CellType growth playbook

Prerender or statically generate every marketing route (the site is on Vercel, so the build already supports this). Each page should send its full text, its own title and meta description, and Open Graph tags in the first HTML response. Then rebuild the sitemap to list every product, blog, science, benchmark, careers and UDI page, keep it referenced from robots.txt, and submit it in Search Console. Start with the five pages that carry the proof: the three product posts, /biosimbench and /udi/private.

CellType's best assets are written down: validated predictions, benchmarks and product explanations. A buyer or an assistant can only cite what it can read. One build change turns about a dozen invisible pages into pages that can be indexed, quoted and previewed, without new writing.

[ ] /blog/post/ct-perturb, /ct-tox, /ct-translate send full text in HTML
[ ] /science, /about, /careers, /contact, /cli, /changelog prerendered
[ ] /biosimbench, /biosimgym, /udi/private prerendered
[ ] Unique <title> and meta description on every page (no "The Biological World Model" default)
[ ] Open Graph title, description and image per page
[ ] Sitemap lists every page above, referenced from robots.txt
[ ] Sitemap submitted in Search Console
[ ] Build check: fail on empty body or default title

• Interactive charts or demos break when prerendered

  Prerender the text and hydrate the interactive parts on the client.

• Old URLs or titles linger in search results

  Request re-indexing of the key pages and check again after two weeks.

• Your websiteSubpages read without JavaScript (fetched twice on November 9, 2026): /blog, /blog/post/ct-perturb, /blog/post/ct-tox, /science, /about, /careers, /contact, /changelog, /cli, /biosimgym and /udi/private all return the same title, "CellType | The Biological World Model", with no body text and no meta description; /biosimbench has its own title but no body text either. Only the homepage and /udi return readable content, and the homepage title has moved on to "Simulate biology. Predict how your drugs work."Source
• Your websiterobots.txt points to a single sitemap, /udi/sitemap.xml, which lists one URL (/udi, last modified November 3, 2026). No product, blog, science, benchmark or careers page is listed.Source
• Hacker NewsA post titled "Rat-to-human liver gene response: 6× lower error than ortholog copy" links to celltype.com/blog/post/ct-tox; it had 2 points, posted about a month ago.Source
• PageSpeedHomepage report (November 2026): mobile performance 96 and desktop 100; real visitors are fast (75th-percentile LCP 836 ms, INP 62 ms, CLS 0.02, with LCP 30% faster than the previous period); the largest element is the homepage's intro paragraph; about 550 ms of render-blocking requests; one color-contrast issue; no social share image on the homepage.Source

Make /llms.txt a site-wide index. Add one short section per product (what it predicts, inputs, outputs, validated results, deployment on-prem or in a VPC, how to start a partnership), a BioSimBench section, and links to the full product pages once they are readable (idea 1). Move the UDI guide to /udi/llms.txt only and link to it from the index.

Scientists increasingly ask assistants which tools fit a problem. A plain-text summary the company controls is the cheapest way to make sure CT-Tox is the answer to "predict human liver toxicity from rat data".

• Assistants ignore llms.txt

  Pair it with readable pages (idea 1), which crawlers do read.

• #1 Make CellType's product, science and blog pages readable without JavaScript, and list them in the sitemap

• Your website/llms.txt and /udi/llms.txt (last updated November 3, 2026) are full guides for people and AI agents to UDI, its benchmark and its setup in each assistant, but neither mentions CT-Perturb, CT-Tox, CT-Translate or BioSimBench.Source
• Your websiteSubpages read without JavaScript (fetched twice on November 9, 2026): /blog, /blog/post/ct-perturb, /blog/post/ct-tox, /science, /about, /careers, /contact, /changelog, /cli, /biosimgym and /udi/private all return the same title, "CellType | The Biological World Model", with no body text and no meta description; /biosimbench has its own title but no body text either. Only the homepage and /udi return readable content, and the homepage title has moved on to "Simulate biology. Predict how your drugs work."Source

Under each UDI result and in the MCP tool's responses, add one line: "Want this alongside your own assays? Private UDI" and "Predict what your compound does in these cells: CT-Perturb". Both lead to a short partner form that keeps the query as context. On the UDI page, add a "From search to prediction" block linking to the three product pages.

The highest-intent moment is right after a useful answer. Carrying the query into the form gives the partnerships lead a specific opening ("you searched IPF single-cell data; here's what CT-Perturb would do with your screen").

• Users find the prompts pushy

  One line, only after a result, and never in the middle of an answer.

