Biotech R&D · 案例研究
The three-year-old failure the search box couldn't find, because the compound got renamed
它能带来什么Across 6,200 lab-notebook entries, a question like "every experiment we've run on this target with this scaffold" comes back with the internal code name and the new name folded into one compound, and the reason each program stopped quoted from the note that says it.
A new scientist brings in a screening plan. One target, one scaffold, six weeks, ₩18M in reagents. The head of research types the compound name from the plan into the company search box. Zero results. Only after the meeting does he remember running experiments on that same scaffold three years ago. Back then the compound's internal name was CPD-0417, and it doesn't share a single character with the name on the plan.
This person runs research at a 45-person biotech. The average scientist stays 2.8 years, and the one who ran the CPD-0417 experiments left two years ago. What he has: 6,200 exported lab-notebook entries, 40 project post-mortems, 210 notes on outside literature, and compound names split across two generations. This piece shows what happens when that material goes into a Consilience workspace — what he asks, and what shape the answers come back in. It isn't a real customer story. It's a hypothetical scenario.
What's on the desk
The three-year-old experiments are on the record, and the reason they were shelved is written out in sentences in a post-mortem. The only problem is that the name attached to those records isn't the name on today's plan.
- 6,200 exported lab-notebook entries. One experiment per entry, with protocol, conditions, result values, and observation notes.
- 40 project post-mortems. The only place where why it stopped, and what the wall was, is written in sentences.
- 210 literature notes. Target and scaffold summaries pulled from papers, with links to the source.
- Two generations of compound names. Internal code names through 2023 (CPD-0417), a new naming scheme after that. There's no mapping table.
- Notes written by four people who've left. You can't go ask the author.
How it works today
Today the procedure is one search. Put the compound name from the plan into the company search box; if it returns nothing, try the target name instead. Searching by target pulls up more than 300 entries. Sorting out which of those used the same scaffold means opening them one at a time and eyeballing the structure.
So the actual judgment comes from human memory. If a scientist who's been around three years is in the room, someone says "we tried that already." If not, nobody does. A 2.8-year average tenure means that memory doesn't survive three years. The new scientist's plan gets approved, and six weeks later it stops at the same wall.
The person who ran that experiment left two years ago. Nowhere does it say when the name changed.
Getting this material in
The general procedure is in the academy lesson. What follows is only what's specific to this material. The scale — 6,200 entries — and two generations of names that don't resemble each other at all are what make this workspace unusual.
Drop in the whole export folder
Go to "Extract sources" (Cmd+Shift+I) in the sidebar and drag in the folders holding the 6,200 lab-notebook entries and the 40 post-mortems. For the paper links in the literature notes, paste them one per line under "Add manually" → "Add URL". 6,200 entries won't finish in one sitting. Kick it off and go do something else.
Set the domain to medicinal chemistry
Open "Domain vocabulary" in the left rail (⌘⇧G). Put medicinal chemistry in "Domain", put "Read these like a medicinal chemist" in "How the AI should read your notes", and press "Teach the AI your field". You don't need to hand-fill a compound list. As the panel says, the app learns the rest of the vocabulary from your notes on its own, and what you change applies to notes from now on.
Re-read the 6,200 older entries
For what you taught the app to reach the older notes, they have to be read again. Press "Re-extract everything" in the "Extraction" panel. If you send a chat while it's running, the "Knowledge graph still building" dialog asks whether you want "Send when extraction finishes" or "Send now". If you need an answer that swept all 6,200, pick the first one.
Confirm the two-generation names yourself
The "Synonyms" panel merges near-identical labels like "k8s" and "Kubernetes" on its own. CPD-0417 and the new name share no characters, so that path won't catch them. The pair shows up in "Relationships" instead. Pick "same" as the type and the two names merge into one compound, and the pair lands in "Synonyms" as "merged". "same" always waits for your confirmation, regardless of your settings. If it's wrong, flip it back to "kept apart" in "Synonyms".
What you ask
"Why did we stop the CPD-0417 work?"
→ 23 experiment entries, 1 post-mortem. The reason it stopped arrives quoted as
"2023-04 project post-mortem · line 34".
"Show me every experiment we've run on this target with this scaffold"
→ 4 compounds, 61 experiment entries. CPD-0417 and the new name come back on one row.
Every row carries a [[note]] link.
"Any past experiments that overlap this plan's screening conditions?"
→ 3. On one, a later note changed the conclusion — the entity page's "Revisions" has it.Confirm "same" once in "Relationships" and search treats the two names as one compound. Ask using the new name from the plan and the 23 entries from the CPD-0417 era come with it, and when you point at that compound with an @mention, whichever name you type lands in the same place.
The reason each program stopped comes back as sentences rather than a summary because the decisions, claims, and quotes pulled from the 40 post-mortems collect under "In your words" on the entity page. Those are the sentences a person wrote, unchanged, and pressing a [[bracketed]] reference in the answer takes you to that note. If you're going to keep coming back to this compound, right-click the chat → "Pin as case…" to keep it. Reopen it and the "New evidence" strip tells you which notes landed while you were away.
What changes
| Before | Now |
|---|---|
| Search the new name → zero results. The three-year-old experiments may as well not exist. | Whatever the name, it gathers under one compound, and the 23 experiment entries come with it. |
| Search the target → 300 results. Sorting out the same scaffold means opening entries and eyeballing them. | Ask for "this target with this scaffold" and 4 compounds and 61 entries come back with links. |
| Why it stopped lives in the memory of whoever stayed three years. No one there, no answer. | The reason it stopped arrives in the answer, quoted as "2023-04 project post-mortem · line 34". |
| Plan approved → six weeks later it stops at the same wall. | The 3 overlapping past experiments show up while you're still reviewing the plan. |
One confirmation covers exactly one pair. If several compounds have names split across two generations, each one has to be confirmed in "Relationships", and any pair that never gets suggested stays unmerged.