David Friedberg on AI x Biology: Where Founders Should Build Next in SynBio and Biomanufacturing

‍

David Friedberg has pushed many of us to ask a simple question that hides a hard truth: where should founders actually build next in AI and biology to create real value now, not in ten years.

I wrote this roadmap to answer that question with practical steps founders can use in synthetic biology and biomanufacturing today.

You will get a no-nonsense view of data, lab automation, machine learning in biotech, and the exact choke points where a startup can win.

I will also show how to sequence your work from first demo to GMP, and how to keep capital efficiency while you scale.

‍

Why AI x Biology Is Having Its David Friedberg Moment

I see founders riding a wave that finally has enough data, compute, and automation to matter.

Machine learning models are better at structure, sequence, and process prediction, but the bottleneck is still getting high-quality experimental data fast.

That is where lab automation and smart experimental design turn AI from a deck slide into a product.

I think of this moment as applied biology’s cloud era, where the winners will ship workflows, not just science.

When people reference David Friedberg, they often mean a focus on systems thinking, unit economics, and building full-stack solutions across agriculture, food, materials, and climate.

Those are the markets where AI and biology can compound and create defensible moats.

For more on founder operating systems and venture readiness, see our blog post: Capitaly.vc Blog.

The Founder Roadmap: From Idea to GMP Biomanufacturing

I map every AI x biology company against five milestones.

This is the shortest path I know from idea to revenue-grade production.

• Milestone 0: Problem selection and TEA.
• Milestone 1: Design-Build-Test-Learn loop at micro scale.
• Milestone 2: Process and product proof at liter scale with PAT.
• Milestone 3: Pilot scale with digital twin and QbD.
• Milestone 4: GMP readiness and reliable economics at target scale.

I keep the bar simple at each step.

Can you prove a unit model, validate cost curves, and show repeatability.

Can you show data density per dollar and time-to-learning improving each month.

That is the heartbeat of an AI-first bio company.

What Datasets Actually Matter for Machine Learning in Biotech

I see many teams underestimating dataset design and overestimating off-the-shelf models.

Your edge is not just the model but the dataset you can generate repeatedly under controlled conditions.

• Sequence-function maps: Deep mutational scans, promoter libraries, enzyme variants.
• Process-parameter-response matrices: Fed-batch strategies, pH, DO, feed timing, and yield or titer.
• Multi-omics under perturbation: Transcriptomics, proteomics, metabolomics linked to flux and product quality.
• Analytical readouts: LC-MS, HPLC, Raman, NIR, and in-line sensors tied to ground truth assays.

I plan these datasets like products.

Each has a schema, metadata, controls, and a path to become a reusable asset.

Labeling quality and experimental confounders will make or break your model performance.

For more on structuring data for venture-grade progress, see our blog post: Capitaly.vc Blog.

Closing the Design-Build-Test-Learn Loop with Automation

I start with throughput and control, not fancy robots.

If you double your design cycles per week, you double your learning velocity.

Automation is about cutting human-in-the-loop variance and latency.

• Design: Generative design for sequences, pathways, and media formulations.
• Build: DNA assembly workflows, strain construction, CRISPR edits, and QA checks.
• Test: High-throughput screening with reliable assays and normalization.
• Learn: Active learning pipelines that update priors and suggest next experiments.

I aim for a 48-hour DBTL cycle for enzymes and 7 to 14 days for strain improvements early on.

That speed compounds into data density and better models.

Lab Automation That Pays For Itself in 12 Months

I recommend putting automation dollars where the hands are bleeding today.

Start with the top three repetitive steps causing delays or errors.

• Liquid handling: Plate prep, media, and assay setup.
• Colony picking: Remove manual bottlenecks that kill throughput.
• Sample tracking: Barcoding and automated logging into ELN or LIMS.

Measure ROI by time saved, variability reduced, and experiments per week.

A lean stack with an Opentrons, a plate reader, an incubator shaker, and barcode tracking can generate outsized returns if aligned with your DBTL loop.

Do not buy robots you cannot program or maintain.

The Case for Cloud Labs and Biofoundries

I use cloud labs to de-risk capex and to standardize workflows.