• Your websiteUDI page: "UDI, the Universal Data Index" searches "what public biomedical datasets measured, not just what they are called": 1.9M+ public studies from 67 sources, 640K+ signatures, 230K+ single-cell studies, 41M+ papers and 600K+ clinical trials; on 443 dataset requests from 2026 papers it put a right dataset in the top 10 for 92% of requests, against 73% for an AI with web search; "Free during the beta, 10,000 calls a month, no card", with no paid tier listed; it works in the browser or through a hosted MCP server with read-only sign-in to a free CellType Cloud account; the page's browser link is celltype.com/udi while its setup guide gives cloud.celltype.com/udi; it links to Private UDI ("your private data beside public data in one search, inside your environment") and back to the homepage, but not to the Partner page.Source
• Your websiteHomepage (checked November 9, 2026): H1 "Simulate biology. Predict how your drugs work."; three products, CT-Perturb ("Predict cellular responses to compounds"), CT-Tox ("Translate animal safety signals to humans") and CT-Translate ("Connect preclinical data to patient outcomes"), plus BioSimBench and BioSimGym; pre-trained on "the world's largest repository of single-cell, spatial and bulk RNA data", fine-tuned on a partner's own assays, and run on-prem or in the customer's VPC; 4,266 oncology compounds screened, 23 prioritized and 3 novel hits validated in ex-vivo human tumor tissue; the only logos are Yale and Y Combinator and no customers are named; every call to action is "Partner with us", and the closing section asks for "a held-out dataset, a quantitative endpoint, and the baseline to beat".Source

Write a Private UDI page with a clear offer: which data types it ingests, deployment options (on-prem or in the customer's VPC, as the homepage already promises), security and data handling, a fixed-length pilot (for example four weeks on one internal data collection) and a request form. Prerender it (idea 1) and link it from every UDI touchpoint.

R&D informatics buyers need to forward a page to IT and procurement. A specific offer turns curiosity from free users into scoped pilots.

• Pilots consume a five-person team

  Cap concurrent pilots at two and qualify on data readiness.

• #1 Make CellType's product, science and blog pages readable without JavaScript, and list them in the sitemap

• Your websiteUDI page: "UDI, the Universal Data Index" searches "what public biomedical datasets measured, not just what they are called": 1.9M+ public studies from 67 sources, 640K+ signatures, 230K+ single-cell studies, 41M+ papers and 600K+ clinical trials; on 443 dataset requests from 2026 papers it put a right dataset in the top 10 for 92% of requests, against 73% for an AI with web search; "Free during the beta, 10,000 calls a month, no card", with no paid tier listed; it works in the browser or through a hosted MCP server with read-only sign-in to a free CellType Cloud account; the page's browser link is celltype.com/udi while its setup guide gives cloud.celltype.com/udi; it links to Private UDI ("your private data beside public data in one search, inside your environment") and back to the homepage, but not to the Partner page.Source
• Your websiteSubpages read without JavaScript (fetched twice on November 9, 2026): /blog, /blog/post/ct-perturb, /blog/post/ct-tox, /science, /about, /careers, /contact, /changelog, /cli, /biosimgym and /udi/private all return the same title, "CellType | The Biological World Model", with no body text and no meta description; /biosimbench has its own title but no body text either. Only the homepage and /udi return readable content, and the homepage title has moved on to "Simulate biology. Predict how your drugs work."Source
• Your websiteHomepage (checked November 9, 2026): H1 "Simulate biology. Predict how your drugs work."; three products, CT-Perturb ("Predict cellular responses to compounds"), CT-Tox ("Translate animal safety signals to humans") and CT-Translate ("Connect preclinical data to patient outcomes"), plus BioSimBench and BioSimGym; pre-trained on "the world's largest repository of single-cell, spatial and bulk RNA data", fine-tuned on a partner's own assays, and run on-prem or in the customer's VPC; 4,266 oncology compounds screened, 23 prioritized and 3 novel hits validated in ex-vivo human tumor tissue; the only logos are Yale and Y Combinator and no customers are named; every call to action is "Partner with us", and the closing section asks for "a held-out dataset, a quantitative endpoint, and the baseline to beat".Source

Choose one address for the search app (likely cloud.celltype.com/udi, with celltype.com/udi as the marketing page), make every link and the setup guide use it, and redirect or canonicalize the other. Tag links from the MCP setup, the CLI and social posts so each source can be told apart.

One clear path removes a step of hesitation for new users and makes the funnel in idea 13 measurable.

• Breaking saved bookmarks

  Use permanent redirects, never removal.

• Your websiteUDI page: "UDI, the Universal Data Index" searches "what public biomedical datasets measured, not just what they are called": 1.9M+ public studies from 67 sources, 640K+ signatures, 230K+ single-cell studies, 41M+ papers and 600K+ clinical trials; on 443 dataset requests from 2026 papers it put a right dataset in the top 10 for 92% of requests, against 73% for an AI with web search; "Free during the beta, 10,000 calls a month, no card", with no paid tier listed; it works in the browser or through a hosted MCP server with read-only sign-in to a free CellType Cloud account; the page's browser link is celltype.com/udi while its setup guide gives cloud.celltype.com/udi; it links to Private UDI ("your private data beside public data in one search, inside your environment") and back to the homepage, but not to the Partner page.Source

Design one share-image template (product name, the key number, CellType mark) and generate an image for each page: "6× lower error, rat-to-human liver response" for CT-Tox, "4,266 compounds, 3 validated hits" for the oncology screen, and the 92%-vs-73% result for UDI. Add Open Graph and X card tags with per-page descriptions.

