How do you design AI for drug discovery?一款 B 端 AI 药物研发软件， 怎么设计？
Signals AI 帮科学家查实验、筛候选、对照数据。我负责 AI 对话与结果体验，让每次提问都能接上正在做的研究。Signals AI helps scientists find experiments, filter candidates and compare data. I designed the conversations and results that connect each question to their research.
43 个候选找到了 科学家能放心往下走吗？43 candidates found Can the scientist move forward?
Signals AI 能从实验里整理出一份名单。接下来，科学家还要决定查什么、验证什么。我负责从拿到答案到继续研究的这段体验。Signals AI can assemble a shortlist from an experiment. Scientists still need to decide what to inspect and test. My work focuses on the experience between receiving an answer and continuing the research.
他们在实验室做研究，也在电脑前查记录、读结果。我为药企的科学家和研发人员设计，让 Signals AI 接上他们正在做的工作。They run experiments at the bench, then return to a screen to inspect records and results. I design for pharmaceutical scientists and R&D teams, bringing Signals AI into the work already in progress.
01在实验台前，验证一个想法。At the bench, test an idea.02回到屏幕前，判断下一步。At the screen, decide what comes next.
Signals 的客户SIGNALS CUSTOMERS
他们来自这些药企The companies behind the researchers
辉瑞Pfizer
默沙东Merck
葛兰素史克GSK
罗氏Roche
诺华Novartis
阿斯利康AstraZeneca
拜耳Bayer
莫德纳Moderna
我怎样和团队合作HOW I WORK WITH THE TEAM
我的设计， 和三地团队一起完成My design work Built with a team across three locations
01
美国 · 产品经理UNITED STATES · PRODUCT MANAGER
Larry
客户调研与部分数据，从他这里进入设计讨论。Customer research and data inform the design discussion through Larry.
02
高级产品设计师SENIOR PRODUCT DESIGNER
Joey Zhao
我负责 AI 入口、筛选，以及对话和结果体验。I design AI entry points, filtering, conversations and results.
03
上海 · 美国SHANGHAI · UNITED STATES
研发工程师Engineering
与研发团队一起完成产品实现。Working with engineering to bring the product to life.
04
印度 · 质量测试INDIA · QUALITY ASSURANCE
QA
由印度团队负责质量测试。Quality assurance with our team in India.
01 / 行业观察 · 2026.1001 / INDUSTRY REVIEW · OCT 2026
同样叫 AI 药物研发， 要解决的问题差得很远AI for drug discovery Very different design problems
Insilico 和 Recursion 把 AI 用在候选药物与实验研究上。Benchling、Signals 则把 AI 放进科学家每天使用的软件。我的问题更具体，实验已经在这里，怎样让 AI 帮人接着做下去？Insilico and Recursion apply AI to drug candidates and experimental research. Benchling and Signals bring AI into scientists’ everyday software. My question is more specific. The experiment is already here. How can AI help someone continue the work?
01研发候选药物Develop drug candidates
02帮助团队完成研究Help teams do research
Insilico Medicine设计候选分子Design drug candidates
让 AI 提出值得测试的分子Propose molecules worth testing
Pharma.AI 帮助寻找研究目标和设计分子，再由实验与人体试验检验。Pharma.AI supports target discovery and molecule design, followed by experiments and human trials.
正进入更大规模的人体验证Moving into larger human trials
2025 年发表早期人体试验结果。公司于 2026 年 9 月披露，Rentosertib 已在中国进入三期临床，继续验证安全性与疗效。Early human results were published in 2025. In September 2026, the company reported entry into Phase 3 in China for further testing of safety and efficacy.
研究者需要比较候选方案，再决定做哪些实验。Scientists need to compare candidates before deciding what to test.
Recursion把实验接进 AIConnect AI and experiments
用实验数据，反过来指导研究Let experiments inform the next question
把自动化实验、细胞图像与 AI 分析连起来，研究疾病和潜在药物。Connect automated experiments, cell images and AI analysis to investigate disease and potential medicines.
已有多个临床研究项目Multiple clinical programs
官网列出的 REC-4881 处于二期临床，即在患者中继续检验安全性与作用。进入临床仍需要后续验证。Its pipeline lists REC-4881 in Phase 2, testing safety and activity in patients. Clinical entry still requires further validation.
设计需要连接实验与计算结果，让每轮研究能接着上一轮继续。Design must connect experiments and computational results across successive research cycles.
Benchling科研工作助手Assist scientific work
让 AI 进入科学家每天的工作Put AI into the scientist’s daily work
在实验记录里查历史、分析数据、写报告，也能调用科学模型。Find prior work, analyze data, write reports and run scientific models from the notebook.
从回答问题，走向执行任务From answers to actions
2026 年 8 月发布了能执行任务的 AI 助手，可登记分子、创建实验记录、处理结果。其产品同样提供带来源的回答。Agents announced in August 2026 can register molecules, create notebook entries and process results. The product also provides source-linked answers.
查得到、带引用，已经是这类产品的共同方向。还要考虑如何接上用户的下一步。Search and citations are shared directions in this category. The next question is how they connect to the user’s next action.
