🔶 基于智能体的SOTA全基因组解读：间质性肺病案例研究

过去几年里，越来越多的证据表明，在遗传病高度富集的新生儿重症监护室（NICU）中推广全基因组测序能够改善临床和经济结局。然而，测序仍主要集中在顶级学术中心，原因在于：

• 开具检测与解读结果的复杂性
• 尽管几乎所有的支付方都覆盖报销，但申请报销的难度和费用高昂
• 获取结果或进行再分析所需的时间

Gamow Labs 的假设是，正确驾驭的 AI 智能体能够以人类水平或更高的水平完成遗传病分子诊断任务，且成本仅为极少部分、时间大幅缩短，并能实现大规模推广。本文中，我们展示了支持这一假设的早期证据。

通过与贝勒医学院 Pawel Stankiewicz 实验室进行的一项盲法合作，我们获得了 46 名个体的原始测序读段（FASTQ 文件），这些个体包括死于间质性肺病的婴儿及其健康家庭成员。在所有病例中，一线临床基因组学实验室均未能给出最初的分子诊断，因此升级至该领域顶尖专家 Stankiewicz 博士进行重新分析。

我们将 Gamow Labs 系统 George-0.1 与原始 FASTQ 文件，以及 Stankiewicz 博士提供的一个简单表型描述一同输入：“患者患有致命性肺疾病，如肺泡毛细血管发育不良、腺泡发育不良或先天性肺泡发育不良”（此描述对健康家庭成员显然是错误的，但我们当时并不知道这些成员被包括在内）。该智能体按照 ACMG 指南生成了分子诊断。

随后我们将结果交由 Stankiewicz 博士评分。真阳性定义为变异位点与其答案一致且评分为致病性或可能致病性。真阴性定义为未给出诊断或评分为意义不明确的变异（VUS）（该系统针对高召回率进行调校，因此在呈现疾病表型时对健康人报告一些 VUS 并不意外）。

结果令人鼓舞。除健康家庭成员外，所有病例均被一线临床实验室误诊（这也是它们被升级至 Stankiewicz 博士的原因）。Stankiewicz 博士的实验室随后解决了 26 例中的 19 例，仍有 7 例未能确诊。Gamow Labs 的 George-0.1 达到了与 Stankiewicz 博士实验室同等的水平，成功复现了全部 19 例的分子诊断，并额外贡献了 2 个新诊断。这些初步结果汇总于表 1，表明像 George-0.1 这样的智能体 AI 能够超越临床全基因组测序实验室的表现，并匹敌甚至超越人类专家。

此外，我们利用这一数据集独立验证了波士顿儿童医院与 OpenAI 近期发表在 NEJM AI 上的论文。他们展示了消费者聊天机器人——本研究中使用的是 ChatGPT 5.5 Pro，他们的研究中是 o3 Deep Research——在罕见病诊断方面具有显著的实用性。我们观察到了相同的现象。尽管 ChatGPT 5.5 Pro 的表现落后于 Gamow Labs 的 George-0.1，但在这些挑战性病例上，上传表型和基因组至 ChatGPT 5.5 Pro 的表现优于临床实验室。

我们还尝试将实验扩展到 Claude.ai（Opus 4.8）和 Gemini.google.com（Gemini 3.5 Flash）这两个流行的消费者工具，但两个系统在前三次尝试中均未成功。

表 1：三种方法在全基因组解读中的性能总结。这些病例均涉及复杂的结构变异或内含子区变异，最初被多家临床实验室漏检。ChatGPT 5.5 Pro 相较临床工作流显著提高了诊断率，而 Gamow Labs George-0.1 则进一步提升了诊断率。

• 传递给聊天机器人的是预先检出的 vcf 文件，而 Gamow Labs George-0.1 处理的是原始测序读段，因此前者的任务更简单
  ** 5 例未知病例中有 4 例得到了分子诊断，但明显错误，原因在案例研究 #2 中讨论