Scientists scroll past bare links. A preview with the result in it earns the click, which matters most for a company with no paid acquisition.

• Overstated numbers in previews

  Use only results stated on the page, with the same wording.

• #1 Make CellType's product, science and blog pages readable without JavaScript, and list them in the sitemap

• PageSpeedHomepage report (November 2026): mobile performance 96 and desktop 100; real visitors are fast (75th-percentile LCP 836 ms, INP 62 ms, CLS 0.02, with LCP 30% faster than the previous period); the largest element is the homepage's intro paragraph; about 550 ms of render-blocking requests; one color-contrast issue; no social share image on the homepage.Source
• Hacker NewsA post titled "Rat-to-human liver gene response: 6× lower error than ortholog copy" links to celltype.com/blog/post/ct-tox; it had 2 points, posted about a month ago.Source
• Your websiteSubpages read without JavaScript (fetched twice on November 9, 2026): /blog, /blog/post/ct-perturb, /blog/post/ct-tox, /science, /about, /careers, /contact, /changelog, /cli, /biosimgym and /udi/private all return the same title, "CellType | The Biological World Model", with no body text and no meta description; /biosimbench has its own title but no body text either. Only the homepage and /udi return readable content, and the homepage title has moved on to "Simulate biology. Predict how your drugs work."Source

Update the Product Hunt links to the current repository, give /cli real content (what it does, install, the BixBench result, a link to the repository), and add a "Search public data with UDI" section and a "Run on your own data" link to the README. Pin a short post on the repository about Private UDI and partnerships.

Developers and computational biologists who star the repository are CellType's warmest free audience. Every dead or blank link loses some of them.

• README turns into marketing

  Keep it to two short sections at the end.

• Product HuntCellType CLI listing ("The biological reasoning engine for your terminal"): 2 points, 2 followers, no reviews; it links to github.com/celltype/cli and celltype.com/cli, while the GitHub organization's only public repository is now celltype-agent.Source
• GitHubcelltype-agent (the CellType CLI): open source under MIT, 240 stars and 47 forks, 8 open pull requests; "190+ drug discovery tools", "30+ database APIs", a persistent Python and R sandbox and HTML report export; reports 90% on BixBench-Verified-50; the README links to celltype.com/cli and [email protected] for enterprise, on-premise use, and does not mention UDI.Source
• Your websiteSubpages read without JavaScript (fetched twice on November 9, 2026): /blog, /blog/post/ct-perturb, /blog/post/ct-tox, /science, /about, /careers, /contact, /changelog, /cli, /biosimgym and /udi/private all return the same title, "CellType | The Biological World Model", with no body text and no meta description; /biosimbench has its own title but no body text either. Only the homepage and /udi return readable content, and the homepage title has moved on to "Simulate biology. Predict how your drugs work."Source

Ship a UDI tool in the agent that is used by default for dataset and evidence lookups, with sign-in to a free CellType Cloud account on first use. Show the source of each dataset in the HTML report with a "found with UDI" link. Publish a before-and-after BixBench run, with UDI and without.

It joins the two free products into one loop: agent users become UDI accounts, and UDI accounts are the audience for Private UDI and partnerships.

• Open-source users dislike required accounts

  Keep the agent usable without UDI; make the account optional and the benefit clear.

• #7 Fix the Product Hunt and README links so CLI users land on a live page and on UDI

• GitHubcelltype-agent (the CellType CLI): open source under MIT, 240 stars and 47 forks, 8 open pull requests; "190+ drug discovery tools", "30+ database APIs", a persistent Python and R sandbox and HTML report export; reports 90% on BixBench-Verified-50; the README links to celltype.com/cli and [email protected] for enterprise, on-premise use, and does not mention UDI.Source
• Your websiteUDI page: "UDI, the Universal Data Index" searches "what public biomedical datasets measured, not just what they are called": 1.9M+ public studies from 67 sources, 640K+ signatures, 230K+ single-cell studies, 41M+ papers and 600K+ clinical trials; on 443 dataset requests from 2026 papers it put a right dataset in the top 10 for 92% of requests, against 73% for an AI with web search; "Free during the beta, 10,000 calls a month, no card", with no paid tier listed; it works in the browser or through a hosted MCP server with read-only sign-in to a free CellType Cloud account; the page's browser link is celltype.com/udi while its setup guide gives cloud.celltype.com/udi; it links to Private UDI ("your private data beside public data in one search, inside your environment") and back to the homepage, but not to the Partner page.Source

Make /biosimbench a readable leaderboard: tasks, held-out splits, metrics, baselines and current scores, with a submission process, a public results table and a paper or preprint. Invite academic groups and companies to submit. Keep BioSimGym as the training counterpart for teams that want to improve their score.

Benchmarks get cited in every paper that uses them, and "tops BioSimBench" becomes a phrase other companies repeat, pointing back to CellType.

