Discover
Data in Biotech
Data in Biotech
Author: CorrDyn
Subscribed: 30Played: 360Subscribe
Share
© 2023 CorrDyn
Description
Data in Biotech is a fortnightly podcast exploring how companies leverage data to drive innovation in life sciences.
Every two weeks, Ross Katz, Principal and Data Science Lead at CorrDyn, sits down with an expert from the world of biotechnology to understand how they use data science to solve technical challenges, streamline operations, and further innovation in their business.
You can learn more about CorrDyn - an enterprise data specialist that enables excellent companies to make smarter strategic decisions - at www.corrdyn.com
82 Episodes
Reverse
Protein engineering pays off only when your model ventures past its training data without proposing sequences that can't fold.
Every data point costs hundreds to tens of thousands of dollars and weeks of lab time. One bad plate can burn half a million dollars and set a program back three months. So how do you get one costly experiment to teach your model as much as possible?
Stef van Grieken is co-founder and CEO of Cradle, which designs antibodies and enzymes for scientists to test in their own labs. At Google he built infrastructure to train and serve large models, including those that supported AlphaFold. Cradle now works with ten of the top 25 global biopharma companies on 80+ active programs.
You'll hear how Cradle's generator and predictor models stay honest with each other, why training procedures matter more than architecture in biotech data science, and what raw instrument data reveals that a summary number hides. You'll also get a practical checklist for making assay data fit for machine learning.
This episode is for data leaders, ML engineers and protein scientists working on drug discovery and life sciences data pipelines. Expect technical detail on active learning, multi-property optimization and preference models.
Key Takeaways
Cradle's generator never sees measured values. Stef explains why letting it memorize what the scorer knows makes both models confidently wrong.A surprising share of reported KDs come from invalid SPR curves. Capturing on-rate, off-rate and fit quality changes what your model can trust.Some teams set objectives below their instrument's detection limit, so the model ends up optimizing noise.Separate predictors for binding and stability can't trade off against each other. One model with multiple heads makes those trade-offs explicit before the lab does.
Chapter Markers
00:00 Exploring beyond the training data
01:30 Stef's path from Google to Cradle
04:06 Paying the first customer for a case study
06:08 How a scientist runs a Cradle project
11:31 Debating plates and explicit trade-offs
14:53 Why Transformers fit protein sequences
17:08 Training procedures in a data-poor field
19:18 EvoFlow and the three-step training process
26:06 Bringing historical pharma data into custom models
29:18 Why raw SPR curves beat a single KD
34:17 Diversification, multi-property optimization, active learning
40:23 Case study: cross-reactive snake antivenom antibody
44:51 How Cradle uses its internal wet lab
51:49 Cradle's per-project SaaS business model
54:36 Build versus buy for pharma ML teams
58:30 What good biopharma data practice looks like
62:25 Two-person biotechs in five years
66:57 Ross's three takeaways
Useful Links & Resources
Cradle: https://www.cradle.bio
Connect With the Show
Ross Katz on LinkedIn: https://www.linkedin.com/in/b-ross-katz/CorrDyn on LinkedIn: https://www.linkedin.com/company/corrdyn/CorrDyn: https://corrdyn.com
Is your lab capturing full binding curves, or just the KD? Tell us in the comments how your team decides what assay data gets stored.
Visit corrdyn.com to see how CorrDyn can help your organization extract value from data. And subscribe to Data in Biotech so you don't miss an episode.
#DataInBiotech #ProteinEngineering #MachineLearning #DrugDiscovery #CorrDyn
Four rival pharma companies build AI startups together. The science is rarely what decides which ones succeed.
You've built a model that works on public data. Now pharma wants a security review, a test you can't design yourself, and proof your tool fits a program they already run.
Yair Benita is CTO of AION Labs, a venture studio owned by Teva, Merck, Pfizer, AstraZeneca and AWS that builds AI drug discovery companies from scratch. His career runs from immuno-oncology at Merck in the early Keytruda days to immune checkpoint discovery at Compugen and patient data work at CytoReason.
Ross Katz and Yair break down the barriers between good technology and a pharma team willing to use it. You'll hear how blind validation won over skeptical chemists, why a design tool sells where a prediction stalls, and how a year of compliance work became a one-click AWS deployment.