Cloud facilities like Emerald Cloud Lab and Strateos can compress time to data when I have clear protocols and acceptance criteria.

I still validate critical steps in-house to keep IP knowledge and edge cases close.

Think of cloud labs as elastic capacity for methods that are already locked.

When your process is still moving targets, in-house control often wins.

Foundation Models for Proteins, Pathways, and Cells

I treat protein language models, structure prediction, and diffusion models as senior interns, not oracles.

They are fast and creative, but they need a manager called the wet lab.

Use ESM-style embeddings, AlphaFold-style structure, and generative models to propose candidates.

Then couple that to rigorous screening and negative controls.

The same logic applies to metabolic pathway design and regulatory sequences.

Foundation models are multipliers when coupled with thoughtful priors and Bayesian updates.

Active Learning and Bayesian Optimization in the Wet Lab

I set up active learning early, even with small datasets.

A simple loop of propose, test, update, and repeat lowers the number of experiments needed to reach a target.

• Bayesian optimization: Tune media and process parameters with Gaussian processes and expected improvement.
• Bandit strategies: Allocate plates across promising variants and exploration.
• Closed-loop automation: Trigger next runs based on last run results in the same week.

Make the optimizer a product feature, not a research project.

Show investors you learn faster than peers, not just that you can run more plates.

Building a Bioprocess Digital Twin Early

I build a digital twin at the 1 to 5 liter stage with mass balances and kinetic models, and I update it weekly.

It does not need to be perfect, but it must be useful.

The twin informs feed strategies, oxygen transfer, and scale-down models.

Link it to your PAT signals like Raman and off-gas, and you will prevent failures later.

A rough but updated twin beats a glossy model that never touches reality.

From Bench to 10,000 L: Scale-Up Without Burn

I recommend a scale-up ladder that proves transferability before you order expensive steel.

• Step 1: 250 mL to 2 L runs to nail kinetics and control loops.
• Step 2: 5 L to 20 L to validate oxygen transfer and mixing.
• Step 3: 50 L to 200 L with PAT, fail-studies, and robustness mapping.
• Step 4: 500 L to 2,000 L at a pilot plant or CDMO with QbD baked in.

Use scale-down models to mimic large-vessel gradients and test worst cases.

Build a transfer package with clear CPPs, CQAs, and acceptable ranges.

That package is your passport to any CDMO or internal plant.

Quality by Design and PAT as a Software Product

I think of QbD as a live model with guardrails.

It is not a PDF.

Your software should map inputs, parameters, and outputs to product quality in real time.

Put alarms and playbooks on top of that map so operators know what to do when drift happens.

That is how you turn process understanding into margin.

Regulatory as a Feature: GxP from Day One

I do not bolt on GxP at the end.

I write SOPs, validation plans, and data integrity rules while the system is small.

Even if you are not pharma, food and materials customers care about traceability.

Build your ELN, LIMS, and MES with audit trails, user roles, and change control from day one.

This saves months at commercialization.

The Data Layer for Biology: ELN, LIMS, and MES Unbundled

I separate three layers so I can scale without replatforming.

• ELN: Human-readable context and decisions.
• LIMS: Sample, assay, and inventory tracking with schemas and barcodes.
• MES: Execution layer for processes, recipes, and equipment states.

I add a fourth layer that many skip.

That is the analytics and model hub that pulls from all three and pushes recommendations back.

Use open standards where possible like SiLA 2 to reduce custom glue code.

For more on systems that speed up fundraising diligence, see our blog post: Capitaly.vc Blog.

Hardware Choices: Opentrons vs Tecan vs Hamilton vs Custom

I pick gear based on software ecosystem, support, and the fit to my workflows, not just specs.

Opentrons is flexible and affordable for early loops.

Tecan and Hamilton shine for robust, validated high-throughput operations.

Custom rigs only make sense when your workflow is your moat and you can service it.

I avoid vendor lock-in by keeping protocols in version-controlled repositories and using adapters where possible.

Go-To-Market: Who Buys First and Why

I look for customers with a burning platform and a short time-to-value.

These are usually teams with revenue pressure or regulatory deadlines.