• Seen as rigged if CellType always wins

  Publish splits and evaluation code, and accept results where CellType loses.

• #1 Make CellType's product, science and blog pages readable without JavaScript, and list them in the sitemap

• Your websiteHomepage (checked November 9, 2026): H1 "Simulate biology. Predict how your drugs work."; three products, CT-Perturb ("Predict cellular responses to compounds"), CT-Tox ("Translate animal safety signals to humans") and CT-Translate ("Connect preclinical data to patient outcomes"), plus BioSimBench and BioSimGym; pre-trained on "the world's largest repository of single-cell, spatial and bulk RNA data", fine-tuned on a partner's own assays, and run on-prem or in the customer's VPC; 4,266 oncology compounds screened, 23 prioritized and 3 novel hits validated in ex-vivo human tumor tissue; the only logos are Yale and Y Combinator and no customers are named; every call to action is "Partner with us", and the closing section asks for "a held-out dataset, a quantitative endpoint, and the baseline to beat".Source
• Your websiteSubpages read without JavaScript (fetched twice on November 9, 2026): /blog, /blog/post/ct-perturb, /blog/post/ct-tox, /science, /about, /careers, /contact, /changelog, /cli, /biosimgym and /udi/private all return the same title, "CellType | The Biological World Model", with no body text and no meta description; /biosimbench has its own title but no body text either. Only the homepage and /udi return readable content, and the homepage title has moved on to "Simulate biology. Predict how your drugs work."Source
• Competitor researchRivals' own homepages: Noetik, "an AI-native biotechnology company" with "one of the largest collections of multimodal tumor data" and trials "enriched for likely responders"; Tahoe Labs, "Building the ground truth for biological superintelligence", with perturbative single-cell data, response-prediction models and the Tara research; Recursion, "The AI engine uncovering novel medicines that matter".See the research Source

Package a named offer, for example a "Baseline Challenge": a fixed four-to-six-week pilot where CellType fine-tunes on the partner's data inside their environment, predicts the held-out set, and reports against their baseline with a pre-agreed success threshold. Publish what the partner provides, what they get, the timeline, the price range or credit-against-contract terms, and the security posture. Use the form to collect the three items up front.

Pharma buyers approve scoped, measurable pilots far faster than open collaborations. A head-to-head against the partner's own baseline is the proof that converts to a contract.

• The model fails to beat a strong baseline

  Qualify the endpoints in advance and treat losses as roadmap input, not public results.

• Your websiteHomepage (checked November 9, 2026): H1 "Simulate biology. Predict how your drugs work."; three products, CT-Perturb ("Predict cellular responses to compounds"), CT-Tox ("Translate animal safety signals to humans") and CT-Translate ("Connect preclinical data to patient outcomes"), plus BioSimBench and BioSimGym; pre-trained on "the world's largest repository of single-cell, spatial and bulk RNA data", fine-tuned on a partner's own assays, and run on-prem or in the customer's VPC; 4,266 oncology compounds screened, 23 prioritized and 3 novel hits validated in ex-vivo human tumor tissue; the only logos are Yale and Y Combinator and no customers are named; every call to action is "Partner with us", and the closing section asks for "a held-out dataset, a quantitative endpoint, and the baseline to beat".Source
• HiringFounding Research Scientist, Biological Foundation Models (New York, $150K-$250K, 0.50%-2.00% equity, visa sponsorship, open on Y Combinator's job board): "Own modeling for live pharma and biotech programs, from problem framing through partner-facing results"; "Design evaluations that biologists trust"; "Decide what data to generate next with academic and pharma collaborators"; "Turn project results into reusable capabilities and product workflows"; partners are pharma, biotech and diagnostics companies, and the aim is "the core intelligence layer for biology".Source
• Y CombinatorLaunch post: the team "screened 4,000+ drugs" and predicted a candidate to make "cold" tumors more visible to the immune system, validated in cell lines and human tumor microenvironments; it says it is working with Top 10 pharma with every deal inbound, and asks for pharma and biotech teams de-risking discovery, "data partners" and biology and ML builders.Source

Write one long-form, prerendered case study: the question, the data, how the screen was run, the 23 picks, the validation assays, the hit rate compared with a random or standard screen, and what it means for a partner's own library. Include a section on the CX-4945 program as far as the partner agrees to publish. Offer a downloadable PDF for R&D committees.

Pharma scientists need methods, not slogans. A case study they can forward internally does the selling between meetings.

• Partner confidentiality

  Publish the public research results first and add partner details only with written approval.