Key Takeaways
A startup's model ranked a pharma compound library it had never seen, against targets it didn't know. The chemists hoped for two binders in the top 100. The results kept them testing for ten targets.Telling a team its CRISPR guide is toxic after months and $1 million of work sends them back to the start. Yair's rule: if you can predict it, you can design it.Splitting antibody discovery and optimization into two companies made scientific sense. Commercially, pharma wanted the whole step.If pharma partners can't agree on a nonclinical test for success, don't start the company. That's why AION Labs never entered drug combination.
Chapter Markers
00:00 Introduction: Rival pharma companies building AI companies together
01:30 From pharmacology to computational biology
03:13 Merck and Compugen: discovering novel immune checkpoints from genome structure
07:00 Go it alone or reveal early? Lessons on proving a computational discovery
11:14 Cytoreason: deconvolving cell composition to solve pharma problems
15:18 Start with low-hanging fruit: indication selection vs. target discovery
17:30 Business models: why tech-for-pharma is hard, and why that is changing
19:58 What AION Labs is and how four pharma companies and AWS built it
23:10 Limitations of the model: shared challenges and the data-for-training problem
27:13 Blind validation: the small molecule ranking story
32:30 Baby steps, blinded competitions and building pharma trust
35:39 How seed-stage startups innovate on limited data and funding
41:54 Lesson: predictions are useless without a fix (Cassidy Bio)
44:00 Lesson: pharma wants the full process, not a piece (antibody discovery and optimization)
46:29 Defining the problem and the test: why no drug combination company was started
51:00 Eon’s evolving model: access to pharma, co-development and funding partners
54:40 The compliance bottleneck: ISO 27001 and one-click compliant AWS
58:07 Vision: an end-to-end ecosystem of specialized AI companies
01:00:27 Who should reach out to AION Labs
01:01:24 Four barriers of data, product, proof and access
Useful Links & Resources
AION Labs: https://www.aionlabs.comCASP (Critical Assessment of Structure Prediction): https://predictioncenter.orgISO/IEC 27001 information security standard: https://www.iso.org/standard/27001AlphaFold Protein Structure Database: https://alphafold.ebi.ac.ukYair Benita on LinkedIn: https://www.linkedin.com/in/ybenita/
Connect With the Show
Ross Katz on LinkedIn: https://www.linkedin.com/in/b-ross-katz/CorrDyn on LinkedIn: https://www.linkedin.com/company/corrdyn/CorrDyn website: https://www.corrdyn.com
Which barrier has slowed your team most: data access, proof, product fit, or security review? Tell us in the comments. If you've run a blind validation with a pharma partner, we'd like to hear how you set it up.
Visit corrdyn.com to see how CorrDyn can help your organization extract value from data.
#DataInBiotech #DrugDiscovery #BiotechDataScience #MachineLearning #LifeSciences
AI can design a molecule in seconds, however someone still has to build it and test it before it means anything.
You're pouring resources into AI-driven molecule design, but nobody on your team has answered the harder question: can you actually make what the model just designed?
Patrick Finn is President and Chief Operating Officer of Twist Bioscience, a company writing DNA on silicon chips at a scale of roughly 32 million distinct oligonucleotides a day. Under his leadership, Twist has become the manufacturing layer behind antibody discovery, liquid biopsy, and molecular residual disease testing programs across biotech.
You'll get a clear picture of what happens after an AI model produces a candidate molecule, and why the synthesis step decides whether that design ever becomes usable data. Paddy breaks down how pooled DNA synthesis and display selection turn hundreds of thousands of candidates into model feedback in weeks instead of months, and where the manufacturability gap bites hardest as AI pushes sequences into territory nature never explored.
This episode covers high-throughput DNA synthesis, target enrichment for liquid biopsy, tumor-informed molecular residual disease testing, and the economics of scaling antibody discovery. It's built for data and R&D leaders investing in AI-driven drug discovery who haven't stress-tested the physical layer underneath their models, plus anyone in biotech data science curious how life sciences data pipelines actually generate their evidence.
