AI-designed drugs, proteins and materials: pretty lab slides do not pay
AI-designed drugs need wet-lab proof, regulation, manufacturing and buyers before they become companies. Use this founder filter before building.
A beautiful molecule is not a company.
That sounds rude only if you have never watched a founder confuse a lab slide, a model output and a business. AI-designed drugs, proteins and materials are one of the most serious openings in deep tech, but the hard part starts after the model says yes.
TL;DR: AI-designed drugs, proteins and materials use machine learning, protein design, molecular modeling, simulation and lab data to suggest new molecules, binders, enzymes, catalysts, battery materials, polymers or other scientific assets. The startup trap is treating the design as the win. A founder still needs wet-lab proof, safety records, IP clarity, regulation work, manufacturing logic, partner access, buyer budget and cash discipline. Use the article below as a founder filter before you build a company around a molecule.
I am Violetta Bonenkamp, founder of Mean CEO, CADChain, and F/MS Startup Game. CADChain works close to hard technology, CAD data, intellectual property, machine learning, R&D and manufacturing reality. That makes me allergic to startup advice that treats science like a prettier SaaS pitch.
AI can propose.
The market still asks: can you prove, make, protect, sell and survive it?
What AI-Designed Drugs, Proteins And Materials Means
AI-designed drugs means software helps choose or create a drug candidate. That can include target choice, hit discovery, molecular generation, binding prediction, toxicity checks, clinical trial planning or patient matching.
AI-designed proteins means software helps create or alter amino acid sequences so a protein folds into a wanted shape or performs a wanted job. That can mean enzymes, binders, antibodies, vaccines, sensors or protein materials.
AI-designed materials means AI helps propose new inorganic crystals, polymers, battery materials, catalysts, coatings, semiconductors or other materials with desired properties.
The field sits between AI for science, computational biology, chemistry, materials science, lab automation, IP, regulation and industrial sales.
That is why founders need a business lens early.
The model output may be:
- A structure prediction.
- A generated molecule.
- A ranked candidate list.
- A protein sequence.
- A material recipe.
- A simulation result.
- A wet-lab test plan.
- A patent search lead.
None of that is revenue by itself.
The company begins when a buyer has a reason to pay before the founder runs out of money.
Why The Model Is Not The Company
The 2024 Nobel Prize in Chemistry for computational protein design and protein structure prediction made protein AI visible far beyond research circles. David Baker was awarded for computational protein design, while Demis Hassabis and John Jumper were awarded for protein structure prediction.
That prize matters.
It also creates a dangerous founder illusion: if the science is now famous, the business must be near.
No.
Drug candidates still face biology, toxicity, dosage, trial design, patient access, regulator questions, payer logic, manufacturing, patents and sales. Materials still face synthesis, yield, purity, supply chain, testing standards, buyer qualification and factory fit.
AlphaFold 3 expanded structure prediction beyond proteins into interactions among proteins, DNA, RNA, ligands and other molecules. That is a serious research tool. It is also not a magic cash machine.
The founder question is painfully plain:
Who pays for the next proof point, and why now?
If the answer is "investors will understand the science," you may have a research story, not a business.
The Founder Filter Table
Use this before building, pitching or applying for public money.
Pharma R&D team, biotech partner, CRO
Target logic, binding data, toxicity screen, patent room
Treating a generated molecule as a finished asset
Biotech team, diagnostics firm, research lab
Structure, affinity, selectivity, wet-lab repeat tests
Showing model confidence without bench evidence
Chemicals firm, food tech team, industrial R&D
Activity, stability, yield, process fit
Ignoring manufacturing cost per batch
Pharma team, clinical biotech, university spinout
Mechanism, safety signals, assay data, IP position
Pitching therapy before proof of biology
Medtech firm, device maker, materials lab
Biocompatibility, strength, ageing, sterilisation path
Forgetting qualification cycles
Energy company, cell maker, materials supplier
Stability, cycle life, manufacturability, raw material access
Assuming lab performance survives production
Chemicals plant, climate tech buyer, industrial partner
Conversion rate, durability, feedstock fit, recovery cost
Selling science without plant economics
Chip supplier, photonics lab, hardware company
Purity, process window, device fit, supply chain path
Underestimating buyer qualification time
CRO, university lab, pharma team, materials lab
Clean records, access control, repeatable audit trail
Building a dashboard nobody trusts
Deep tech founder, grant-backed lab, spinout
Source trail, lab record, IP note, buyer memo
Writing documents that do not move a sale
The table has one job: force the molecule back into a buying system.