• #1 Make CellType's product, science and blog pages readable without JavaScript, and list them in the sitemap

• Your websiteHomepage (checked November 9, 2026): H1 "Simulate biology. Predict how your drugs work."; three products, CT-Perturb ("Predict cellular responses to compounds"), CT-Tox ("Translate animal safety signals to humans") and CT-Translate ("Connect preclinical data to patient outcomes"), plus BioSimBench and BioSimGym; pre-trained on "the world's largest repository of single-cell, spatial and bulk RNA data", fine-tuned on a partner's own assays, and run on-prem or in the customer's VPC; 4,266 oncology compounds screened, 23 prioritized and 3 novel hits validated in ex-vivo human tumor tissue; the only logos are Yale and Y Combinator and no customers are named; every call to action is "Partner with us", and the closing section asks for "a held-out dataset, a quantitative endpoint, and the baseline to beat".Source
• PressGoogle blog (October 15, 2025): C2S-Scale 27B, "a new 27 billion parameter foundation model designed to understand the language of individual cells", from a research collaboration with Yale, predicted that silmitasertib would boost antigen presentation only with low-level interferon; lab tests showed "a roughly 50% increase in antigen presentation"; the preprint, model and code are public.Source
• PressBioSpectrum Asia (March 4, 2026): Senhwa Biosciences signed an MoU with CellType to apply its AI platform to silmitasertib (CX-4945) for clinical development and licensing; CellType will feature CX-4945 as the flagship case study of its platform and Senhwa becomes its founding strategic partner; no financial terms were disclosed.Source

Use beta usage to set three tiers: Free (individuals, current limits), Team (shared workspaces, higher limits, a usage dashboard, priority support) and Private UDI (on-prem or VPC, the user's own data). Announce the tiers on the UDI page well before any change and grandfather active beta users for a period.

A visible ladder lets small biotechs pay by card, and it turns the biggest teams into qualified Private UDI leads.

• Pricing slows adoption

  Keep a generous free tier; charge for teams and private data, not individuals.

• #4 Give Private UDI a real offer page: what is deployed, how a pilot works, and how to start

• Your websiteUDI page: "UDI, the Universal Data Index" searches "what public biomedical datasets measured, not just what they are called": 1.9M+ public studies from 67 sources, 640K+ signatures, 230K+ single-cell studies, 41M+ papers and 600K+ clinical trials; on 443 dataset requests from 2026 papers it put a right dataset in the top 10 for 92% of requests, against 73% for an AI with web search; "Free during the beta, 10,000 calls a month, no card", with no paid tier listed; it works in the browser or through a hosted MCP server with read-only sign-in to a free CellType Cloud account; the page's browser link is celltype.com/udi while its setup guide gives cloud.celltype.com/udi; it links to Private UDI ("your private data beside public data in one search, inside your environment") and back to the homepage, but not to the Partner page.Source

Define five events: first search, account created, MCP connected, Private UDI or product page viewed, partner form submitted. Tag the source (browser, MCP, CLI, social, Hacker News) and report the funnel weekly. Record the source on every partner conversation as well.

A five-person team has to choose where to spend effort. The funnel shows whether the CLI, UDI or BioSimBench earns the next hour of work.

• Tracking researchers' queries feels intrusive

  Track events and sources, not query text, and say so in the privacy policy.

• #5 Pick one browser address for UDI and use it everywhere

• ResearchSite scan (November 2026): one product-analytics tool (80 KB, 4 requests) and Google Fonts are the only third parties on the homepage; no newsletter, no case studies, no testimonials and no customer logos beyond Yale and Y Combinator were found; a trust center at /trust.Source
• Your websiteUDI page: "UDI, the Universal Data Index" searches "what public biomedical datasets measured, not just what they are called": 1.9M+ public studies from 67 sources, 640K+ signatures, 230K+ single-cell studies, 41M+ papers and 600K+ clinical trials; on 443 dataset requests from 2026 papers it put a right dataset in the top 10 for 92% of requests, against 73% for an AI with web search; "Free during the beta, 10,000 calls a month, no card", with no paid tier listed; it works in the browser or through a hosted MCP server with read-only sign-in to a free CellType Cloud account; the page's browser link is celltype.com/udi while its setup guide gives cloud.celltype.com/udi; it links to Private UDI ("your private data beside public data in one search, inside your environment") and back to the homepage, but not to the Partner page.Source

Offer a no-cost sample: a safety team sends one completed rat study with known human outcome (or a public one), and CellType returns a short human-translation report. It compares the predicted human liver response with ortholog mapping and flags where the two disagree. Promote it to toxicology leads with the CT-Tox post and a sample report built from public data.

Retrospective studies with known answers let a skeptical safety team test the model without risk. A correct call on their own history is the most persuasive demo possible.

• Data-sharing hurdles at pharma

  Offer to run inside the customer's VPC, as the homepage already promises.

