Building UHBSE: a clinical simulator to compare possible biological futures before intervention. Grounded in longitudinal patient data spanning 7 years.
Wilmington, DE
Joined August 2026
- Tweets394
- Following120
- Followers38
- Likes167
Pinned Tweet
Simulate before you intervene.
We’re building UHBSE, a clinical simulator designed to compare possible biological futures for an individual before a real-world intervention.
We’re starting with a clinic’s long-term monitoring programme. The goal: give its care team an evolving view of the patient, from what changed between visits to possible future trajectories and, as models are developed and validated, alternative intervention scenarios.
Today, our research foundation reconstructs patient states from longitudinal clinical records and evaluates forecasts of future measurements, preserving the source of each observation. Development is grounded in real-world patient data spanning seven years across two clinical sources.
Our next product milestone: a focused clinic prototype, tested retrospectively for a predefined measurement and forecast horizon, alongside an evaluation of care-team review time.
Research stage. Clinical validation and intervention-response modelling remain ahead.
Explore the synthetic concept demo: uhbse.tech/#engine
Investors and clinic design partners: let’s discuss the evidence, product roadmap and first pilot.
uhbse.tech/#partners
Genomics helps frame what to watch. Longitudinal records show what changes.
A clinical simulator must then be tested on patients it has not seen, for a defined outcome and time horizon.
Combining data is a starting point. Demonstrating usefulness is the next test.
We are building UHBSE as a clinical simulator for exploring possible biological outcomes before intervention.
A useful test must name the patient group, outcome and time horizon and show when the evidence is too weak to support a prediction.
Simulation has been used in engineering for decades. Yet validating even a narrow model can take years.
Now we're attempting something far more complex: modeling human biology.
The challenge isn't generating plausible outcomes. It's proving when those outcomes deserve scientific trust.
We can model molecules. We can edit genes.
But can we reliably model how one person’s biological state might change under different interventions?
That’s a different scientific challenge.
At UHBSE, we’re building toward it using mechanistic simulation and longitudinal clinical data.
Validation comes before claims.
Tele-EEG and AI-Enabled Digital Twins for Epilepsy: A Scoping Review mdpi.com/4137946 #mdpicomputation через @ComputationMDPI
A digital twin is not simply a model personalized once.
The real challenge is continuous updating, longitudinal prediction, uncertainty and independent validation.
Until we can demonstrate those properties, we should be careful about calling a biological simulation a digital twin.
Drug discovery is accelerating. But designing a molecule and predicting how a human biological state will change are fundamentally different problems.
We need longitudinal evidence, reproducible models and explicit uncertainty.
At UHBSE, we’re building toward that missing layer.
52,236 cancer patients. Longitudinal clinical data. Independent validation.
A new Nature Medicine study reveals systemic physiological changes over time.
The next frontier: modeling alternative responses to interventions.
That’s the research question behind UHBSE.
AI is making biological hypotheses cheap. Validation remains expensive.
But there’s another missing layer: the transition from a molecular result to a human trajectory. A model that predicts binding has not yet predicted response in a person.
At UHBSE, we’re building research infrastructure for source-aware longitudinal states and reproducible predictions. Intervention-response modeling is a goal, not a validated capability today.
Scientific AI needs more than a strong benchmark.
Does the result hold in a new biological context? Can the model show when it is outside its limits?
Those questions matter when building patient-specific biological simulators.
AI is expanding what we can discover in biology.
But discovering a target is not the same as predicting what happens to a human physiological system after an intervention.
At UHBSE, we’re exploring longitudinal simulation of those changes.
The decisive milestone: independent validation.
The next frontier in computational biology isn’t just identifying what drives disease.
It’s understanding how biological states evolve under different interventions.
This is the problem we’re exploring at UHBSE through longitudinal physiological simulation. Validation remains the critical test.
Today, @insitro announces our new collaboration with @AstraZeneca worth up to $1.5 billion.
The partnership pairs insitro's AI and causal human biology platform with AstraZeneca's drug development expertise to pursue new medicines for cardiometabolic diseases.
This is what applying AI to the life sciences is all about: the opportunity to understand what drives a disease and turn those insights into medicines that make a difference for patients.
Congrats to @DaphneKoller, @im_dlight, @mmrozenman, Sharon Barr, Aradhana Sarin, and the entire insitro and AstraZeneca teams.
