BostonGene Corporation, a Waltham, Massachusetts-based AI oncology company, presented new clinical data at the United States and Canadian Academy of Pathology 115th Annual Meeting this week showing that its multimodal AI framework, trained on approximately 20,000 tumour profiles, accurately identified tumour origin and uncovered actionable therapeutic targets in more than 65 percent of real-world cancer of unknown primary cases. The study, delivered as an oral presentation by Juan Miguel Mosquera, MD, MSc, of Weill Cornell Medicine, integrated whole exome sequencing and transcriptomic data to address one of the most diagnostically resistant categories in oncology. The findings signal a broadening application of BostonGene’s foundation model beyond traditional companion diagnostics toward AI-driven disease modelling at the clinical frontier.
Why CUP remains one of oncology’s most intractable diagnostic problems
Cancer of unknown primary is not a rare edge case. It has historically accounted for roughly one to two percent of all malignancies and is characterised by an almost uniformly poor prognosis. Average survival following diagnosis sits between nine and twelve months across all subtypes, and for the majority of patients falling into the unfavourable category, which represents approximately 80 percent of CUP presentations, empiric chemotherapy has remained the default despite decades of inadequate results. The absence of a confirmed primary site creates a cascading problem: patients are excluded from tissue-specific treatment guidelines, biomarker-gated therapies, and most precision medicine pathways that have transformed outcomes in defined tumour types. The case for better diagnostic resolution is not theoretical. It is arithmetical.
Even with advances in gene expression profiling and next-generation sequencing, the field has struggled to close the gap between identifying a likely tissue of origin and meaningfully improving survival through site-specific therapy. Two landmark randomised trials conducted before the widespread adoption of immunotherapy and molecular-guided therapy failed to demonstrate survival benefit from site-directed treatment over empiric chemotherapy, a finding that dampened enthusiasm for tissue-of-origin classifiers for years. More recent evidence, including randomised data from the GEFCAPI consortium, has begun to shift that calculus, suggesting that improved precision therapies guided by molecular diagnosis can yield better outcomes. BostonGene’s USCAP presentation enters a field that is actively revising its assumptions about what diagnostic resolution can and cannot deliver.
What the BostonGene framework does differently from existing tissue-of-origin tools
Most tissue-of-origin classifiers currently in clinical use rely on a single molecular layer, typically gene expression profiling or genomic sequencing, to infer the primary site. The BostonGene framework described at USCAP integrates whole exome sequencing with transcriptomic data, creating a combined genomic and gene expression profile for each tumour. This is not a novel idea in principle, but doing it at scale in real-world CUP cases, where tissue quality is often compromised and clinical history is ambiguous, is operationally different from laboratory-controlled analyses. The Weill Cornell Medicine collaboration adds institutional credibility to the real-world validation claim, though the presentation is a conference abstract rather than a peer-reviewed publication, and independent replication across different testing environments has not yet been reported.
The distinction between identifying a tumour’s origin and identifying actionable therapeutic targets is also analytically important. The BostonGene approach does not stop at classification. It generates treatment-relevant molecular signals in parallel, which is where the 65 percent figure derives its clinical weight. Industry observers note that classifiers which provide origin predictions without therapeutic guidance have limited utility in the absence of site-specific approvals, particularly for patients whose presumed primary falls into categories with restricted treatment access. The combined genomic-transcriptomic approach is designed to surface both the origin inference and the molecular rationale for therapy selection in the same analytical pass.
How a 20,000-tumour training dataset shapes clinical confidence in AI-driven classification
Training a multimodal AI framework on approximately 20,000 tumour profiles is a significant dataset by the standards of most oncology machine learning studies, but the composition of that cohort and how closely it reflects the heterogeneity of real-world CUP presentations matters considerably. CUP is not a single disease entity. It is a constellation of histologically and molecularly distinct tumours that share the common feature of diagnostic opacity. Adenocarcinomas represent the majority of presentations, but squamous cell carcinomas, poorly differentiated neoplasms, and undifferentiated tumours each behave differently and carry distinct molecular landscapes. Whether the training data is proportionally representative of this heterogeneity, and whether the model performs consistently across rare subtypes rather than only on the most common presentations, are questions the published abstract does not fully address.
The training sample size also raises questions about how accuracy was measured and over what patient population. Real-world accuracy in CUP is notoriously difficult to benchmark because there is often no confirmed ground truth for the primary site. Validation methodologies that use known-primary cases as surrogates for CUP are common but imperfect, since patients with a confirmed primary may have a molecularly cleaner tumour profile than a genuine CUP case. Clinicians tracking the field are likely to scrutinise the validation design closely before the framework can be positioned as a clinical standard rather than a research-stage tool.
Actionable targets in more than 65 percent of patients: what this figure actually means
The headline figure from the USCAP presentation, actionable targets identified in more than 65 percent of patients including FDA-approved options, requires careful contextualisation. The presence of a genomic alteration with a matched approved therapy does not guarantee clinical benefit. Tumour mutational burden, immune microenvironment characteristics, and co-occurring genomic events all influence whether a patient is likely to respond to a nominally matched agent. In CUP specifically, the available evidence linking molecular matching scores to survival improvement remains encouraging but contested. Data from earlier NGS-guided CUP studies has shown that patients with higher matching scores between their genomic alterations and available targeted therapies achieve significantly longer progression-free survival than those with lower scores, but overall survival differences have not consistently reached statistical significance.