• #1 Make CellType's product, science and blog pages readable without JavaScript, and list them in the sitemap

• Hacker NewsA post titled "Rat-to-human liver gene response: 6× lower error than ortholog copy" links to celltype.com/blog/post/ct-tox; it had 2 points, posted about a month ago.Source
• Your websiteHomepage (checked November 9, 2026): H1 "Simulate biology. Predict how your drugs work."; three products, CT-Perturb ("Predict cellular responses to compounds"), CT-Tox ("Translate animal safety signals to humans") and CT-Translate ("Connect preclinical data to patient outcomes"), plus BioSimBench and BioSimGym; pre-trained on "the world's largest repository of single-cell, spatial and bulk RNA data", fine-tuned on a partner's own assays, and run on-prem or in the customer's VPC; 4,266 oncology compounds screened, 23 prioritized and 3 novel hits validated in ex-vivo human tumor tissue; the only logos are Yale and Y Combinator and no customers are named; every call to action is "Partner with us", and the closing section asks for "a held-out dataset, a quantitative endpoint, and the baseline to beat".Source
• Your websiteHomepage safety section: "Cross-species transfer" to "kill unsafe programs before they enter the clinic"; the translate section covers responder signatures, resistance mechanisms and biomarkers.Source

Publish a data-partner page: which data types are wanted (perturbation screens, paired animal-human, spatial, clinical outcomes), and what partners receive: early model access, co-authorship, a fine-tuned model on their own data, or credits. Set out how data is protected and attributed. Start with labs and biotechs that already have unpublished perturbation screens.

Better data is the moat for a foundation model. Clear terms turn one-off favors into a pipeline of datasets, each bringing an organization that may become a customer.

• Data rights disputes later

  Use one clear template with ownership and use terms spelled out.

• Y CombinatorLaunch post: the team "screened 4,000+ drugs" and predicted a candidate to make "cold" tumors more visible to the immune system, validated in cell lines and human tumor microenvironments; it says it is working with Top 10 pharma with every deal inbound, and asks for pharma and biotech teams de-risking discovery, "data partners" and biology and ML builders.Source
• HiringFounding Research Scientist, Biological Foundation Models (New York, $150K-$250K, 0.50%-2.00% equity, visa sponsorship, open on Y Combinator's job board): "Own modeling for live pharma and biotech programs, from problem framing through partner-facing results"; "Design evaluations that biologists trust"; "Decide what data to generate next with academic and pharma collaborators"; "Turn project results into reusable capabilities and product workflows"; partners are pharma, biotech and diagnostics companies, and the aim is "the core intelligence layer for biology".Source
• Competitor researchRivals' own homepages: Noetik, "an AI-native biotechnology company" with "one of the largest collections of multimodal tumor data" and trials "enriched for likely responders"; Tahoe Labs, "Building the ground truth for biological superintelligence", with perturbative single-cell data, response-prediction models and the Tara research; Recursion, "The AI engine uncovering novel medicines that matter".See the research Source

Prerender /careers with each open role in full and job-posting structured data, so search engines show it as a job. Add a short "why join" section built on what is already public: the Cell2Sentence papers, the validated predictions, live pharma programs. Link the YC posting and add a referral note asking the research community to forward it.

Readable job pages with structured data appear in job search results for free, and a strong founding research hire speeds up every partner program.

• Stale roles stay listed

  Generate the page from one source of truth and remove filled roles the same day.

• #1 Make CellType's product, science and blog pages readable without JavaScript, and list them in the sitemap

• HiringFounding Research Scientist, Biological Foundation Models (New York, $150K-$250K, 0.50%-2.00% equity, visa sponsorship, open on Y Combinator's job board): "Own modeling for live pharma and biotech programs, from problem framing through partner-facing results"; "Design evaluations that biologists trust"; "Decide what data to generate next with academic and pharma collaborators"; "Turn project results into reusable capabilities and product workflows"; partners are pharma, biotech and diagnostics companies, and the aim is "the core intelligence layer for biology".Source
• Your websiteSubpages read without JavaScript (fetched twice on November 9, 2026): /blog, /blog/post/ct-perturb, /blog/post/ct-tox, /science, /about, /careers, /contact, /changelog, /cli, /biosimgym and /udi/private all return the same title, "CellType | The Biological World Model", with no body text and no meta description; /biosimbench has its own title but no body text either. Only the homepage and /udi return readable content, and the homepage title has moved on to "Simulate biology. Predict how your drugs work."Source
• Y CombinatorYC profile: "The agentic drug company. We simulate human biology." Winter 2026 batch, founded 2025, team of 5, New York; one open job listed.Source

Write one page per decision, each answering five to eight real questions in plain language, with CellType's evidence (the oncology funnel, the CT-Tox error result, the publications) and a link to the matching product and pilot. Prerender them, add FAQ structured data, and include them in llms.txt.

Decision-level pages match how buyers search and give assistants a citable source with CellType's results in it.

• Thin, generic content

  Every answer cites a CellType result or paper.

