Aug 1, 2024 · 30m · a16z
AI in Pharmaceutical R&D with Kim Branson
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of Raising Health, host Vijay Pande interviews Kim Branson, SVP and Global Head of AI and Machine Learning at GSK, about the transformative role of artificial intelligence in pharmaceutical R&D. Branson discusses scaling AI within big pharma, strategies for biotech startups, active learning in target selection, and the future of computational biomarkers.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 3.5 Guest teaching 4.0 Guest disagreement 1.4 The host pushing back 0.4
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 18:45 Dismissing talent-only startup pitches
Kim bluntly shuts down a common startup pitch deck narrative, stating that offering smart people without proprietary data generation won't work because pharma already has smart talent internally.
Hardest push from the host ▶ 18:58 Pressing on data differentiation criteriaVijay immediately challenges Kim to define exactly what constitutes differentiated data in his mind after Kim dismisses talent-driven pitches.
Biggest teaching moment ▶ 16:20 High-dimensional data requirement breakdownKim educates the audience and host on the core paradox of modern biology: cheap measurement technologies create high-dimensional noise that human intuitive analysis cannot parse without ML.
The host holds their own ▶ 19:45 Vijay explaining single-layer neural net equivalence to logistic regressionVijay actively demonstrates his deep ML knowledge by pointing out that complex models often collapse back to logistic regression baselines, predicting the industry will come full circle.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Kim Branson's Journey into Computational Biology | 2 | 3 | 1 | 0 | Vijay asks a friendly introductory question about Kim's trajectory into computational biology. Kim explains his background in structural biology and early computational drug design on Relenza 25 years ago. | |
| Comparing Startup Agility with Big Pharma Scale | 3 | 3 | 1 | 0 | Vijay inquires about the contrast between startup agility and large pharma scale. Kim outlines the trade-offs of speed versus capital and explains what convinced him to join GSK. | |
| Organizational Communication and Driving Cultural Change | 3 | 4 | 1 | 0 | Vijay asks how to drive cultural change in a massive enterprise. Kim applies Amdahl's Law to corporate communication ratios and describes building a sheltered bubble to establish capability. | |
| AI Applications in Target Selection and Functional Genomics | 4 | 5 | 2 | 1 | Kim dismisses overly hyped AI promises like virtual human testing, then delivers a detailed breakdown of real-world AI applications at GSK from target selection to computational pathology. Vijay intermittently offers domain terms like GWAS. | |
| Evaluating Speed, Cost, and High-Dimensional Biological Data | 4 | 5 | 1 | 0 | Vijay asks for the quantitative impact of ML compared to a decade ago. Kim explains how high-dimensional, cheap measurement technology is useless without machine learning algorithms to perform feature compression. | |
| Strategic Guidance for AI Startups and Data Advantage | 5 | 4 | 2 | 1 | Kim emphasizes that proprietary data generation is the true moat for AI startups rather than algorithm complexity, rejecting pitches based solely on talent. Vijay pushes on data differentiation and chimes in on simple model baselines. | |
| Model Robustness, Integration, and Enterprise Adoption | 4 | 4 | 2 | 1 | Vijay asks what algorithm features make Kim excited to evaluate a model. Kim criticizes papers claiming marginal decimal-place gains and highlights enterprise integration realities. | |
| Five-Year Outlook: Computational Biomarkers and Immune Programming | 3 | 4 | 1 | 0 | Vijay asks for a five-year prediction on computational biology. Kim foresees companion software alongside every drug and hybrid mechanistic-ML models replacing brute-force screening. |
Statements from this episode (18)
Branson notes Relenza was designed using early computational drug discovery methods
“So, so, I was working with Joseph Aghese and Peter Coleman and Brian Smith, and these are the people behind the first neuro-interdates drive, which actually was designed actually with computation. So, they used a program called Goodford This grid to map the pr…”
Branson warns biological models showing 0.9 accuracy usually contain underlying errors
“In biology, as you know, like, the best model you can build has, like, a figure of merit of accuracy, like.6, if it's .9, anything that looks great, there's something wrong.”
GSK made AI core to its strategy via in-house functional genomics
“They're fully investing in the things I thought were very important at the time, which was, you know, these large genetic databases coming online and functional genomics. So he was way ahead in the, in sort of the thinking about CRISPR and like the gene pertur…”
Branson applies Amdahl's law to companies balancing productive work and communication
“Armdahl's Law isn't just for CPUs, it's for companies, so that ratio of, like, computation to communication, right, is really important.”
Branson says big pharma's scale turns AI into an unstoppable juggernaut
“You're surprised at the amount of time it takes for a message to percolate, but eventually everyone, everyone gets behind it, you know, then you've turned the ocean tank and then it's this unstoppable juggernaut, right? Yeah. Because there's a lot of capital a…”
GSK's active learning target discovery is 20% faster than random screening
“That's about 20% faster we've shown than doing a random screen.”
Branson argues AI's medical utility depends entirely on cheap measurement technology
“And so it's interesting because AI has become so useful because we've got so much more measurement technology, right? Cheap measurement technology. So we can do, you know, we can do generalized sequencing, cheap, we can do RNA seq, we can do the single cell ty…”
Branson notes high-dimensional biological data is useless without machine learning
“We've got more measurement technology, but it's in such high dimensionality on people that, like, you, I can't make sense of an array of, like, a whole bunch of expression changes which just fluctuate around baseline versus a disease patient. Right? Like, who …”
GSK rejects AI biotech startups pitching only talent and algorithms
“And if your pitch is just like, oh, we've got smart people, give us your data, we'll do cool stuff. I'm like, well, I've got smart people too, right? That's not gonna happen.”
Branson argues random forests should be the baseline for all ML models
“They should be the baseline everyone should start with for everything, right? And you should always compare your information gain above a random forest that could be depressing for a long time.”
Branson notes simple ML models yield most gains over complex ones
“Basically the simple stuff gets you there, and for the extra five percent gained, you need all the extra complexity. I think that is still true now.”
Branson warns small model improvements do not justify enterprise switching costs
“For a lot of people's stuff, the 10% difference, they're like, I don't care. It's not worth the cost of me buying it and stuff.”
Branson argues precision curves matter more than AUC for enterprise AI
“The precision curve, right is really where it's at over AUC, but you gotta work out what you need to be at to make it meaningfully different”
Branson advises enterprise AI startups to win over engineers, not just users
“Because often there's the people who use it in a company, and then there's the people who have to operate it and run it, and how do we control for model drift and everything else, and so you're selling to a user But the people buying it and installing it that …”
Branson urges tech executives to test internal tools to uncover bottlenecks
“I think you still have to kind of build things, and I think also to, it keeps you thinking about aspects, and I also think it also helps you know when you're, when, that your infrastructure is a company, right? If you're kicking the tires on it regularly, or i…”
Branson predicts every drug launched by GSK will have accompanying software
“So every, every software, every drug VSK will launch will have software that sits around it. I think that's gonna be true of everybody, and there's, we're going to see, there's software that will sit around every drug, and we'll say, you know, who should use i…”
Branson predicts pharma will conduct fewer but more informative experiments by 2029
“So I think in five years time, what will happen is I think we'll be doing fewer experiments, but more informative experiments.”
Branson calls human outcome data big pharma's biggest competitive AI advantage
“Outcome data is, is, is the gold standard data to have, right? It's the rarest data that's really, like, one of the advantages of pharmaceutical companies have, because they can generate data in humans at scale in a safe, ethical manner.”