA patients, who typically harbor Alzheimer’s disease pathology, anti-amyloid and emerging anti-tau treatments offer new therapeutic possibilities particularly when initiated early, before extensive cortical damage occurs. For PPA variants associated with FTLD pathologies, several innovative molecular or gene-based strategies have entered clinical trials, including tau-directed therapies （Dunning et al., 2024）, progranulin gene therapies （Sevigny et al., 2024）, or agents targeting TDP-43 proteinopathy （Ljubenkov et al., 2022）. Even if these therapies yield only modest effect in slowing disease progression, their combined use with speech and cognitive retraining interventions could help extend functional independence and improve quality of life for individuals with PPA, whose primary challenges center on communication （Volkmer et al., 2020）.
In recent years, many PPA researchers have moved beyond documenting language deficits to exploring positive or compensatory phenomena in PPA. Patients with semantic variant PPA might develop heightened reliance on phonological or visuospatial skills （Borghesani et al., 2020；Montembeault et al., 2023）. Some individuals develop new artistic or musical talents, suggesting that the reorganization of one brain network can sometimes release creative potential or shift cognitive resources in unexpected ways （Friedberg et al., 2023）. Incorporating this perspective into therapy could yield innovative strategies that harness preserved or even newly enhanced cognitive abilities.
As the field of PPA advances, the integration of basic and clinical research has become increasingly essential to improving care. Clinically, efforts are focusing on refining diagnostic protocols, identifying biomarkers specific to neuropathology, and personalizing treatment strategies. Concurrently, advanced neuroimaging is used to map functional connectivity, characterize network dynamics, and explore developmental vulnerabilities that may predispose individuals to atypical Alzheimer’s or frontotemporal pathologies. （Battistella et al., 2020；Illán-Gala et al., 2022；Mandelli et al., 2018）. Bridging theoretical neuroscience with clinical practice is particularly important given the resources required to recruit, follow, and study patients with atypical neurodegenerative syndromes.
The field of PPA care stands at a potentially transformative juncture. Disease-modifying therapies, particularly those targeting Alzheimer’s pathology, may shift the landscape for the logopenic variant PPA, while tau- and TDP-focused treatments continue to advance for the non-fluent/agrammatic and semantic variants. With improved biomarkers, earlier recognition of language decline, and cross-linguistic diagnostic frameworks, clinicians are better equipped to identify PPA variants early and enroll patients in targeted therapies before significant neurodegeneration occurs.
Equally important is maintaining a comprehensive view of patient and caregiver experiences. PPA deprives individuals of one of their most essential faculties：the ability to communicate fluidly. This impairment has implications for emotional well-being, social identity, and even safety in daily activities. Families and caregivers often shoulder the burden of interpreting speech and assisting with fundamental tasks that require verbal proficiency. Integrating linguistically informed speech therapy with pharmacological and psychosocial support brings us closer to person-centered model of care. The collaborative ethos that has defined the PPA field―from foundational researchers like Mesulam, Hodges, Gorno-Tempini, and Grossman, to emerging investigators expanding research across languages and cultures―remains essential. Through shared data, knowledge, and intervention strategies, we can accelerate progress toward more effective treatments and improved outcomes for individuals living with PPA.
Over the past several decades, the study of PPA has profoundly reshaped our understanding of language and cognition. Once considered a rarity within the spectrum of neurodegenerative diseases, PPA has since emerged as a powerful lens into how neural networks support and organize fundamental speech and language functions. Semantic variant PPA has revealed the important role of the anterior temporal lobes in supporting semantic knowledge, while the logopenic variant has highlighted the vulnerability of posterior temporoparietal phonological loop to Alzheimer’s pathology. The non-fluent/agrammatic variant PPA has illustrated the intricate interplay among motor planning, grammatical structuring, and executive control that underpins fluent speech production. Each of these syndromes demonstrates how disease propagates in a predictable manner through large-scale connectivity networks, reaffirming that language is not an isolated system, but an intricate ensemble of interdependent sub-systems.
The evolving conversation in the PPA field continues to gain momentum, with new challenges and promising opportunities on the horizon―particularly in the areas of clinical characterization, patient care, and cross-linguistic collaborations. As we refine models of PPA and expand available treatment options, our collective effort is poised to benefit not only the relatively small population affected by these rare syndromes, but also the broader scientific community striving to understand how the human brain encodes, retrieves, and communicates knowledge.