Revvity Signals我参与的产品THE PRODUCT I WORKED ON
把回答接回正在做的实验Connect the answer to the experiment
AI Explorer 可以围绕实验查找、筛选、分析。Signals One 同时承载实验记录、化学绘图与数据分析。AI Explorer supports search, filtering and analysis around experiments. Signals One also brings together records, chemical drawing and data analysis.
已有工作平台上的 AIAI within an existing research platform
产品官网已展示 AI Explorer；本案例进一步用真实数据检查“提问、筛选、回到记录”这段流程。The product site demonstrates AI Explorer. This case checks the question, filter and return-to-record flow with real data.
我的重点是入口、筛选依据和回查路径，让平台已有的能力能在一次任务里用起来。My focus is the entry point, filtering criteria and route back to evidence, making the platform’s capabilities usable in one task.
SIGNALS 的产品基础WHAT SIGNALS CAN BUILD ON
实验和工具已经在这里 AI 怎样接进去？The work is already here How should AI fit in?
对已经在 Signals 里工作的团队，AI 可以沿用已有实验，并连接 ChemDraw 绘图和 Spotfire 分析。我的设计重点，是让提问、看结果和回查记录连得上。For teams already working in Signals, AI can draw on existing experiments alongside ChemDraw and Spotfire. My design focus is connecting the question, the result and the route back to the record.
本节为 2026 年 10 月的公开产品对照，按主要任务归类，各家公司业务有重叠。它补充作品集背景，不作为历史用户访谈或四款产品的同条件性能测试。设计启发是本案例的分析。This October 2026 public-product review groups companies by the tasks discussed; their businesses overlap. It adds portfolio context, not historical interviews or a controlled benchmark. Design implications are case-study analysis.
先给 Signals AI 一个具体问题Give Signals AI a concrete question
94 个分子，哪些满足筛选条件？这次用公开实验走一遍，从提问开始，一直查到原始数据。Which of these 94 molecules meet the criteria? This public-data review follows the task from the first question all the way back to the source.
⬡
分子Molecule
这次被测试的化合物，还不能叫药。A compound being tested. It is not yet a medicine.
◎
靶点Target
希望影响的对象。这次是名为 CDK2 的蛋白质。The object of the experiment. Here, a protein called CDK2.
½
IC50
抑制一半活性所需的浓度。同一实验条件下，数值越小，所需浓度越低。The concentration needed for half-maximal inhibition. Under the same assay conditions, a lower value means less is needed.
nM（纳摩尔每升）是浓度单位。带“>”的记录只说“大于某个数”，没有准确值，先排除。The unit is nM (nanomoles per litre), a concentration. Values marked “>” give only a lower bound, so we exclude them.
Signals AI 完成了这次筛选。对照原始文件，43 条全部一致。Signals AI completed this selection. All 43 matched the original file.
继续追问，问题出现了THE FOLLOW-UP EXPOSES A PROBLEM
第一份名单全对 追问时，算错了The first shortlist was right The next answer miscounted
换一个阈值继续问，AI 把 78 条说成了 77 条。比较分子时，又补出了来源没有的实验条件。A follow-up with a different threshold returned 77 instead of 78. A compound comparison then inferred experimental conditions absent from the data.
继续问 IC50 < 100 nMFOLLOW-UP · IC50 < 100 nM
77→78
小于 100 nM 的精确值数量，复核后才改对。The count of exact values below 100 nM was corrected after checking.
答案看起来很完整， 人该从哪里发现问题？When an answer looks complete, where can a person check it?
Signals AI · 把数据画出来SIGNALS AI · SEE THE DISTRIBUTION
纠正之后， 把分布放回眼前After the correction, look at the distribution
图表里有 88 条精确值。小于 100 nM 的三个区间相加是 78，科学家可以顺着图继续核对。The chart includes 88 exact values. The three bins below 100 nM add up to 78, giving the scientist another way to check the answer.
88条精确值进入图表exact values plotted
Signals AI · 对照候选SIGNALS AI · COMPARE CANDIDATES
还有一个小符号， 不能被 AI 省掉One small symbol must survive the answer
“大于 1000”只告诉我们一个下限。把它当成“等于 1000”，比较就变了。我要让这个区别在结果里留得住。“Greater than 1000” gives only a lower bound. Treating it as “equal to 1000” changes the comparison. That distinction needs to remain visible in the result.
CHEMBL1964259
0.54nM
IC50 · 精确值Exact valueCHEMBL1964249
1.0nM
IC50 · 精确值Exact valueCHEMBL271778
> 1000nM
IC50 · 仅有下限Lower bound only
接下来还缺什么？What is still missing?细胞中的作用Cellular activity对其他靶点的影响Effects on other targets
这份实验没有这些数据，不能据此决定哪个分子更适合做药。This experiment contains neither, so these values alone cannot determine which molecule is a better drug candidate.
Signals AI
比较可以继续 未知也要留下Keep comparing Keep the unknowns visible
我参与的一次方向讨论A DIRECTION I ARGUED FOR
再做一个 AI 工具， 科学家就要重新开始吗？A new AI tool Should scientists have to start over?