案例研究 #1：临床实验室假阴性

ACD209.3，发表于 2022 年（此处仅讨论已发表案例），最初由 Rady Children’s Institute for Genomic Medicine 进行测序。比对和变异检出使用的是标准 Illumina DRAGEN 管线，并结合了一些开源 CNV/SV 检出工具。由于工具组合会生成大量变异，因此应用了 CCDS 过滤器，以去除距目标编码区超过 1 千碱基的变异。然而，这一决定导致漏掉了真正的分子诊断，因为受累的 FOXF1 增强子位于受影响的基因上游 1 兆碱基处。

这类错误并不少见。Deciphering Developmental Disorders (DDD) 研究通过评估 13,462 名先证者的外显子组数据中的 ClinVar 致病/可能致病变异，直接探讨了这个问题。在该队列检出的 ClinVar 致病变异中，83.9%（1,134/1,352）被标准自动过滤流程成功标记，意味着 16.1%（218/1,352）因后果、遗传模式或其他自动过滤器而被滤除。在对这些被滤除的变异进行临床复核后，发现有 112 个变异（涉及 107 名先证者，占队列的 0.8%）被确定为遗漏的潜在诊断。

由于 AI 无需对变异进行预过滤，这类系统可以完全避免此问题。Gamow Labs 和 ChatGPT 5.5 Pro 均成功解决了该病例。

案例研究 #2：ChatGPT 5.5 Pro 假阳性

假阳性是 ChatGPT 5.5 Pro 失败占比最高的类别。虽然可以争辩说提供一个错误的表型是不公平的，但两个系统接收到的错误表型是相同的，并且在临床环境中，由于初诊错误或病情演变，这种情况十分常见。当临床医生时间有限时，过度诊断可能带来问题。

这种表现形态多样，但根源在于未经证据支持或未考虑全部信息就过早下结论。例如，ACD130.3（2016 年发表）的一位健康家庭成员被自信地诊断为一个致病性的 308 碱基对缺失，位于导致许多 ACD 病例的 FOXF1 调控区域。该检出是真实存在的，且位于可疑区域，但此特定缺失并非致病性，因为人群频率高达 40%，而 ChatGPT 5.5 Pro 并未核查这一信息。

每个假阳性都是独一无二的，但错误主要可归为三类：

• 在基因与疾病之间建立过于牵强的联系
• 忽略可排除变异致病性的数据，如人群频率
• 以微妙的方式幻觉出变异检出的错误，例如错误地应用坐标系统或误解转录本

案例研究 #3：ChatGPT 5.5 Pro 假阴性

ChatGPT 5.5 Pro 在未发表病例中的错误分布与假阳性类似，但两个已发表病例的错误可归咎于简单失误。

在 ACD144.3 病例中（2016 年发表），ChatGPT 5.5 Pro 的结论是“没有高置信度的遗传根本原因”，并解释了为何无法给出诊断。我对此有些困惑，因为该病例的解决方案是涉及 FOXF1 增强子的大片段缺失，与 ChatGPT 5.5 Pro 之前正确解决的病例类似。

当 Gamow Labs George-0.1 提出可信的假设后，我返回相同的 ChatGPT 5.5 Pro 对话并询问其正确性。ChatGPT 5.5 Pro 回应此诊断是正确的，之前的错误是由于未能将已发表的 hg19 坐标转换为上传的 hg38 vcf 文件坐标所致。

这是一个微不足道且意想不到的错误。尽管这些模型具备令人印象深刻的智能，但在未得到最佳驾驭时，这类错误相对普遍。这正是 Gamow Labs 使命的重要驱动力。

如果这项工作令你感兴趣，欢迎加入 Gamow Labs！

如果你是一位对生物学充满热情的顶尖 AI 工程师，或者是一位坚信 AGI 的计算生物学家，并且对此项工作感兴趣，请与我们联系！我们正在寻找杰出人才加入团队。

Over the past few years, overwhelming evidence has emerged that expanding access to whole genome sequencing improves clinical and economic outcomes in the NICU, a setting highly enriched for genetic disease. However, sequencing is still largely limited to top academic centers due to the:

