learning about agents x ml x bio @generate_biomed | prev @ginkgo @brownuniversity
Boston, MA
Joined July 2020
- Tweets419
- Following493
- Followers2.4K
- Likes2.6K
Liam Bai retweeted
The "Age of Miracles" era:
I took a look inside Anthropic's latest progress in the life sciences, including working with a ton of biopharma partners and a new acquisition disclosure
All coming before what is likely the business world's largest IPO:
endpoints.news/anthropic-pit…
Liam Bai retweeted
This was a surprise and updated my thinking. Would have expected this level of performance improvement to have come from a fine-tuned protein model and lots of training data specific to the task, not off-the-shelf frontier LLM.
Great work @generate_biomed and @AnthropicAI !
Liam Bai retweeted
Love reading about these projects that would have taken entire PhDs to do previously. We're going to see a lot more elegant, self-contained solutions like this in the coming months...
What will drug discovery look like when intelligence is abundant?
Our team at @generate_biomed began with this question, and we’re excited to share some results!
We took on in vivo CAR-T and designed a scaffold 70% smaller than ones in the clinic, with 3x the potency 🧵
Liam Bai retweeted
We have barely scratched the surface of ICD design.
Shout out to @liambai21 for leading this work.
What will drug discovery look like when intelligence is abundant?
Our team at @generate_biomed began with this question, and we’re excited to share some results!
We took on in vivo CAR-T and designed a scaffold 70% smaller than ones in the clinic, with 3x the potency 🧵
Liam Bai retweeted
Incredibly cool to see. This + the one-man biotech acquisition by Anthropic are lead indicators of the future
AI will spawn a thousand flowers in biology. “Why not me?”
What will drug discovery look like when intelligence is abundant?
Our team at @generate_biomed began with this question, and we’re excited to share some results!
We took on in vivo CAR-T and designed a scaffold 70% smaller than ones in the clinic, with 3x the potency 🧵
Liam Bai retweeted
A smaller CAR T-cell receptor could make room for a more capable therapy.
In our collaboration with @AnthropicAI, six of our scientists used @claudeai alongside our own models, computational tools, and wet lab data to refactor a CAR T-cell receptor for solid tumors.
The redesigned signaling component was about 70% shorter than a standard design, making room for new functions and improving cancer-cell killing in laboratory tests. Watch how our scientists turned years of work into months, at a fraction of the cost, expanding what we can engineer for patients.
@AnthropicAI's case study → anthropic.com/research/early…
Our preprint → generatebiomedicines.com/pub…
Our blog → generatebiomedicines.com/blo…
Liam Bai retweeted
I've been hanging out with this rock star team at @generate_biomed for the past few weeks and while this result is worth a deep dive, the way they work is the real game changer. They've allowed AI to remove the friction in the organization. As AI-pilled as I am, it's taken some adapting to transform to the way the future will look, but there is no catering to the 'old' way.
And that's the only way the true impact of AI will be seen, not just in the frontier labs, but also in the AI native biotechs working to bring real cures to patients fater.
Things are going to change quickly and the hardest part is going to be how people and systems adapt to move at speed. And this team is doing it.
Congrats @jasonyi96, @liambai21, @wwang002, @KaggenA!
What will drug discovery look like when intelligence is abundant?
Our team at @generate_biomed began with this question, and we’re excited to share some results!
We took on in vivo CAR-T and designed a scaffold 70% smaller than ones in the clinic, with 3x the potency 🧵
Liam Bai retweeted
Less really is more. Next for us: target selection for early stage solid tumors with @AnthropicAI
What will drug discovery look like when intelligence is abundant?
Our team at @generate_biomed began with this question, and we’re excited to share some results!
We took on in vivo CAR-T and designed a scaffold 70% smaller than ones in the clinic, with 3x the potency 🧵
Liam Bai retweeted
Lab work in science remains artisanal rather than industrialized.
Much of the training is still an apprenticeship in acquiring "good hands.” Molecular techniques advance slowly because practical knowledge is passed from one lab member to the next through observation and imitation.
What surprises me is how little effort is being made to change this. Existing technology could revolutionize laboratory training through scale AI-style efforts to collect short, tiktok-like videos of scientists performing essential techniques. AI could then organize, compare, and extract the practical knowledge embedded in them.
The tools already exist. I haven’t seen anyone try to develop this.
As someone coming from a computational background to biology, a big culture shock is realizing just how much *improvisation* happens in the lab.