If you cannot name the buyer, proof and trap, pause.
The Boring Layer Is Where Revenue Often Starts
Most founders want to own the miracle.
Bootstrappers should look at the layers around the miracle.
There are sellable wedges around AI-designed molecules:
- Literature and patent scans for one therapeutic or material niche.
- Lab data cleanup for assay or synthesis records.
- Wet-lab validation planning for AI-generated candidates.
- IP risk notes for generated molecules and CAD-like design files.
- Buyer-ready proof folders for grant-backed science teams.
- Screening workflow setup for one repeated lab process.
- Manufacturing readiness checks for one material class.
- Partner search and due diligence support for university spinouts.
This is where my CADChain experience matters. In hard technology, IP, access rights, version history and technical proof are not decorations. They are part of whether anyone dares to work with you.
The founder who sells the evidence layer may reach cash before the founder promising to own a whole drug pipeline.
Less glamorous, yes.
More survivable, also yes.
Drug Discovery Still Has Regulation, Trials And Buyers
The European Medicines Agency’s reflection paper on AI in the medicinal product lifecycle covers AI use across drug discovery, non-clinical work, clinical trials, manufacturing and post-authorisation activity.
Read that sentence slowly.
Drug discovery is one part of the path.
AI-designed drugs still need:
- A disease area with real unmet need.
- A target with biological support.
- A candidate that can be made and tested.
- Assays that answer the right question.
- Toxicity and safety work.
- A regulator-facing evidence trail.
- Partners who can run trials or pay for assets.
- A reimbursement or buyer path after approval.
This is why computational biology startups cannot pretend biology has become optional. Models can make search faster, but biology still gets a vote.
For bootstrapped founders, the honest path may be a tool, dataset, service or workflow that helps pharma, biotech, CROs or spinouts reduce one painful step.
Trying to become a full drug company on a tiny budget is brave.
It can also be financial self-harm.
Protein Design Needs Wet-Lab Proof, Not Hero Slides
Protein design is tempting because it feels closer to software than normal biology. You define a function, generate a sequence, check a predicted structure and show a beautiful 3D image.
Here is why that is dangerous.
Proteins live in messy systems. They fold, misfold, degrade, bind off-target, trigger immune reactions, lose activity, break during production or behave differently outside the model’s assumptions.
The American Chemical Society’s summary of the 2024 chemistry Nobel explains why protein design and structure prediction changed what scientists can attempt. The founder lesson is not "skip the lab." The lesson is "ask better questions before the lab eats the budget."
A protein startup should be able to answer:
- What function are we designing for?
- Which assay proves that function?
- Who owns the data and sequence rights?
- What failure mode kills the idea fastest?
- Who pays for validation?
- Which partner can manufacture or test it?
- What evidence would make a buyer lean in?
If you sell to research teams, they may pay for speed, search, assay planning or data handling.
If you sell to pharma, they will ask about repeatability, IP, safety and fit with their existing pipeline.
Either way, the 3D image is the invitation, not the invoice.
Materials AI Has A Different Revenue Path
AI-designed materials can be easier to explain than AI-designed drugs because the buyer may be an industrial company with a material problem: better battery life, lower heat, stronger coating, lighter alloy, new catalyst, cheaper solar material or cleaner process.
It is still slow.
Google DeepMind’s GNoME materials work predicted 2.2 million new crystals and highlighted 380,000 stable candidates. Berkeley Lab also described Google DeepMind adding nearly 400,000 compounds to the Materials Project, a research resource for materials data.
Those numbers are huge.
They do not remove the old industrial question:
Can we make this material, test it, source it, qualify it and use it in a real product?
The Berkeley A-Lab materials synthesis work is a useful signal because it connects AI-guided prediction to robotic synthesis. A-Lab and GNoME were used together to synthesize 41 new inorganic materials in 17 days.
That is the kind of stack founders should study:
- Prediction.
- Synthesis.
- Characterisation.
- Lab record.
- Selection loop.
- Human review.
- Buyer fit.
If you are a bootstrapper, you probably do not start by building all of that.
You start with the part one buyer can pay for.
Manufacturing Is Where Fantasy Goes To Die
AI-designed molecules look cheap inside a model.
Manufacturing can make them expensive again.
A drug candidate may be hard to synthesize, unstable, costly to purify or painful to formulate. A protein may need a production host, purification path, storage rules and batch testing. A material may need rare inputs, high temperatures, tight process control, clean-room conditions or a supply chain that does not exist near your buyer.
This is why founders should talk to manufacturing people early.
Ask them:
- What would make this impossible to make?