• complexity of ordering and interpreting results
• expense and difficulty filing for reimbursement despite near universal payer coverage
• time required to receive results or reanalysis

We at Gamow Labs hypothesize that correctly harnessed AI agents can perform the task of diagnosing a genetic disease at or above human performance at a fraction of the cost and in a fraction of the time at great scale. In this blog post, we present early evidence that this hypothesis is true.

Through a blinded collaboration with Pawel Stankiewicz at Baylor College of Medicine, we accessed the raw reads (FASTQ files) of 46 individuals, who were either infants lost to interstitial lung disease or their healthy family members. In all cases, the initial molecular diagnosis was missed by the first-line clinical genomics lab and escalated to Dr. Stankiewicz, the leading expert on the genetic causes of these diseases, for reanalysis.

We provided the Gamow Labs system, George-0.1, with the raw FASTQ files alongside a simple phenotype provided by Dr. Stankiewicz: “the patient suffered from a lethal lung disorder like alveolar capillary dysplasia, acinar dysplasia, or congenital alveolar dysplasia” (this phenotype was obviously wrong for the healthy family members, but we had no idea that these were included). The agent generated a molecular diagnosis classified using the ACMG guidelines.

We then shared the results with Dr. Stankiewicz for scoring. A true positive would be marked as correct if the variant matched his answer key with pathogenic or likely pathogenic scoring. A true negative would be marked correct if there were either no diagnosis or variant of uncertain significance (VUS) (the system is tuned for recall, so it is not surprising to find some VUSes called for healthy people when presented with a diseased phenotype).

The results were encouraging. Every case, with the exception of the healthy family members, was incorrectly diagnosed by the first-line clinical labs (this is why they were escalated to Dr. Stankiewicz). Dr. Stankiewicz’s lab subsequently solved 19/26, leaving 7 as mysteries. Gamow Labs’ George-0.1 matched the performance of Dr. Stankiewicz’s lab, reproducing the molecular diagnosis of all 19 cases while contributing 2 additional solutions. These initial results, summarized in Table 1, suggest that agentic AI like George-0.1 can exceed the performance of clinical whole genome sequencing labs and match or even exceed specialized human experts.

Furthermore, we independently validated the recent NEJM AI paper authored by Boston Children’s and OpenAI on this dataset. They showed that consumer chatbots–ChatGPT 5.5 Pro in our study and o3 Deep Research in theirs–have significant utility in rare disease diagnosis. We observed the same. While performance lagged Gamow Labs’ George-0.1, uploading a phenotype and genome into ChatGPT 5.5 Pro outperforms clinical labs on these challenging cases.

We attempted to extend the experiment to Claude.ai (Opus 4.8) and Gemini.google.com (Gemini 3.5 Flash), two other popular consumer tools, but neither system succeeded in the first three attempts.

| Total | Standard of care (clinical lab) correct | Gamow Labs correct | ChatGPT 5.5 Pro correct*
---|---|---|---|---
Healthy family members | 20 | N/A | 20 (100%) | 6 (30%)
Published known diagnosis | 11 | 0 (0%) | 11 (100%) | 9 (82%)
Unpublished known diagnosis | 8 | 0 (0%) | 8 (100%) | 5 (63%)
Previously unknown diagnosis | 2 | 0 (0%) | 2 (100%) | 2 (100%)
Still unknown diagnosis | 5 | 0 (0%) | N/A | <=1 (<=20%)**

Table 1: summary of performance of three approaches to genome interpretation. These are all hard cases involving complex structural or intronic variants that were initially missed by various clinical labs. ChatGPT 5.5 Pro significantly improved diagnostic yield over clinical workflow and Gamow Labs’ George-0.1 improved further.