Small observations like the viscosity of the liquid, firmness of the pellet, resistance while pipetting can all drive on-the-fly adjustments: spin a little longer, mix a few more times, or aspirate with a bit more force…
These details are critical to troubleshooting and data interpretation - and can make the difference between signal and noise. But they’re rarely recorded.
Without knowing the messy physical reality of experiment execution, the AI scientist, no matter how brilliant, flies blind. This observability gap decouples AI progress from scientific progress.
Excited to see this great team take on one of the most important problems in science!
Liam Bai retweeted
OpenBind intends to collect 10,000s of protein-ligand structures & affinities. To prioritize what we collect next, we need cofolding models trained on the latest data. Today we're releasing OpenBind-0 and 717 new ligand-bound structures.
Liam Bai retweeted
This. Anyone who has built in biotech knows how important great RA's are and how disruptive it is when they leave.
I remember talking to a guy at DropBox about "rock engineers", people who are just like, happy doing their job, know how everything works, are hyper productive, and have no desire to climb a ladder because they love what they do. Solid and reliable. So much of biotech runs on people like this. More should be done to incentivize them.
Liam Bai retweeted
A sparse autoencoder aims to learn a dictionary of interpretable features from a model's activations — but a lot can come out "dead," never firing once. On some models this is rare; on others, >70% die even with fixes like AuxK. We went down a rabbit hole to understand why...
Liam Bai retweeted
Progress in generative biology becomes real when a molecule performs exactly as it was designed.
Today, we announced the initiation of SOLAIRIA, the first global Phase 3 studies of a long-acting anti-TSLP antibody for severe asthma. These studies will evaluate GB-0895, which blocks TSLP — an upstream driver of the inflammatory response in asthma. Using our generative optimization technology stack, GB-0895 was computationally designed to improve potency and half-life while retaining specificity, supporting a potential every-six-month dosing schedule.
This move into Phase 3 demonstrates what programmable biology can enable — design translating into development.
Learn more about SOLAIRIA → solairia.com?utm_source=x-li…
Read our full announcement → generatebiomedicines.com/med…
Liam Bai retweeted
Published! 🎉
Paper now has more feature analysis and higher quality figures - thanks to great reviewer feedback!
Code also got a major upgrade - v1.0.0 is way more modular so you can easily swap in different protein embeddings or SAE architectures: github.com/ElanaPearl/InterP…
How do protein language models (PLM) think about proteins?🧬 We answer this w/ #InterPLM, just published in @naturemethods!
Using sparse autoencoders + LLM agent, we identify 1000s of interpretable concepts learned by PLMs, pointing to new biology 🧵
Liam Bai retweeted
Entering my PhD, de novo antibody design was a grand challenge I thought would not be solved without huge increases in affinity data and Ab-Ag structure. Only 2 years later, we provide the first open-source recipe to get antibody binders, almost magically, out of a computer (1/3)
Liam Bai retweeted
Can protein LMs reveal scientific knowledge? We start by asking how pLMs turn sequences into structure signals. We map a contact prediction circuit: early motif features gate later domain features. Spurious or science? We can now test. 🧵(1 of N)
Liam Bai retweeted
Anthropic's recent blog post provides a good summary of some of the recent work applying mechanistic interpretability to biology. I am very bullish on this research direction!
Liam Bai retweeted
MechInterp for pLMs has been considered a “dark art” for some time due to spurious correlations and pseudo-insights into biological mechanisms
Happy to have contributed a bit to this ongoing effort, containing many stellar, honest results. Plenty of work yet to be done!
Update to our ICML paper on interpreting features in protein language models: We ask human raters to assess the interpretability of SAE feature activations and ESM activations. Human raters found SAE features far more interpretable!
nitter.cf/etowah0/status/1888977…
Liam Bai retweeted
Update to our ICML paper on interpreting features in protein language models: We ask human raters to assess the interpretability of SAE feature activations and ESM activations. Human raters found SAE features far more interpretable!
nitter.cf/etowah0/status/1888977…
Can we learn protein biology from a language model?
In new work led by @liambai21 and me, we explore how sparse autoencoders can help us understand biology—going from mechanistic interpretability to mechanistic biology.
Liam Bai retweeted
You can take a ESM's final-layer projection matrix (used to get logits) and apply it to intermediate layer embeddings. This can roughly tell you what the model “thinks” the sequence is at different layers.
nitter.cf/liambai21/status/19256…