What the 65 percent figure does accomplish clinically is frame the access problem. In current practice, without molecular profiling and tissue-of-origin resolution, a substantial proportion of CUP patients receive empiric platinum-based chemotherapy with limited actionable information guiding the choice. A framework that surfaces FDA-approved therapeutic options for the majority of patients it evaluates changes the clinical conversation from what we can try empirically to what the biology supports. The extent to which this changes outcomes depends heavily on whether treating clinicians act on the molecular data and whether payers support the pathway to access those therapies.
The gap between origin identification and survival improvement in CUP
The history of tissue-of-origin testing in CUP is instructive. Earlier gene expression profiling platforms that entered clinical use with strong classification accuracy data ran into a fundamental obstacle: demonstrating that patients treated with site-specific therapy based on predicted origin actually lived longer than those who received empiric chemotherapy. The GEFCAPI 04 trial, which compared empiric treatment against gene expression profiling-guided therapy, did not demonstrate a significant improvement in one-year overall survival in its primary analysis, even as secondary endpoints and subgroup analyses showed signals of benefit in some patient populations. This history creates a structural challenge for any new CUP diagnostic tool. Classification accuracy is a necessary but insufficient condition for clinical validation. The question the field is waiting to have answered is whether improved diagnostic resolution, delivered through a more sophisticated multimodal framework, translates into a survival signal robust enough to change standard-of-care guidance.
BostonGene’s USCAP data does not answer that question. It demonstrates real-world accuracy in origin identification and actionable target discovery, which are meaningful first steps, but the company has not presented randomised survival data from a CUP-specific therapeutic intervention study. Regulatory watchers suggest that without prospective outcomes data, integration into national guidelines will be slow regardless of the technical sophistication of the underlying platform.
What AI disease modelling changes about patient stratification in routine pathology
The broader conceptual shift BostonGene is articulating at USCAP, from traditional tumour classification toward what it describes as AI-driven disease modelling, has implications beyond CUP diagnostics. Traditional pathology classifications are largely morphology-based and have remained structurally stable for decades. The integration of genomic, transcriptomic, and immune data into a unified analytical framework creates a classification space that is richer, more dynamic, and capable of capturing biological relationships that conventional histology cannot resolve. For pathologists, this is both an opportunity and a workflow challenge. The interpretive layer required to translate AI-generated tumour portraits into clinical decisions requires new competencies and new institutional infrastructure that most pathology departments are still in the early stages of building.
At USCAP specifically, the framing of AI as a tool for deepening biological understanding rather than replacing diagnostic judgment reflects a deliberate positioning choice. The conference is primarily a pathologist audience, and presentations that frame AI as augmenting expert interpretation rather than competing with it typically land differently than those that emphasise automation. The collaboration with Weill Cornell Medicine serves a similar function: academic partnership signals scientific rigour and institutional validation in a way that internal company data alone does not. Industry observers note that the decision to present at USCAP, rather than a purely clinical oncology venue, also reflects BostonGene’s strategy of embedding its platform within the pathology workflow as a point-of-care analytical tool rather than a separate specialty service.
Unresolved questions around validation, reimbursement, and real-world scalability
Several structural challenges stand between the USCAP findings and routine clinical integration. Reimbursement for multimodal genomic-transcriptomic testing in CUP remains inconsistent across payers and geographies. Current coding frameworks for complex molecular diagnostics do not cleanly accommodate combined whole exome and transcriptomic profiling as a single clinical product, creating billing uncertainty that affects laboratory adoption even when the clinical case is strong. The turnaround time and per-sample cost of the integrated framework have not been publicly disclosed, but both will influence whether community oncology and community pathology settings, which handle a large proportion of CUP diagnoses, can practically implement the workflow.
There are also unresolved questions about clinical decision support integration. A multimodal tumour portrait is analytically rich but requires structured clinical decision support to translate into actionable treatment guidance at the point of care. BostonGene’s existing Tumor Portrait product is designed to serve this function, but the degree to which its outputs are integrated into electronic health record workflows at scale, as opposed to being delivered as standalone reports, is unclear from publicly available information. Scalability also depends on tissue quality and quantity. CUP cases often present with limited biopsy material obtained under suboptimal conditions, and the minimum tissue requirements for the combined whole exome and transcriptomic assay may limit applicability in exactly the cases where diagnostic resolution is most needed.
BostonGene’s broader platform strategy and what USCAP signals about commercial direction
The USCAP presentation is one of several high-profile scientific appearances BostonGene has made in early 2026. The company presented at the Precision Medicine World Conference in March, announced a strategic collaboration with AstraZeneca in January to apply its foundation model to early clinical trial design, and entered a separate collaboration with Daiichi Sankyo in February focused on antibody-drug conjugate development. Taken together, these moves sketch a platform strategy that spans CUP diagnostics, drug development partnership, and clinical trial optimisation, with the foundation model positioned as the connective technology across all three use cases.
For the pathology and oncology communities watching BostonGene’s conference activity, the USCAP oral presentation carries specific significance. CUP is a condition where the clinical need for better diagnostics is unambiguous and the competitive landscape for AI-driven solutions is thinning rather than crowding. Establishing clinical credibility through Weill Cornell Medicine collaboration and presenting real-world performance data in a forum dominated by pathologists positions the company’s Unknown Primary test as a serious contender for guideline consideration, provided it can generate the prospective clinical outcomes data that the field has historically demanded before changing standard-of-care practice. The USCAP findings are an important proof point. They are not yet the proof.
BostonGene AI framework identifies actionable targets in 65% of cancer of unknown primary cases at USCAP 2026 added by Pallavi Madhiraju on
View all posts by Pallavi Madhiraju →