• pre-called vcf files passed to chatbot vs. raw reads to Gamow Labs’ George-0.1, making the task easier
  ** 4/5 unknown cases received molecular diagnoses that were obviously wrong due to the reasons discussed in Case Study #2

Case study #1: Clinical lab false negative

ACD209.3, published in 2022 (only published cases discussed here), was originally sequenced by Rady Children’s Institute for Genomic Medicine. Alignment and variant calling were performed using a standard Illumina DRAGEN pipeline in addition to a combination of open-source CNV/SV callers. Because this ensemble of callers generates a significant number of variants, a CCDS filter was applied to remove variants >1 kilobase from the coding regions of interest. However, this decision resulted in discarding the molecular diagnosis because the impacted FOXF1 enhancer is one megabase upstream of the affected gene.

This type of error is not uncommon. The Deciphering Developmental Disorders (DDD) study directly addressed this question by evaluating ClinVar pathogenic/likely pathogenic variants in exome data from 13,462 probands. Of ClinVar pathogenic variants identified in the cohort, 83.9% (1,134/1,352) were successfully flagged by standard automated filtering pipelines, meaning 16.1% (218/1,352) were filtered out by consequence, inheritance, or other automated filters. After clinical review of these filtered-out variants, 112 variants in 107 probands (0.8% of the cohort) were identified as potential diagnoses that had been missed.

Because there is no need to pre-filter variants for an AI, this problem can entirely be avoided with these types of systems. Both Gamow Labs and ChatGPT 5.5 Pro solve this case.

Case study #2: ChatGPT 5.5 Pro false positive

False positives were the largest cohort of ChatGPT 5.5 Pro failures. While it can be argued that it is unfair to provide an incorrect phenotype, the same incorrect phenotype was provided to both systems and, in a clinical setting, this can be common both due to incorrect initial assessment and evolving conditions. When clinician time is limited, overdiagnosing can be problematic.

This behavior came in many flavors, but the root cause was jumping to conclusions not supported by evidence or considering all information. For example, one of the healthy family members of ACD130.3, published in 2016, was confidently diagnosed with a pathogenic 308 base pair deletion in the FOXF1 regulatory region responsible for many ACD cases. The call is real and in a suspicious region, but this particular deletion is not pathogenic, as evidenced by the 40% population frequency, which ChatGPT 5.5 Pro did not bother to check.

Each false positive was unique, but the errors largely fell into three buckets:

• Drawing overly creative links between gene and disease
• Ignoring data, e.g. population frequency, that could disqualify a variant as pathogenic
• Hallucinating calls in subtle ways, e.g. incorrectly applying wrong coordinate systems or misunderstanding transcripts

Case study #3: ChatGPT 5.5 Pro false negative

The distribution of ChatGPT 5.5 Pro’s errors over the unpublished cases was similar to the false positives, but the two published cases could be chalked up to simple errors.

In ACD144.3, published in 2016, ChatGPT 5.5 Pro concluded that “there is no high-confidence genetic root cause” and explained why none of the calls could be diagnostic. I was a bit puzzled by this because the solution is a large deletion across the FOXF1 enhancer similar to cases ChatGPT 5.5 Pro solved correctly.

Once the Gamow Labs’ George-0.1 generated a plausible hypothesis, I returned to the same ChatGPT 5.5 Pro session and asked if it was correct. ChatGPT 5.5 Pro responded that it was and its previous error was due to not lifting over the published hg19 coordinates for the hg38 vcf file uploaded.

This is a trivial and unexpected error. Despite the impressive intelligence these models possess, these errors are relatively common when not harnessed optimally. This is a large motivation of the Gamow Labs mission.

If this work interests you, join Gamow Labs!

If you are a cracked AI engineer passionate about biology or an AGI-pilled computational biologist and this work seems interesting, please reach out! We are looking for exceptional people to join the team.