Key Takeaways
Twist's silicon chip platform pushes roughly 32 million oligonucleotides through synthesis every day, and that scale is what turns a single AI-designed antibody into a dataset instead of a guess.Liquid biopsy tests need sequencing depth around 50,000x coverage to catch a signal in the noise, and target enrichment panels can cut total sequencing costs by roughly half.AI models are pushing sequence designs into territory that's genuinely hard to manufacture: homopolymers, long repeats, extreme GC content. Twist's acceptance rate on submitted sequences climbed from 96% to 99.5%, and that gap matters more than it sounds.Memorial Sloan Kettering sent Twist 100,000 antibody designs and had the entire set synthesized as a single pooled library, tested through phage or yeast display, instead of one sequence at a time.
Chapter Markers
00:00 Introduction to Twist Bioscience and Paddy Finn
01:12 Overview of Twist's silicon DNA synthesis platform
04:13 How miniaturized chips drive 32 million oligos daily
05:35 Killer applications built on Twist's infrastructure
07:44 Liquid biopsy and the cost of sequencing depth
12:02 How target enrichment lowers diagnostic costs
13:15 Tumor-informed molecular residual disease testing
16:30 What "disease-free" really means in MRD testing
18:02 Producing and characterizing antibodies at scale
22:25 The assay menu behind antibody developability
24:04 Why AI-designed sequences are hard to manufacture
27:36 Predicting manufacturability before synthesis begins
31:16 Going from 96% to 99.5% sequence acceptance
33:16 Wet lab partnership with Amazon and Memorial Sloan Kettering
36:10 Why Twist stays an enabler, not a model builder
38:20 Is AI-driven biology durable, or a short cycle
41:53 What R&D leaders should ask before scaling AI design
43:00 Ross's takeaways: bottlenecks move, they don't disappear
Useful Links & Resources
Twist Bioscience: https://www.twistbioscience.comPatrick Finn on LinkedIn: https://www.linkedin.com/in/patrickjfinn/CorrDyn: https://www.corrdyn.comRoss Katz on LinkedIn: https://www.linkedin.com/in/b-ross-katz/
#DataInBiotech #BiotechDataScience #DrugDiscovery #LifeSciences #AIinBiotech
Flow cytometry can find a cell population in seconds, but naming it consistently across labs is a problem still unsolved.
If you've ever tried to compare cytometry data across studies, CROs, or even two scientists in the same lab, you know the frustration: everyone gates their own way, and the same cell population ends up with three different definitions. That inconsistency is quietly limiting what AI can do with biotech data.
Ryan Brinkman is VP Research Director of Flow Cytometry Bioinformatics at Dotmatics and Founding Director of SOULCAP, the Standard Ontology for Unambiguous Labeling in Cytometry and Phenotyping, built after years as an academic developing automated gating tools that kept running into the same naming problem. He's joined by Brian Wile, who is the General Manager of Flow Cytometry at KCAS Bio, a CRO that feels the cost of inconsistent labeling in client projects every day.
You'll get a clear picture of how flow cytometry data moves from raw signal to labeled cell population, why that last step has resisted automation, and what a shared standard could unlock for machine learning models trained on this data. Ryan and Brian break down the gap between automated gating and consistent labeling, and why agreement, not just data volume, is what AI in biotech actually needs.
This episode covers the mechanics of flow cytometry, an EVE Online citizen science project that trained a gating algorithm on hundreds of millions of human-labeled plots, and why cell population names like "Treg" or "natural killer cell" don't map to one agreed set of markers.
Key Takeaways
Gating automation solved half the problem: computers can now draw boundaries around cell populations, but nothing proves which name belongs on the result.A citizen science project turned hundreds of thousands of EVE Online players into an unlikely training set, collecting roughly half a billion labeled plots from people with zero biology background.The same cell type, like a regulatory T cell, gets defined by entirely different marker combinations depending on the lab or CRO, and those datasets can't simply be merged later.SOULCAP is running a Delphi-style consensus process among scientists worldwide to tie every cell population name to a specific, reproducible set of markers and experiments.
