Sergei Narykov | UHBSE

Research engine for human biology over time.
We model trajectories before anyone treats an intervention as known.
Research prototype, not clinical advice.

[Sergei Narykov | UHBSE @uhbse_engine]

UHBSE is a research layer for comparing possible human biological trajectories before an intervention’s effect is treated as known.The intended product is an auditable human biology simulator. Today, it remains a research prototype.

[Sergei Narykov | UHBSE @uhbse_engine]

Can we move from predicting biology to running useful virtual experiments on it?

[Sergei Narykov | UHBSE @uhbse_engine]

AI may dramatically compress drug discovery. But discovering an intervention and understanding what it will do to an interconnected human system are different problems.
We’re building UHBSE around the second one: simulate and compare biological trajectories under alternative interventions.

[Sergei Narykov | UHBSE @uhbse_engine]

*This is an important step toward longitudinal modeling of human health. Predicting how disease risk evolves over time is one problem; the next is counterfactual: how might the same biological trajectory change under different interventions?
That’s the layer we’re building with UHBSE comparing alternative intervention trajectories over time.

[Sergei Narykov | UHBSE @uhbse_engine]

Really interesting to see context length become an engineering lever in biological modeling. At a different scale, whole-body simulation faces a related challenge: retaining enough interacting biological context while keeping inference tractable. That tradeoff becomes increasingly important as models move toward more complex biological systems.
That’s one of the problems we’re working through with UHBSE.

[Sergei Narykov | UHBSE @uhbse_engine]

This is a fascinating direction. AI agents can dramatically expand the search space for drug discovery. The next layer is equally interesting: once an intervention is proposed, can we model how the human biological state may evolve under it over time?
That’s the layer we’re building with UHBSE

[Sergei Narykov | UHBSE @uhbse_engine]

This is the part of biological AI that matters most: closing the loop with reality. Different biological scale, same principle a model only becomes useful when its predictions survive independent validation.
With UHBSE we’re applying that principle to longitudinal human biology: simulate intervention trajectories, then test them against previously unseen clinical data

[Sergei Narykov | UHBSE @uhbse_engine]

Interesting example of where biology seems to be heading: not bigger models alone, but models constrained by multiple independent layers of biological evidence.
That’s also the architecture we’re pursuing with UHBSE at the whole-body level mechanisms + longitudinal clinical data + intervention trajectories.

[Sergei Narykov | UHBSE @uhbse_engine]

Important framing. The jump from biological prediction to biological simulation needs more than scale: explicit mechanisms, temporal state, and data that capture how systems change under intervention.
We’re taking the same hybrid view at the whole-body level with UHBSE mechanistic structure constrained by longitudinal clinical data, with validation on unseen trajectories as the key test.

[Sergei Narykov | UHBSE @uhbse_engine]

@AnthropicAI Making specialized biology models cheaper to run is a big unlock. Another bottleneck is what we ask them to model: not only molecules, but how a human biological state evolves over time under an intervention. That’s the layer we’re building with UHBSE.

[Sergei Narykov | UHBSE @uhbse_engine]

@biogerontology This is a strong case for domain-specific biology over generic scale. The next frontier may be temporal: moving from interpreting biological states to modeling how those states evolve under different interventions. That’s what we’re building with UHBSE.

[Sergei Narykov | UHBSE @uhbse_engine]

@BiologyAIDaily One lesson here goes beyond antibodies: architecture matters, but the biological data distribution matters just as much. That’s central to UHBSE combining longitudinal clinical trajectories with mechanistic biology to model intervention-dependent human states.

[Sergei Narykov | UHBSE @uhbse_engine]

@biogerontology AI can help design the intervention. The next challenge is predicting the biological trajectory it may trigger.
That’s the layer we’re working on with UHBSE: comparing possible whole-body responses to different interventions from the same baseline.

[Sergei Narykov | UHBSE @uhbse_engine]

Hope you are not used an LLMs😃I’m not sure a single biomarker is the right target. What seems more useful is a standardized longitudinal endpoint: several routinely measured markers interpreted as a trajectory, not isolated snapshots.
The bigger challenge is separating natural variation from the effect of the intervention.

[Sergei Narykov | UHBSE @uhbse_engine]

@BioAI_NeuralNet Interactive papers could be much more useful than static PDFs if they preserve provenance, assumptions and reproducibility. In computational biology, that audit trail may matter as much as the interface itself

[Sergei Narykov | UHBSE @uhbse_engine]

@QuanquanGu AI-native drug discovery expands the space of candidates. The next hard problem may be modeling what those interventions do to longitudinal human biology not just a target, but the biological state that follows over time.

[Sergei Narykov | UHBSE @uhbse_engine]

@fvderop Biological knowledge is getting cheaper fast. Verified biological prediction is a different problem especially when the target is a longitudinal human trajectory rather than a static question.

[Sergei Narykov | UHBSE @uhbse_engine]

@kastacholamine @srikosuri A single-atom change, multiple downstream consequences. This is exactly why modeling interventions as changes to one endpoint feels insufficient the interesting problem is the coupled trajectory of the system that follows.

[Kate Stafford @kastacholamine]

I feel like this paper didn't get enough airtime when the preprint came out. Happy to see it published now! Great experiment (may annoy your local medchemist) https://t.co/MuKWTlJyIQ

[Sergei Narykov | UHBSE @uhbse_engine]

@BiologyAIDaily Sequence → function is becoming remarkably modelable. The next difficult layer is function → longitudinal biological state, where context, interacting systems and time start to matter.

[Sergei Narykov | UHBSE @uhbse_engine]

@adaptyvbio @TwistBioscience @Anthropic Designing the molecule is one challenge. Modeling what happens to the biological system after the intervention may be the next one especially when downstream effects unfold across multiple systems and over time.