• Biopharma R&D: Enzyme engineering, cell-line optimization, and analytical automation.
• Industrial bio: Strain and process optimization for chemicals and materials.
• Food and ag: Protein functionalization, fermentation of novel ingredients, and crop trait development.

I lead with a pilot that has a clear ROI and a success metric tied to money.

That can be cost per variant screened, time to hit target titer, or reduced batch failure rate.

Business Models: Software, Services, or Cells-as-a-Service

I see three models that work, and hybrids among them.

• Software: Sell design tools, LIMS, or PAT analytics with clear integration paths.
• Services: Offer DBTL cycles or bioprocess optimization with SLAs and outcomes.
• Cells-as-a-Service: Provide engineered strains or enzymes with performance-based pricing.

Services are a bridge to product and a data engine if you treat metadata like gold.

Software multiplies margin and reach if you can standardize integrations.

Cells-as-a-Service gives you recurring revenue but requires careful IP and quality control.

Economic Moats: Data, Workflows, and Integration

I do not rely on novelty alone as a moat in synbio.

I build moats from data quality, proprietary workflows, and integrations that become switching costs.

• Data: High signal-to-noise datasets under known protocols and controls.
• Workflows: Protocols that compress cycle time and improve outcomes.
• Integration: Hardware and software stacks that work out of the box with validation.

When a customer says switching will cost them months of validation, you have a moat.

Document your edge like a product, not folklore.

Team Design: Dry Lab Meets Wet Lab

I staff small, cross-functional squads aligned to outcomes.

Each squad owns a loop from design through test and analysis.

Put an ML engineer next to an automation scientist and a fermentation lead.

Give them a weekly cadence and a dashboard tied to milestones.

Hire for people who automate themselves out of a job and then go find the next bottleneck.

Capital Efficiency: A Practical Budget for the First 18 Months

I build budgets around learning velocity, not vanity equipment.

Here is a rough blueprint I have used.

• People first: Five to eight core hires covering automation, ML, fermentation, and analytics.
• Core gear: One liquid handler, incubator shakers, a plate reader, and essential analytics.
• Cloud capacity: Use cloud labs and CDMOs for spikes and confirmatory runs.
• Software: Start with ELN and LIMS that you can script, and add MES later.

Track burn per learning milestone, not just runway in months.

If your cost per unit learning drops each quarter, you are on the right path.

For more on capital planning and investor expectations, see our blog post: Capitaly.vc Blog.

Where I Would Build Next in SynBio and Biomanufacturing

I would build in three areas that feel under-served and near-term valuable.

• Process intelligence for mid-scale fermentation: A plug-in PAT and optimization layer for 50 to 2,000 L that reduces batch failures and speeds tech transfer.
• Sequence-to-function for industrial enzymes with guaranteed cycles: A DBTL service with SLAs that hits performance targets and returns data products to clients.
• Data layer for regulated bio manufacturing: A unified LIMS to MES bridge with audit-ready analytics and digital twin support.

Each of these has clear buyers, measurable ROI, and a path to defensible data assets.

They align with the practical mindset I associate with the best AI x biology founders.

How to Pick Your First Product Like an Operator

I pick the first wedge by using two filters.

High pain and low adoption friction.

I validate with five design partner calls and one paid pilot before I start building.

If I cannot define a single number that improves within 90 days, I do not ship it.

Examples include cutting analytical turnaround by 50 percent or improving titer by 20 percent in six weeks.

Designing the Right Assays to Power Machine Learning in Biotech

I bias toward assays with high repeatability, strong dynamic range, and automation potential.

If the assay is noisy, the model will be blind.

I run gage R and control charts even in research to spot drift early.

I standardize sample prep and normalization and I annotate everything that could confound learning.

That discipline compounds into model performance.

IP, Data Rights, and Vendor Lock-In

I negotiate data and IP rights upfront with any CDMO, cloud lab, or instrumentation vendor.

Your assay outputs, metadata, and derived models should be yours to use and improve.

I avoid proprietary file formats that lock me into a tool.

Export to open schemas and mirror your data to your own warehouse.

Long-term leverage comes from portability.

De-Risking Tech Transfer with Playbooks and Digital Twins

I write tech transfer playbooks that an external team can follow without tribal knowledge.