• #1 Make CellType's product, science and blog pages readable without JavaScript, and list them in the sitemap
• #2 Extend llms.txt from UDI to the whole platform: CT-Perturb, CT-Tox, CT-Translate and BioSimBench

• Your websiteHomepage safety section: "Cross-species transfer" to "kill unsafe programs before they enter the clinic"; the translate section covers responder signatures, resistance mechanisms and biomarkers.Source
• Your websiteUDI example questions include "Find human single-cell RNA-seq of IPF lung with at least 10 donors", "What human genetic evidence links IL23R to psoriasis?" and "Where does SPP1 go up in IPF lung?".Source
• Your websiteSubpages read without JavaScript (fetched twice on November 9, 2026): /blog, /blog/post/ct-perturb, /blog/post/ct-tox, /science, /about, /careers, /contact, /changelog, /cli, /biosimgym and /udi/private all return the same title, "CellType | The Biological World Model", with no body text and no meta description; /biosimbench has its own title but no body text either. Only the homepage and /udi return readable content, and the homepage title has moved on to "Simulate biology. Predict how your drugs work."Source

Launch a limited self-serve CT-Perturb in CellType Cloud: upload up to N compounds (SMILES), pick cell types from public data, get predicted expression changes and pathway scores with confidence bands, and pay per run. Keep fine-tuning on private data as the partnership upgrade. Start invite-only with the most active UDI users.

It turns free UDI users into paying users and creates a steady stream of results that feed case studies and partnership leads.

• Self-serve results are over-trusted

  Show confidence and recommend wet-lab validation on every report.

• It undercuts partnership pricing

  Limit self-serve to public cell states; private fine-tuning stays a partnership.

• #12 Price UDI before the beta ends: keep a free tier and add a team tier that leads to Private UDI
• #13 Track one funnel from UDI search to partner call, with the analytics already on the site

• Your websiteUDI page: "UDI, the Universal Data Index" searches "what public biomedical datasets measured, not just what they are called": 1.9M+ public studies from 67 sources, 640K+ signatures, 230K+ single-cell studies, 41M+ papers and 600K+ clinical trials; on 443 dataset requests from 2026 papers it put a right dataset in the top 10 for 92% of requests, against 73% for an AI with web search; "Free during the beta, 10,000 calls a month, no card", with no paid tier listed; it works in the browser or through a hosted MCP server with read-only sign-in to a free CellType Cloud account; the page's browser link is celltype.com/udi while its setup guide gives cloud.celltype.com/udi; it links to Private UDI ("your private data beside public data in one search, inside your environment") and back to the homepage, but not to the Partner page.Source
• Your websiteHomepage (checked November 9, 2026): H1 "Simulate biology. Predict how your drugs work."; three products, CT-Perturb ("Predict cellular responses to compounds"), CT-Tox ("Translate animal safety signals to humans") and CT-Translate ("Connect preclinical data to patient outcomes"), plus BioSimBench and BioSimGym; pre-trained on "the world's largest repository of single-cell, spatial and bulk RNA data", fine-tuned on a partner's own assays, and run on-prem or in the customer's VPC; 4,266 oncology compounds screened, 23 prioritized and 3 novel hits validated in ex-vivo human tumor tissue; the only logos are Yale and Y Combinator and no customers are named; every call to action is "Partner with us", and the closing section asks for "a held-out dataset, a quantitative endpoint, and the baseline to beat".Source
• Y CombinatorYC profile: "The agentic drug company. We simulate human biology." Winter 2026 batch, founded 2025, team of 5, New York; one open job listed.Source

Turn the deal into a named program for clinical-stage assets: a standard analysis (indications, responder biomarkers, combinations) delivered under fixed terms, a fee plus milestone or royalty share, and an option for deeper co-development. Publish the program page and approach companies with assets that match CellType's validated strengths (immuno-oncology first).

Asset-linked deals grow with the value created, and each one is a case study written in clinical outcomes, the proof pharma buyers trust most.

• Revenue depends on slow clinical milestones

  Always include an upfront analysis fee.

• #11 Publish the 4,266-to-3 oncology screen as a full case study with methods

• PressBioSpectrum Asia (March 4, 2026): Senhwa Biosciences signed an MoU with CellType to apply its AI platform to silmitasertib (CX-4945) for clinical development and licensing; CellType will feature CX-4945 as the flagship case study of its platform and Senhwa becomes its founding strategic partner; no financial terms were disclosed.Source
• Y CombinatorYC profile: "The agentic drug company. We simulate human biology." Winter 2026 batch, founded 2025, team of 5, New York; one open job listed.Source
• PressGoogle blog (October 15, 2025): C2S-Scale 27B, "a new 27 billion parameter foundation model designed to understand the language of individual cells", from a research collaboration with Yale, predicted that silmitasertib would boost antigen presentation only with low-level interferon; lab tests showed "a roughly 50% increase in antigen presentation"; the preprint, model and code are public.Source

Publish a yearly report: benchmark results across public models, where simulation works and fails (by tissue, perturbation type, species gap), and aggregate, anonymized trends in what scientists search for (top diseases, data types, unmet data needs). Release it as a prerendered page, a PDF and a preprint, with charts built for sharing.

An authoritative, data-backed annual report earns links, press and inbound interest every year, from people who would not read a product page.

• Users worry about query privacy

  Publish only aggregates above a threshold and explain the method.