- Which input cost can kill the business?
- Which purity rule matters?
- Which batch test takes the longest?
- Which equipment is already available?
- Which process change would a buyer refuse?
- What makes this hard to qualify?
Do not wait until after a pitch deck wins compliments.
Manufacturing people have a special talent for making founder fantasies very quiet.
Where Bootstrapped Founders Can Enter
If you do not have a lab, a pharma partner or years of runway, you can still build in this space. You just need to choose a narrow service or tool that reaches paid proof faster.
Possible entry paths:
- A search workflow for AI-designed candidates in one disease or material niche.
- A data room for AI-generated molecule evidence.
- A wet-lab partner matching service for one research field.
- A patent and prior-art scan for generated proteins or materials.
- A screening plan builder for small biotech teams.
- A manufacturing-readiness checklist for AI-designed materials.
- A grant-to-buyer proof package for deep tech spinouts.
- A lab record and audit trail tool for AI-planned experiments.
- A buyer memo service for scientific founders who cannot explain the commercial path.
The F/MS AI workshop makes a point I use often: AI and automation help only when the founder knows the workflow, the inputs, the checks and the cost logic. In science markets, that rule becomes stricter because weak checks can waste lab time or create safety risk.
The F/MS Startup Game is built around moving from problem to first customer. AI-designed drugs and materials need the same discipline. Do not ask whether the molecule is cool. Ask who pays for the next proof.
The Grant Can Help, But It Can Also Hide Weak Demand
AI-designed drugs, proteins and materials often need public money, university partners, research grants and corporate pilots.
That is normal.
Deep tech is expensive.
The trap is letting the grant become the customer.
The Mean CEO guide to public-private funding for European deep tech fits this topic because molecule and materials startups may need a funding stack: grants for technical proof, corporate pilots for buyer access, private capital for speed and revenue for clean market evidence.
Before applying, write this sentence:
This money will help us prove one asset for one buyer type through one paid or partner-backed next step.
If you cannot fill it in, you may be about to create paperwork instead of a company.
The Evidence Pack Buyers Need
In this category, buyers need proof they can inspect.
A serious evidence pack should include:
- Problem and buyer definition.
- Target, molecule, protein or material description.
- Source data list.
- Model method and limits.
- Wet-lab test plan.
- Assay or synthesis results.
- Failed candidates and why they failed.
- IP notes.
- Safety and ethics notes.
- Manufacturing assumptions.
- Partner conversations.
- Buyer questions and objections.
- Next proof point with budget.
If the work touches regulated AI or regulated health markets, the evidence trail matters even more. That is where AI governance tools for audit trails become relevant, because science buyers do not want vibes. They want receipts.
This is also why lab robotics and autonomous experimentation belongs in the same topic cluster. Robots, models and labs create data. A company needs to know which data can be trusted, repeated and shown to a buyer.
The Founder SOP Before You Build
Use this simple operating flow before you start building an AI-designed molecule startup.
Choose drug candidate, protein, enzyme, catalyst, battery material, polymer, coating or lab data layer. Do not pitch the whole scientific universe.
Write the exact buyer type: pharma licensing team, CRO, materials lab, battery firm, diagnostic company, university spinout, industrial R&D head or grant-backed founder.
Choose the smallest proof someone may fund: assay plan, screening report, data pack, synthesis test, IP note, manufacturing review or partner memo.
Biology, toxicity, yield, stability, purity, patents, regulation, safety, data rights, buyer trust, manufacturing and cash.
Ask what they would pay for, what they would never trust and what proof makes them move to the next call.
Start with the workflow that creates proof, not the grand platform.
Every failed molecule, assay, material recipe or buyer call teaches the next move. Do not delete the ugly data.
If nobody pays for a report, service, screen, dataset or pilot, the problem may be admiration without demand.
Mistakes To Avoid
- Pitching AI-designed drugs as if trials disappeared.
- Treating protein structure as proof of therapeutic value.
- Showing material candidates without synthesis and buyer fit.
- Hiding behind grants because buyers are hard to reach.
- Ignoring IP until a partner asks.
- Building a platform when a paid service would teach faster.
- Hiring a wet-lab team before the buyer path is clear.
- Using open data without checking rights and provenance.
- Assuming pharma will buy because the science is beautiful.
- Confusing scientific novelty with startup timing.
The expensive mistake is not being wrong.
The expensive mistake is being wrong with a payroll, a lab lease and no buyer.
What To Do This Week
If you are tempted by AI-designed drugs, proteins or materials, do this before touching the product plan:
- Write a one-page buyer memo.