Sergei Narykov | UHBSE retweeted
Today, @insitro announces our new collaboration with @AstraZeneca worth up to $1.5 billion.
The partnership pairs insitro's AI and causal human biology platform with AstraZeneca's drug development expertise to pursue new medicines for cardiometabolic diseases.
This is what applying AI to the life sciences is all about: the opportunity to understand what drives a disease and turn those insights into medicines that make a difference for patients.
Congrats to @DaphneKoller, @im_dlight, @mmrozenman, Sharon Barr, Aradhana Sarin, and the entire insitro and AstraZeneca teams.
AI drug discovery is accelerating. But there’s a question we still struggle to answer:
What happens to the entire human biological system after an intervention not just to a single target?
At UHBSE, we’re building toward that question.
The next milestone: independent validation.
Building a clinical simulator means making each step testable.
At UHBSE, we have built a longitudinal research foundation across two clinical sources, including reconstructed individual states and technical checks for data accounting, repeatable execution and source consistency.
Our next product milestone is a focused prototype for one clinic workflow. The planned evaluation will compare a forecast for a predefined measurement and time horizon against a simple baseline, alongside an assessment of care-team review time.
We want to establish how reliably the model performs and whether the workflow delivers practical value to the people using it.
Research stage. Clinical validation remains ahead.
Explore our current evidence and product vision:
uhbse.tech/
One starting state. Several possible paths.
UHBSE’s public concept demo illustrates the experience we’re working toward: start from a biological state, move through time, and inspect how different scenario branches evolve.
The interface brings together a starting state, biomarkers, a timeline and illustrative branches, so visitors can explore the idea of comparing possible futures from the same starting point.
The demo uses synthetic data and illustrative response rules. It is separate from our research models and does not demonstrate a validated treatment effect.
Our research foundation already reconstructs individual states from longitudinal clinical records and evaluates technical forecasts of future measurements. Reliable intervention-response modelling remains ahead.
Explore the synthetic concept demo:
uhbse.tech/#engine
Predicting structure was a breakthrough. Understanding dynamics is another challenge entirely.
At UHBSE, we’re exploring this question at the physiological level: how biological states evolve over time under different interventions.
The essential next step is independent validation on unseen longitudinal data.
Google DeepMind's Pushmeet Kohli says AlphaFold's useful structure predictions did not solve protein dynamics:
"Proteins are extremely complex. They are disordered. Their shape might change depending on the context."
"What we are trying to do is basically someone got a structure, deposited it in the PDB, and we are trying to replicate and get the same structure. That's what we did, right? And it just so happens that it's useful. But that does not mean that we have understood all of protein dynamics."
"Let's not stop the funding of protein structure prediction and protein dynamics because we are just getting started."
- Pushmeet Kohli, Google DeepMind, at Biohub, via Latent Space (@swyx)
One of the hardest questions in deep tech is how to distinguish genuine technical progress from a compelling narrative.
In computational biology, a working demo is not enough. Independent validation, reproducibility, and meaningful biological endpoints matter far more.
Are we rewarding the right milestones?
MakerMaker.AI co-founders Dhaval Adjodah and Owen He with a16z's Martin Casado on recursive self-improvement, and how to tell when it's real:
A lot of RSI pitches are really just people using AI to build better AI. Tech has always worked that way. You needed computers to build chips. Steam engines got better once you could get coal faster.
Dhaval and Owen's bar is higher. The model has to improve itself: its weights, its harness, even the prompt that tells it what to work on next.
And the gains have to hold up when you count what they cost. They track everyone's experiment spend on a "bleeder board," and one engineer, the Viceroy of Variance, has the job of breaking their results.
Over the past few months, they've started seeing early signs of their system getting better and cheaper at the same time.
0:55 Why most RSI pitches miss
2:25 Optimizer vs. optimizee
5:10 The RSI equation: performance over cost
6:50 Why a closed loop does not mean takeoff
8:40 Humans are agents inside the system too
9:25 The denominator trap: progress divided by capital raised
11:40 Training GPT-2 for $100
13:40 Most RSI papers fail when replicated
16:40 Why you have to co-optimize
17:40 Why inference is the first problem
20:15 The agent rewrites its own prompt
22:20 "It's not a belief. It's happening."
25:00 Why they haven't published
27:50 Intelligence is a vector and Amdahl's law
29:45 The garden hose theory of RSI
YouTube: youtu.be/3Q1mdn5ZGvM
@makermaker_ai @_dval_ @martin_casado