Exploring the performance of a novel machine learning classifier for minimal-invasive CNS lymphoma diagnosis through ultrasensitive profiling of circulating tumor DNA from cerebrospinal fluid and blood plasma – a prospective oligo-center trial (DETECT_CNS

中枢神经系统淋巴瘤(CNSL)的诊断需要侵入性神经外科手术，在某些高危情况下(例如，在老年/虚弱患者或当病变位于有能力的脑结构时)通常不能安全地执行，或者由于同时进行皮质类固醇或抗血小板治疗而被推迟。传统的脑脊液(CSF)细胞病理学或流式细胞术分析和诊断性MRI的灵敏度和鉴别能力不佳，不能进行非手术的CNSL诊断。因此，克服这些局限性并允许对CNSL进行可靠的微创识别的改进方法将对这些患者的临床护理产生变革。我们已经建立了一种新的机器学习方法，可以从通过超灵敏的下一代脑脊液或血浆循环肿瘤DNA(CtDNA)测序(NGS)所描绘的突变环境中进行稳健的CNSL识别。在训练验证方法中，我们已经证明我们的分类模型正确地从CSF-ctDNA中识别了60%的CNSL病例，显示出100%的特异性和阳性预测值(PPV)。然而，在这种方法可以用于临床常规之前，这项回顾性研究仍有几个局限性需要克服。首先，机器学习分类器需要在更大的患者队列和真实世界条件下进行进一步的前瞻性测试，使样本收集、处理和样本量标准化。此外，该分类器的PPV和特异性仅基于16名非CNSL患者进行评估；因此，需要更多的非淋巴瘤患者包括更广泛的恶性和非恶性实体来确认其性能，以进行稳健的CNSL识别。因此，我们在这里提出了一项前瞻性、诊断性、非随机、少中心的临床试验，旨在探索和验证我们用于正确和稳健识别CNSL的新型微创分类器(DETECT_CNSL)的性能。我们将招募120名患者(36名CNSL患者，84名非CNSL患者)，他们有一种新的脑部病变和立体定向活检的指征，CNSL是一种鉴别诊断。这些患者将接受腰椎穿刺术和抽血，然后进行神经外科活检，通过靶向NGS从脑脊液和血浆中提取ctDNA的基因图谱。使用我们的新型机器学习算法，我们将把CNSL和非CNSL的样本进行分类，并将我们的结果与金标准组织病理学进行比较，最终目标是正确地从CSF-ctDNA中对CNSL进行80%的分类，假设特异性和PPV为100%。如果成功，Detect_CNSL可能会在高危情况下或在诊断延迟的情况下对疑似CNSL患者的临床管理产生实践改变的影响，并可能为未来在拟议方案中测试这一应用的干预试验提供信息。

The diagnosis of CNS lymphoma (CNSL) requires invasive neurosurgical procedures that often cannot be safely performed in certain high-risk situations (e.g., in elderly/frail patients or when lesions are located in eloquent brain structures) or are delayed due to concurrent corticosteroid or anti-platelet therapies. Conventional analysis of cerebrospinal fluid (CSF) by cytopathology or flow cytometry and diagnostic MRI have demonstrated suboptimal sensitivity and discriminative capacity to allow surgery-free CNSL diagnosis. Therefore, improved methods that overcome these limitations and allow reliable minimal-invasive identification of CNSL would be transformative for the clinical care of these patients. We have established a novel machine learning approach that allows robust CNSL identification from mutational landscapes profiled by ultrasensitive next-generation sequencing (NGS) of circulating tumor DNA (ctDNA) from CSF or blood plasma. In a training-validation approach, we have demonstrated that our classification model correctly identified CNSL in 60% of cases from CSF-ctDNA, showing a specificity and positive predictive value (PPV) of 100%. However, this retrospective study harbors several limitations that remain to be overcome before such an approach can be used in clinical routine. First, the machine learning classifier requires further prospective testing on larger patient cohorts and under real-world conditions, standardizing sample collection, processing, and sample volumes. In addition, PPV and specificity of the classifier was assessed based on only 16 Non-CNSL patients; thus, higher numbers of non-lymphoma patients comprising a wider range of malignant and non-malignant entities are needed to confirm its performance for robust CNSL identification. Therefore, we here propose a prospective, diagnostic, non-randomized, oligo-center clinical trial that aims to explore and validate the performance of our novel minimal-invasive classifier for correct and robust CNSL identification (DETECT_CNSL). We will enroll 120 patients (36 CNSL patients, 84 Non-CNSL patients) with a novel brain lesion and indication for stereotactic biopsy, with CNSL being a differential diagnosis. These patients will undergo lumbar puncture and blood draw before neurosurgical biopsy to genetically profile ctDNA from CSF and blood plasma by targeted NGS. Using our novel machine learning algorithm, we will then classify samples as CNSL vs. Non-CNSL and compare our results to the gold standard histopathology, with the ultimate goal to correctly classify 80% of CNSL from CSF-ctDNA, assuming a specificity and PPV of 100%. If successful, DETECT_CNSL could chave practice-changing impact on the clinical management of patients with suspected CNSL in high-risk situations or when the diagnosis is delayed, and could inform future interventional trials testing this appraoch in the proposed scenarios.