Intro to AI for Scientists
McClain Thiel · Encode Fellow · UCL Barnes Lab
A bit about me
- ● UC Berkeley — Data Science
- ● Tempus AI — clinical genomics
- ● Snorkel AI — the "labels are expensive" problem
- ● WealthHawk — founded, built, acquired
- ● Encode Fellow + PhD @ UCL — AI-driven DNA design
The Hype Wall
The word "AI" appears in ~75% of Series A pitches.
Today
- 1 The basics of ML
- 2 What ML can('t) do
- 3 The business lens
The basics of ML
Clearing the Fog
"A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E."
What can ML actually do?
Labeled examples
Learn a mapping from input to known output.
Structure without labels
Find patterns, clusters, or low-dimensional structure.
A bit of both
Label some examples, infer the rest.
That's it. Everything else is a creative application of this.
Same idea. Different framing.
input
house features (size, location, age)
↓
output
price = $420,000
input
"the quick brown fox"
↓
output
P(next word | context)
input
MKTAYIAKQRQIS...
↓
output
3D atomic coordinates
Image classification
the task that kicked off the deep-learning era
an image
convolutional neural network
class probabilities
AlphaFold
2024 Nobel Prize in Chemistry · Hassabis, Jumper, Baker
MQIFVKTLTGKTITLEVEPSDT IENVKAKIQDKEGIPPDQQRLI FAGKQLEDGRTLSDYNIQKEST LHLVLRLRGG
amino acid sequence
ubiquitin · 76 aa
3D atomic coordinates
PDB 1UBQ
How an LLM actually works
input: all previous tokens → output: probability over next token
What ML can('t) do
Three ingredients
What goes into each box
Gather examples. Pay someone — or something — to say what's what.
Pick a model. Minimize loss on the labeled data.
How well does it work on data it hasn't seen?
When ML works
ML is a pattern-recognition algorithm.
- 1 You can define what "good" means. Labels. An evaluation metric. Ground truth you can measure.
- 2 You have a lot of data. Thousands to millions of labeled examples of the pattern.
- 3 Your goal is to automate a difficult task. Something a human can do — but slowly, expensively, or inconsistently.
When doesn't ML work?
WHAT WE ASSUME
Trial patients ≈ test patients
test accuracy looks great
WHAT ACTUALLY HAPPENS
Trial patients ≠ real-world patients
accuracy drops in deployment
In vs. out of distribution
DERMATOLOGY AI trained on light skin → −15% accuracy on dark skin
SEPSIS MODEL AUC 0.83 at Hospital A → 0.63 at Hospital B
RARE DISEASE 10k common cases in training → near-random on 200 rare cases
Where AI wins. Where it doesn't.
- Consistent, repeatable patterns
- Large labeled datasets
- Clear input / output framing
- Measurable, objective outcomes
- High-volume decisions
- Distribution shift
- Tiny datasets
- Novel biology with no prior data
- Causal reasoning
- High-stakes, irreversible decisions
The value chain
Who actually makes money?
Which layer are they in?
Own the data, fine-tune on top
Proprietary, labeled domain data. The thesis turns on what data you have.
Train a model from scratch
Pre-training at scale. In bio: ESM, AlphaFold, Boltz are open and commoditising.
Build a product on top
Some succeed (FDA-cleared product, real distribution). Some don't (no moat, no buyer budget).
What makes a real data moat
Clinical partnerships, proprietary instruments. Data you only see if you own the pipeline.
Wet-lab experiments, rare patient cohorts. Data that takes real money and real time.
More users → more labels → better model → more users. Usage itself generates training data.
Can't scrape. Can't replicate. Access governed by contracts or regulation.
1. Are you building a model, or using someone else's?
2. What data do you have?
3. What does "good" look like for this task?
The mental model
Questions?
McClain Thiel · UCL Barnes Lab