They include CPPs, CQAs, sampling plans, out-of-spec protocols, and contact trees.

I attach a simplified digital twin and a control strategy page that explains why each parameter matters.

This is how I reduce transfer cycles and build credibility with partners.

Financing Strategy: Milestones Investors Will Pay For

I time raises around milestones that are obvious on one slide.

Pre-seed is about proving your loop and getting first performance gains.

Seed is about model lift, scale-down validation, and one or two paying pilots.

Series A is about pilot scale economics, QbD, and a pipeline of commercial contracts.

I tell the story in cost per learning and cost per unit improvement because it maps to value creation.

For more on fundraising narratives in deep tech, see our blog post: Capitaly.vc Blog.

Hiring the First Ten People

I hire builders who have shipped in messy environments.

I probe for people who wrote their own scripts, automated their job, and taught others.

My first ten include a lead automation scientist, an ML engineer, a fermentation lead, a data engineer, and a scrappy operations owner.

I pick athletes who can cover two positions each for the first year.

Culture is weekly demos, blameless postmortems, and relentless documentation.

Storytelling for Customers and Regulators

I keep the story simple.

We make biology faster, cheaper, and more reliable, and here is the proof.

For regulators, I lead with control strategies and data integrity.

For customers, I lead with time-to-value and risk reduction.

One slide per audience is all you need when the work is real.

Common Failure Modes and How to Avoid Them

I see five patterns that waste time and burn cash.

• Fancy models with bad data: Fix the assay and metadata first.
• Robots without workflows: Start with the bottleneck and write SOPs you can run.
• Skipping PAT: You cannot fix what you cannot see in real time.
• No digital twin: You will relearn the same lessons at each scale.
• Weak integration: Glue code breaks under GxP and customer pressure.

When in doubt, trim scope and tighten loops.

Your cadence is your moat.

FAQs

What is the difference between synthetic biology and biomanufacturing

Synthetic biology is the design of biological systems and parts, while biomanufacturing is the scaled production using those systems.

Think of synbio as design and build, and biomanufacturing as produce and control.

How much data do I need before machine learning helps

You can get value with hundreds to low thousands of labeled examples if your assay is clean and you use active learning.

The right priors beat raw volume early on.

What is the fastest way to get a DBTL loop running

Automate one high-impact bottleneck, standardize one robust assay, and ship a simple active learning loop.

Expand only after you see weekly gains.

Do I need GMP for non-pharma products

No, but GxP principles help with traceability, reliability, and customer trust.

Build the bones early and scale the formality later.

How do I choose between cloud labs and in-house automation

Use cloud labs for standardized, scalable methods and burst capacity.

Keep variable, IP-critical, or evolving workflows in-house.

What tools should I buy first for lab automation

Start with a liquid handler, reliable incubator shakers, a plate reader, and barcode-based tracking.

Add colony picking and analytics as throughput grows.

How do I protect my IP and data when using vendors

Negotiate ownership of raw data, metadata, and derived models.

Export data to open formats and mirror to your own storage.

How do I model unit economics at pilot scale

Combine technoeconomic analysis with pilot data on titer, rate, yield, media cost, and batch success probability.

Build sensitivity ranges and update monthly.

How do I pitch investors on AI x biology

Lead with learning velocity, repeatability, and a clear line from data to dollars.

Show a timeline that reduces risk at each stage with evidence.

What risks kill most synbio startups

Assay drift, unscalable workflows, lack of integration, and underestimating scale-up complexity.

Fix those with PAT, digital twins, and disciplined process design.

Conclusion

The next generation of AI and biology companies will win by running tighter loops, owning their data, and delivering measurable outcomes across design and manufacturing.

Use this roadmap to pick a wedge, build your DBTL engine, and scale with process intelligence as your product.

If you lead with clarity on datasets, automation, and unit economics, you will separate yourself from the noise.

That is the practical path I see for founders in synthetic biology and biomanufacturing who are inspired by the operator mindset often associated with David Friedberg.

Subscribe to Capitaly.vc Substack (https://capitaly.substack.com/) to raise capital at the speed of AI.