Question 1

What is AI-native drug discovery?

Nathan C. Frey · Accepted Answer

AI-native drug discovery integrates artificial intelligence as a core component from the start, using foundation models, active learning, and lab-in-the-loop automation to design and validate therapeutics autonomously rather than adding ML as a tool to traditional workflows.

Question 2

Where does AI actually deliver value in drug discovery?

Nathan C. Frey · Accepted Answer

AI delivers ROI by automating routine decision-making and data synthesis—hit prioritization, protein design optimization, retrosynthetic planning, and experimental management—not the headline-grabbing protein binding prediction that rarely blocks real campaigns.

Question 3

What is lab-in-the-loop machine learning?

Nathan C. Frey · Accepted Answer

Lab-in-the-loop ML creates closed-loop systems where experimental data directly informs model updates, enabling autonomous optimization through cycles of computational design, robotic synthesis, experimental validation, and model retraining.

Question 4

Should we invest in AI drug discovery?

Nathan C. Frey · Accepted Answer

Yes, if you focus on automating decision-making bottlenecks and data synthesis rather than headline-grabbing capabilities like binding prediction; ROI comes from accelerating existing workflows, not replacing medicinal chemistry expertise.

Question 5

What's the ROI of foundation models in pharma?

Nathan C. Frey · Accepted Answer

Foundation models deliver ROI through transfer learning that reduces training data requirements, embedding-based property prediction that accelerates screening, and naturalness scoring that improves candidate prioritization, not through de novo design alone.

Question 6

When should biotech companies build vs. buy AI capabilities?

Nathan C. Frey · Accepted Answer

Build when your therapeutic modality or target space requires custom models (novel antibody formats, rare protein families); buy or partner for general capabilities like retrosynthesis, standard antibody optimization, or small molecule property prediction.

Question 7

How do protein language models work?

Nathan C. Frey · Accepted Answer

Protein language models use transformer architectures trained on millions of protein sequences via masked language modeling to learn evolutionary patterns, structural constraints, and functional relationships encoded in sequence space without explicit structural supervision.

Question 8

What are the limitations of AlphaFold for antibody design?

Nathan C. Frey · Accepted Answer

AlphaFold struggles with CDR H3 loop prediction for sequences distant from PDB training data (RMSD ~4Å for novel sequences), predicts bound conformations by default when unbound structures are needed, and lacks accuracy for antibody-antigen interface modeling.

Question 9

How accurate is antibody structure prediction in 2025?

Nathan C. Frey · Accepted Answer

General models achieve ~2Å RMSD on sequences similar to PDB data but degrade to 4-6Å for novel CDR H3 loops; specialized antibody models (Ibex, IgFold) handle bound/unbound states better but still struggle with out-of-distribution sequences.

Question 10

What is discrete Walk-Jump Sampling?

Nathan C. Frey · Accepted Answer

Walk-Jump Sampling combines local exploration via discrete diffusion (walk steps) with global jumps that escape local optima, operating natively in discrete sequence space to generate high-quality, diverse protein sequences without continuous relaxation.

Question 11

How do I get started in BioML research?

Nathan C. Frey · Accepted Answer

Build strong fundamentals in CS, applied math, and ML (1-2 years), complete one meaningful research project (1 year), develop engineering discipline with documented code, demonstrate high agency by identifying problems independently, and network proactively before job searching.

Question 12

Should I do a PhD for BioML work?

Nathan C. Frey · Accepted Answer

A PhD helps for research-focused roles but isn't required; what matters is completing meaningful projects, demonstrated execution ability, strong engineering practices, and problem-solving agency—skills you can build through industry, self-study, or formal education.

Question 13

Which biotech companies are truly AI-native?

Nathan C. Frey · Accepted Answer

AI-native biotechs build discovery platforms where ML is core to operations from day one (autonomous design loops, continuous learning systems, data-first infrastructure) with value propositions impossible to articulate before ChatGPT (Nov 2022), typically built by teams under 30 people before scaling.

Question 14

Can established biotech companies become AI-native?

Nathan C. Frey · Accepted Answer

No—companies that scaled before November 2022 are structurally 'AI-encumbered' by organizational inertia, incompatible systems, and established hierarchies; true AI-native transformation requires rebuilding from foundations, not rebranding or adding computational teams.

Question 15

Why do most enterprise AI transformations fail?

Nathan C. Frey · Accepted Answer

MIT's NANDA initiative found 95% of enterprise GenAI pilots fail to increase revenue not due to AI technology limitations, but because of organizational learning gaps, incompatible existing systems, cultural resistance, and attempting to bolt AI onto established workflows rather than redesigning processes.

Question 16

What happened to the AI drug discovery hype?

Nathan C. Frey · Accepted Answer

Hype focused on solving wrong problems (binding prediction) that don't block real campaigns; value comes from unglamorous automation of routine decision-making, data synthesis, and process optimization that experienced drug hunters recognize as actual bottlenecks.

Question 17

How do you validate AI-designed antibodies?

Nathan C. Frey · Accepted Answer

Validation requires computational filtering (structure prediction, binding prediction, developability), experimental screening (binding affinity, specificity, expression), and functional assays (neutralization, cell-based assays) in lab-in-the-loop workflows that inform model retraining.

Nathan C. Frey, PhD

FAQ

AI Drug Discovery

Where does AI actually deliver value in drug discovery?

Should we invest in AI drug discovery?

How do I evaluate AI drug discovery vendors?

Foundation Models & Interpretability

How do you make protein AI models interpretable?

What are concept bottleneck models for proteins?

Protein Design & Generation

What is Walk-Jump Sampling?

How do you generate novel protein sequences?

How do you generate protein sequences and structures together?

What is CHEAP and PLAID?

Antibody Design

What are the limitations of AlphaFold for antibody design?

How accurate is antibody structure prediction?

Why is CDR H3 loop prediction so difficult?

Molecular Generation

How do you generate molecules quickly?

What are normalizing flows for molecular design?

Biomarker Discovery

How can machine learning find new biomarkers?

What are the advantages of long-read RNA sequencing?

AI-Native Companies

What is an AI-native company?

Can established companies become AI-native?

Why do most enterprise AI transformations fail?

Open Science & Publishing

Should biotech companies open-source their AI work?

Why is open science important in BioML?

What’s wrong with AI/ML academic publishing?

Careers

How do I get started in BioML?

What skills do I need for BioML roles?

Should I do a PhD for BioML work?

What do hiring managers look for?

Biotech Culture

Why does biotech struggle to attract top talent?

How do I evaluate biotech culture before joining?

What’s the difference between biotech and tech culture?