Chapter Markers
00:00 Why naming cells is harder than naming genes
02:13 What flow cytometry measures, from cell to signal
06:21 The scale and complexity of high-dimensional cytometry data
09:20 How raw data becomes a gated cell population
12:02 Why gating has stayed manual for decades
16:19 Clusters of differentiation and the marker explosion
18:11 Training an algorithm with EVE Online players
25:31 Evaluating accuracy without a gold standard
31:34 Why solving gating doesn't solve labeling
36:32 What genomics got right that cytometry hasn't
38:09 Two definitions of a regulatory T cell, one label
44:28 The cost of inconsistent labeling for CROs and pharma
47:43 Introducing SoCAP and the Delphi consensus process
57:18 What automated labeling could make possible
65:38 How to get involved in SOULCAP
Useful Links & Resources
Ryan Brinkman on LinkedIn: https://www.linkedin.com/in/ryan-brinkman-9bb1103/Brian Wile on LinkedIn: https://www.linkedin.com/in/brianwile/KCAS Bio: https://www.kcasbio.comCorrDyn: https://www.corrdyn.com
Connect With the Show
Ross Katz on LinkedIn: https://www.linkedin.com/in/b-ross-katz/CorrDyn LinkedIn: https://www.linkedin.com/company/corrdyn/
Where do you land on the naming problem? Have you had to merge cytometry datasets that turned out to use different definitions for the same cell type? Tell us about it in the comments.
Visit corrdyn.com to learn how CorrDyn can help your organization extract value from data.
#DataInBiotech #FlowCytometry #BiotechDataScience #Bioinformatics #LifeSciences
Ask a general AI model how many approved drugs hit a target, and it might tell you three when the real answer is six, sounding just as confident either way.
If your team is grounding drug discovery decisions in AI output with no way to trace where the answer came from, you're one regulator's question away from a very expensive problem.
Lisa Downey is CEO of DrugBank, a structured biomedical intelligence platform cited in more than 60,000 papers and used by nine of the top 20 global pharma companies. She previously built Clarivate's genomic and rare disease data business from the ground up and held leadership roles at GlobalData, giving her almost 20 years across healthcare and life sciences data.
Lisa breaks down why the bottleneck in AI-driven drug discovery has shifted from data scarcity to trustworthy grounding, and what that means for teams making target identification and go/no-go calls. You'll hear how DrugBank's knowledge graph separates causation from correlation, why reproducibility matters more than speed, and what questions to ask before building a reference data layer in-house.
This episode covers deterministic versus probabilistic data, human-in-the-loop versus human-over-the-loop curation, and how biopharma teams connect grounding layers to their AI agents through MCP. It's built for data and analytics leaders, R&D teams, and anyone deciding whether to build or buy their biomedical data infrastructure.
Key Takeaways
- A general model asked how many approved drugs hit PD-L1 will answer with total confidence, and total inaccuracy, missing half the real number without any signal that it's wrong.
- Anthropic's own benchmarks found frontier models pulling public genomic data got it right as little as 17% of the time, until a deterministic tool pushed accuracy past 99%.
- DrugBank moved from human-in-the-loop curation to human-over-the-loop oversight once its data was connected enough that one expert validating one relationship could cascade trust across dozens of related facts.
- Before building or buying a reference data layer, Lisa lays out four questions that separate real infrastructure from marketing, starting with whether every fact traces back to a source and a date.
Chapter Markers
00:00 Why data scarcity isn't the real bottleneck anymore
01:22 What drew Lisa to DrugBank's mission
03:04 What DrugBank is and who relies on it
05:03 The grounding layer: completeness and reproducibility
07:28 Anthropic's benchmark on data infrastructure
09:21 The high-stakes decisions DrugBank data informs
12:23 Where lost cycle time actually comes from
14:13 DrugBank versus homegrown knowledge graphs
19:43 Human-in-the-loop versus human-over-the-loop curation
24:12 How DrugBank checks its own data quality
25:37 Deterministic versus probabilistic data explained
28:52 The J&J case: separating causation from correlation
33:06 Connecting DrugBank to your AI stack via MCP
37:28 Four questions to ask before you build or buy
42:15 Where DrugBank fits, and where it doesn't
44:44 AI as an amplifier of both good and bad decisions
Useful Links & Resources
- Connect with Lisa Downey on LinkedIn (https://www.linkedin.com/in/lisaldowney/)
Connect With the Show
- Ross Katz on LinkedIn (https://www.linkedin.com/in/b-ross-katz/)
- CorrDyn on LinkedIn (https://www.linkedin.com/company/corrdyn/)
Have you run into an AI model giving you a confident, wrong answer in your own R&D work? Tell us about it in the comments, we're always looking for real examples for future episodes.
Visit corrdyn.com to learn how CorrDyn can help your organization extract value from data.
#DataInBiotech #BiotechAI #DrugDiscovery #DataScience #LifeSciences