- Pick one asset type and one buyer.
- Ask three scientists what proof they would distrust.
- Ask three commercial people what they would pay for.
- Map the IP risk before sharing files.
- Find one lab partner and ask what validation costs.
- Find one manufacturing person and ask what kills the idea.
- Write a grant plan only after the buyer memo is honest.
- Set a cash stop line before the science becomes addictive.
If personalized medicine startups is your next rabbit hole, keep the same filter: patient benefit, reimbursement, clinical evidence and access matter more than a premium health demo.
Science deserves ambition.
Your bank account deserves discipline.
The Bottom Line
AI-designed drugs, proteins and materials are real. They may change medicine, energy, manufacturing, climate tech and scientific research.
But a startup is not paid for "real."
It is paid for proof that a buyer trusts.
If you are a bootstrapped founder in Europe, do not copy the fantasy version of AI drug discovery. Build the narrow wedge, protect the IP, document the evidence, find the buyer, price the next proof and keep enough cash to survive the slow parts.
Pretty lab slides do not pay.
Paid proof does.
FAQ
What are AI-designed drugs?
AI-designed drugs are drug candidates where software helps identify targets, generate molecules, predict binding, screen toxicity risks or plan experiments. The AI part can make search faster and cheaper, but the candidate still needs lab testing, safety work, regulation review, manufacturing logic, patent room, partners and buyers. A founder should treat the AI output as a starting asset, not as a finished company.
How are AI-designed proteins different from AI-designed drugs?
AI-designed proteins focus on amino acid sequences and protein structures. They may become enzymes, binders, antibodies, vaccines, sensors or protein materials. AI-designed drugs often focus on molecules that affect a biological target. Both need wet-lab proof, but protein design has its own issues: folding, activity, stability, immune response, production and purification. The founder still needs a paying user or partner.
What are AI-designed materials?
AI-designed materials are materials proposed or ranked by AI systems for properties such as stability, conductivity, strength, heat tolerance, catalytic activity or battery performance. They can include crystals, polymers, catalysts, coatings, semiconductors and battery materials. The hard test is whether the material can be synthesized, measured, qualified, sourced and used by an industrial buyer.
Can a startup build AI-designed molecules without a wet lab?
Yes, but only if the startup sells the right layer. A no-lab founder can build software, data workflows, search tools, IP support, partner matching, validation planning, evidence packs or buyer memos. If the company claims to own a drug, protein or material asset, it will still need wet-lab partners and credible testing. No lab does not mean no proof.
Why do AI-designed drug startups still need regulation?
AI-designed drug startups need regulation because patients, clinicians, payers and authorities need safety and efficacy evidence. A model can suggest a candidate, but it does not prove that the candidate works in a body, can be dosed safely, can be manufactured reliably or should be approved. Founders should plan the evidence trail early so regulatory questions do not arrive after money has already been spent.
What proof do investors and pharma buyers want?
Investors and pharma buyers usually want target logic, lab data, repeatable assays, toxicity signals, patent room, manufacturing assumptions, founder credibility, partner access and a clear next proof point. They also want to know who pays for each stage. A vague story about AI speed is weak. A narrow asset with evidence, costs, risks and buyer interest is much easier to discuss.
How can bootstrapped founders enter AI drug discovery?
Bootstrapped founders can enter through narrow services and tools around the drug discovery process. Good wedges include literature search, patent review, assay planning, data room setup, CRO matching, target reports, lab record tools, evidence packs and buyer memos. The aim is to sell something useful before attempting to own a full therapeutic pipeline.
How do AI-designed materials become commercial products?
AI-designed materials become products when they move from prediction to synthesis, testing, qualification, manufacturing and buyer use. An industrial buyer needs performance data, process fit, input cost, supply chain clarity and failure history. A founder should choose one buyer problem, then prove that the material can meet that buyer’s conditions without breaking the budget.
What mistakes should founders avoid in AI-designed drugs and materials?
Founders should avoid pitching model outputs as finished assets, ignoring lab validation, skipping IP checks, building broad platforms too early, hiring before buyer demand is visible, depending on grants as oxygen and treating scientific admiration as sales proof. The best early move is usually a narrow proof workflow that a buyer or partner will pay for.
How should founders price an AI-designed molecule startup offer?
Price the offer around the proof you can deliver now. That may be a fixed-fee evidence pack, target report, lab validation plan, partner search, IP scan, screening workflow or pilot package. Do not price against the fantasy value of a future drug. Price against the buyer’s current cost of delay, uncertainty, missed candidates, messy records or wasted lab work.