• #9 Publish BioSimBench as an open leaderboard that other model builders can submit to

• Your websiteUDI page: "UDI, the Universal Data Index" searches "what public biomedical datasets measured, not just what they are called": 1.9M+ public studies from 67 sources, 640K+ signatures, 230K+ single-cell studies, 41M+ papers and 600K+ clinical trials; on 443 dataset requests from 2026 papers it put a right dataset in the top 10 for 92% of requests, against 73% for an AI with web search; "Free during the beta, 10,000 calls a month, no card", with no paid tier listed; it works in the browser or through a hosted MCP server with read-only sign-in to a free CellType Cloud account; the page's browser link is celltype.com/udi while its setup guide gives cloud.celltype.com/udi; it links to Private UDI ("your private data beside public data in one search, inside your environment") and back to the homepage, but not to the Partner page.Source
• Your websiteHomepage (checked November 9, 2026): H1 "Simulate biology. Predict how your drugs work."; three products, CT-Perturb ("Predict cellular responses to compounds"), CT-Tox ("Translate animal safety signals to humans") and CT-Translate ("Connect preclinical data to patient outcomes"), plus BioSimBench and BioSimGym; pre-trained on "the world's largest repository of single-cell, spatial and bulk RNA data", fine-tuned on a partner's own assays, and run on-prem or in the customer's VPC; 4,266 oncology compounds screened, 23 prioritized and 3 novel hits validated in ex-vivo human tumor tissue; the only logos are Yale and Y Combinator and no customers are named; every call to action is "Partner with us", and the closing section asks for "a held-out dataset, a quantitative endpoint, and the baseline to beat".Source

Give each UDI result set a permanent, public link: the query, the datasets found, their sources and a "Cite this search" snippet (with a version and date). Opening a shared link shows the results and invites the reader to run their own search. Add a "copy citation" button and the same link in MCP responses.

Every share becomes a working demo for a new reader, and citations in papers and decks compound over time.

• Sharing exposes sensitive research directions

  Keep sharing opt-in per search, and never share by default.

• #5 Pick one browser address for UDI and use it everywhere

• Your websiteUDI page: "UDI, the Universal Data Index" searches "what public biomedical datasets measured, not just what they are called": 1.9M+ public studies from 67 sources, 640K+ signatures, 230K+ single-cell studies, 41M+ papers and 600K+ clinical trials; on 443 dataset requests from 2026 papers it put a right dataset in the top 10 for 92% of requests, against 73% for an AI with web search; "Free during the beta, 10,000 calls a month, no card", with no paid tier listed; it works in the browser or through a hosted MCP server with read-only sign-in to a free CellType Cloud account; the page's browser link is celltype.com/udi while its setup guide gives cloud.celltype.com/udi; it links to Private UDI ("your private data beside public data in one search, inside your environment") and back to the homepage, but not to the Partner page.Source
• Your websiteUDI example questions include "Find human single-cell RNA-seq of IPF lung with at least 10 donors", "What human genetic evidence links IL23R to psoriasis?" and "Where does SPP1 go up in IPF lung?".Source

Offer Private UDI (and later self-serve CT-Perturb) free to a capped number of academic labs per year. In return, labs cite CellType in publications, give feedback, and opt in to sharing selected datasets under the data-partner program. Apply through a short form and announce the cohort publicly.

Academic adoption builds citations, data and a talent pipeline cheaply, and graduates carry tool preferences into industry.

• Support load on a small team

  Cap the cohort and require a named lab contact.

• #4 Give Private UDI a real offer page: what is deployed, how a pilot works, and how to start
• #15 Open a data-partner program with clear terms, since the launch post already asks for data partners

• Your websiteHomepage science and team: publications at ICML 2024 (Cell2Sentence), bioRxiv 2025 (scaling large language models for next-generation single-cell analysis) and Nature Methods 2023 (CINEMA-OT); the architecture was "built by the inventors of Cell2Sentence at Yale"; a team of five is listed, including a Chief Strategy & Partnerships Officer; offices in San Francisco and New York.Source
• HiringFounding Research Scientist, Biological Foundation Models (New York, $150K-$250K, 0.50%-2.00% equity, visa sponsorship, open on Y Combinator's job board): "Own modeling for live pharma and biotech programs, from problem framing through partner-facing results"; "Design evaluations that biologists trust"; "Decide what data to generate next with academic and pharma collaborators"; "Turn project results into reusable capabilities and product workflows"; partners are pharma, biotech and diagnostics companies, and the aim is "the core intelligence layer for biology".Source
• Your websiteUDI page: "UDI, the Universal Data Index" searches "what public biomedical datasets measured, not just what they are called": 1.9M+ public studies from 67 sources, 640K+ signatures, 230K+ single-cell studies, 41M+ papers and 600K+ clinical trials; on 443 dataset requests from 2026 papers it put a right dataset in the top 10 for 92% of requests, against 73% for an AI with web search; "Free during the beta, 10,000 calls a month, no card", with no paid tier listed; it works in the browser or through a hosted MCP server with read-only sign-in to a free CellType Cloud account; the page's browser link is celltype.com/udi while its setup guide gives cloud.celltype.com/udi; it links to Private UDI ("your private data beside public data in one search, inside your environment") and back to the homepage, but not to the Partner page.Source