The work is supported by the Global Brain Health Institute （GBHI ALZ UK-19-589585）, Alzheimer’s Association （AACSFD-22-972143）, University of California, San Francisco, National Institutes of Health （NIA R21 AG068757, R01 AG080469, R01 AG083840, R01 AG075775, NIH UF1 NS100608, U01 NS128913, P01 AG019724, U19 AG079774, K24 DC015544, RF1 NS050915）, Alzheimer’s Disease Research Center of California （P30 AG062422）, National Taiwan University Hospital, Taipei, Taiwan （113-N0034, 114-P0003）, and Charles and Helen Schwab Foundations. The authors also acknowledge the use of the AI Scientific Writer, an artificial intelligence （AI） workflow developed by Pedro Pinheiro-Chagas, to draft the initial version of the manuscript based on the two talks delivered respectively by Drs. Marilu Gorno-Tempini and Boon Lead Tee at the 48th Annual Scientific Meeting of the Japan Society for Higher Brain Function. The draft was subsequently critically reviewed and substantially revised by all authors to ensure the scientific content, accuracy, and integrity of the final manuscript. Further details of the methodology are provided in the Supplementary Methods.
The authors declare no conflicts of interest associated with this manuscript.
;%0A%09%09%09newWindow.document.open();%0A%09%09%09newWindow.document.write('<img src=%22./Graphics/45_159_t1.jpg%22>');%0A%09%09%09newWindow.document.close();%0A%09%09)
Artificial intelligence（AI）workflow
The AI Scientific Writer is a comprehensive agentic workflow for converting scientific talks into publication-ready perspective articles developed by Pedro Pinheiro-Chagas. The code is publicly available：https://github.com/pinheirochagas/AI_scientific_writer. All automated computations were managed by a Python 3.11 controller script.Each computational step was encapsulated as a discrete agent function, clearly defined by its prompt, model choice, and structured inputs and outputs. Model inferences were executed using Microsoft Azure OpenAI via UCSF Versa, a secure platform employing stateless endpoints that neither retain nor utilize user data for training purposes.
Source Acquisition and Preprocessing
This manuscript provides a synthesis of two presentations delivered by Drs. Marilu Gorno-Tempini and Boon Lead Tee at the 48th Annual Scientific Meeting of the Japan Society for Higher Brain Function. The source material for the manuscript was derived from two independent datasets：an audio recording of Dr. Maria Luisa Gorno-Tempini’s lecture and the presentation slides from Dr. Boon Lead Tee’s talk. The audio stream was transcribed verbatim using Whisper-3-large, configured specifically for 16 kHz mono audio input. Slides from the PDF were converted into 300 dpi PNG images using a custom preprocessing utility. These images were subsequently processed by GPT-4o-2024-08-06, utilizing a tailored slide-captioning prompt to generate full-sentence descriptions that preserved technical terminology. Captions from individual slides were concatenated and integrated with the audio transcript, producing a comprehensive, sequentially accurate seminar transcript.
Literature Retrieval
Relevant literature was obtained via an automated agent querying the PubMed E-utilities API using the search terms “Boon Lead Tee AND PPA” and “Maria Luisa Gorno-Tempini AND PPA”. Up to 100 records, comprising titles, abstracts, and PMIDs, were retrieved and stored in JSON format to facilitate downstream processes.
Draft Generation
A manuscript draft was produced by an agent ingesting the unified transcript, leveraging the o1-2024-12-17 language model. Instructions specified generating a coherent first draft while maintaining scientific accuracy and the original narrative voice, explicitly excluding any content absent from the original transcripts.
Citation Scaffolding and Review
Initial citations were scaffolded by employing GPT-4o-2024-08-06 to insert “［REF］” tokens immediately after statements necessitating citation. The manuscript then underwent three iterative Review―Improve cycles. Each cycle involved structured feedback on narrative coherence, factual accuracy, and completeness of citations, provided by a review agent. Feedback was addressed by an improvement agent utilizing the o1-2024-12-17 model, ensuring modifications were restricted to structural and stylistic enhancements, aligning strictly with reviewer recommendations.
Reference Resolution and Ranking
After iterative refinements, citation placeholders were algorithmically resolved by matching each “［REF］” placeholder to the most semantically relevant publication from the retrieved PubMed literature. This involved calculating semantic similarity scores between manuscript statements requiring citations and the PubMed abstracts, selecting the best match for each placeholder. Citations were then formatted to include author, year, and PMID details. Additionally, pivotal statements referencing Dr. Gorno-Tempini’s contributions were automatically identified and linked to the three most relevant APA-formatted citations, structured within a standardized JSON schema.
Outputs and Computational Resources
The automated pipeline generated a formatted, citation-rich manuscript, a ranked references JSON file, and comprehensive intermediate outputs for enhanced reproducibility. Processing was executed on an 8-vCPU virtual machine, requiring approximately 14 minutes per complete run （mean±45 s；n＝3）. Audio transcription was the primary computational bottleneck （～6 minutes）, with Review―Improve cycles collectively taking ～5 minutes；all other agent-based tasks individually required less than 60 seconds.
This modular, multi-agent methodology enabled a systematic, transparent, and auditable transformation of seminar materials into a manuscript suitable for author revision and peer review, with each stage―from initial content ingestion to drafting, review, and citation management―clearly delineated. Throughout the entire process, human experts provided critical oversight, reviewing and refining outputs at each stage, with particular attention to manuscript revision, reference assignment and verification of scientific content.