最初，团队讨论过做一个全新的 AI Agent 工具。我更在意另一件事：客户已经在 Signals One 里积累了研究数据，这才是我们最有价值的基础。Early on, the team discussed building an entirely new AI agent tool. I kept coming back to what customers already had: research data accumulated in Signals One. To me, that was our strongest foundation.
我的建议是，让 Signals AI 围绕已有数据工作。科学家不用换个产品从头开始，设计也从“接上正在做的研究”出发。I argued for Signals AI to work with that existing data. Scientists should be able to continue their research without starting again in a new product. That was the direction I wanted the experience to take.
01 / Signals AI
让提问留在实验旁边Keep the question beside the experiment
科学家已经打开了实验，不该再解释一遍“我在看什么”。我把对话留在记录旁边；真正要接上的，是当前实验、应用和可用数据。The experiment is already open. Scientists should not have to explain it again. I keep the conversation beside the record; the AI also needs the current experiment, application and available data.
少一点工作区宽度，保留眼前的实验。Trade some workspace width for the experiment in view.
02 / Signals AI
先看懂条件，再相信结果Show the rule behind the result
“并且”被理解成“或者”，名单就会变。在这次复核里，我把单位、阈值和比较关系一起留在回答中，让科学家能检查 AI 到底按什么筛选。Read “and” as “or” and the shortlist changes. In this review, I keep the unit, threshold and comparison rule in the answer so scientists can check what the AI actually filtered.
多留一行条件，少一次猜测。One extra line makes the interpretation inspectable.
科学家最终要检查的是数据。引用直接接回实验，详细记录留在原处，对话不用越写越长。Scientists ultimately need to inspect the data. A citation returns to the experiment, keeping the conversation brief and the detailed record intact.
摘要留在对话，细节回到记录。A short answer with access to the full record.
03.1
引用，应该能打开A citation should open
从回答里的实验名称，回到 Signals 原始记录。Follow the experiment name in the answer back to its Signals record.
回到实验，决定下一步THE TASK CONTINUES IN THE RECORD
拿到名单以后， 带着依据继续研究Take the shortlist Continue with the evidence
43 条候选，有条件可查，有原始实验可回。细胞中的作用和对其他靶点的影响，仍要补实验。Signals AI 帮人整理到这里，下一步的判断留给科学家。43 candidates, with criteria to inspect and an experiment to return to. Cellular activity and effects on other targets still need further experiments. Signals AI assembles the evidence; the scientist decides what comes next.
01回答里找到来源Find the citation
02回到实验，继续检查Return to the experiment
Sources and review records数据来源与复核记录＋
2026 年公开数据复核，用于说明当前产品行为。来源为 ChEMBL 的 CDK2 实验，非历史用户研究。Public-data review recorded in 2026 to examine current behavior, separate from historical user research.查看来源View sourceIC50 术语来源IC50 reference
比较回答推断了来源没有给出的测试浓度范围，复核后纠正。图表附文把小于 100 nM 的数量算成 77，正确值为 78。完整问答保留在工作档案中。The comparison inferred a concentration range absent from the source. The chart summary also counted 77 values below 100 nM instead of 78. Both were corrected; the full conversation is retained.
材料登记要求数量与纯度，公开数据未提供。因此 94 条数据保存于实验与附件，未登记为实体材料批次。Registration required amount and purity absent from the source. The 94 activities remain in the experiment and attachment; no inventory batches were created.
Source attachment actions来源附件操作Verified return to the experiment已核实的实验跳转Saved activity records已保存的活性记录Source details来源详情Original attachment原始附件Public assay source公开实验来源实验与原始附件Experiment and source file筛选条件局部Filtering detail修正后的比较回答Corrected comparison复核后的分布图Verified distribution公开化合物结构Public compound structures
Signals AI · 产品后续进展SIGNALS AI · PRODUCT MILESTONE
后来，Signals AI 也走进了 ClaudeSignals AI now reaches scientists in Claude, too
Revvity 宣布接入 Anthropic 的 MCP 连接器生态。科学家可以在 Claude 里访问 Signals 中的实验和研究信息，继续查询与分析。Revvity announced its integration with Anthropic’s MCP connector ecosystem. Scientists can access experiments and research context from Signals while working in Claude.
MCP 可以理解为连接 AI 与科研系统的接口：换一个提问入口，仍能访问 Signals 中的研究信息。MCP connects an AI assistant to a scientific system: a different place to ask, with access to the research in Signals.
Josh Bond 确认了产品管理团队给予我的季度表彰与奖励，并肯定了我在 BioDesign、BioReg 和库存移动端的工作。Josh Bond confirmed a quarterly recognition and reward from product management, highlighting my work on BioDesign, BioReg and the Inventory Mobile app.
“谢谢你的创造力、 灵活应变和出色工作。”“Thank you, Joey, for your creativity, flexibility, and outstanding work!”
实验记录Records团队已有的研究Existing research
化学绘图Chemical drawingChemDraw
数据分析Data analysisSpotfire
