AI + Physics-based models for the hardest problems.
Trusted by
02. WHY LUCAI
When data are limited and structure alone cannot explain function, we build methods around the molecular problem.
We combine molecular simulation, physics-based models, and AI to understand binding, dynamics, and activity.
03. WHAT WE SOLVE
We identify the molecular mechanisms that guide drug discovery and protein engineering.
Turning a free-energy number into a validated prediction with a documented error envelope.
Turning proteome-wide evidence into a ranked, auditable shortlist of targets worth pursuing.
Finding actionable interaction information on broad, shallow interfaces with no defined pocket.
Modeling components, proportions, and interactions to prioritize formulations worth testing.
Generating and validating de novo enzyme candidates when nature offers no starting point.
MORE INFORMATION
×WHY THE NUMBER CAN'T BE TAKEN AT FACE VALUE
Benchmark accuracy does not transfer to a new protein family, manual system preparation moves the result more than the sampling does, tight convergence can still be confidently wrong, and results run by hand across a programme are never comparable.
FROM A CALCULATION TO A DECISION
We treat preparation as a parameter, run automated alchemical workflows that monitor and relaunch themselves on the open free-energy stack, validate the protocol against measured affinities on your own targets, and deliver every number with a documented error envelope.
WHAT YOU GET BACK
A deployed workflow running in your environment, a validation report with the measured accuracy on your protein family, predictions you can defend to reviewers, and the conditions under which the protocol stops being reliable.
SCIENTIFIC SCOPE
The approach ranks close analogues well but is weaker across distant scaffolds, and depends on a credible binding pose. Large ring modifications and slow conformational change remain hard, and where validation shows the protocol falls short on a target family, that is reported rather than hidden.
MORE INFORMATION
×WHY TARGET LISTS GO WRONG
Essentiality is not the same as druggability, whole-protein homology misses pocket-level selectivity, evidence layers disagree, and new organisms rarely provide enough labelled data for a transferable model.
FROM A PROTEOME TO A DEFENSIBLE SHORTLIST
We assemble functional, metabolic, essentiality, expression, and structural evidence proteome-wide; compute pocket druggability; evaluate host and off-target risk at the pocket; then expose and tune every weighting with your team.
WHAT YOU GET BACK
A ranked target list with evidence behind every position, a druggability verdict for each candidate, an exposed weighting model that can be adjusted and re-run, and a clear view of which candidates should be dropped.
SCIENTIFIC SCOPE
Prioritisation ranks evidence; it does not prove a target. Confidence depends on the structural and experimental coverage available, and the shortlist defines what is worth validating rather than replacing experimental validation.
MORE INFORMATION
×A broad, shallow interface gives conventional cavity-based docking no defined site to operate on.
A few residues can carry most of the binding energy, even when the surface geometry does not reveal them.
Some waters are costly to displace while others bridge the interacting partners.
A plausible complex does not identify where a small molecule could interfere with the interaction.
Broad surface, no defined cavity.
Probes reveal regions more likely to bind.
A prioritised shortlist of actionable sites.
MORE INFORMATION
×WHY SINGLE-MOLECULE MODELS FAIL
Descriptors capture each compound alone, averaging erases chemical identity, composition changes outcomes, and the number of possible mixtures grows far faster than experimental screening capacity.
FROM MEASURED MIXTURES TO A RANKED SHORTLIST
We represent molecular structures and properties directly, encode pairwise relationships and relative proportions, and use attention to learn context-dependent, non-additive behaviour across binary, ternary, and higher-order systems.
WHAT YOU GET BACK
A validated predictive model, a ranked chemical space of formulations ready to test, an interaction map showing which compounds and pairs matter, and actionable chemical insight into what drives performance.
SCIENTIFIC SCOPE
Reliability depends on the quality and diversity of the measured formulation data. Applicability-domain and uncertainty analysis are part of the output, and model-derived importance indicates predictive relevance—not proof of physical mechanism.
MORE INFORMATION
×WHY CONVENTIONAL ENZYME ENGINEERING FALLS SHORT
Optimizing an enzyme that already works doesn't help when no known enzyme performs the reaction, brute-force screening across sequence space is intractable, and most methods still require a large historical dataset to train from.
FROM SEQUENCE SPACE TO A VALIDATED DESIGN
We generate enzyme candidates de novo with deep-learning structure and sequence design tools, rank them on active-site geometry and predicted expressability, and confirm regio- and stereochemistry through molecular dynamics before any wet-lab work.
WHAT YOU GET BACK
A small, prioritized set of designed candidates instead of a brute-force library, molecular-level reasoning for why each one should work, and a validated, expressible enzyme ready for optimization.
SCIENTIFIC SCOPE
De novo design starts with no historical dataset and no guaranteed hit, so candidates are prioritized computationally and confirmed experimentally rather than assumed. Full IP in every sequence, design, and result stays with the client.
PROVEN IN PRACTICE
LUCAI built reproducible drug–protein simulation infrastructure that reduced molecular design cycles from eight months to two, delivering reliable binding-affinity predictions before wet-lab synthesis.
FEWER CYCLES, SAME RIGOR↗Structure-guided AI design delivered three functional enzyme variants that replace a toxic, high-cost inorganic step in an active pharmaceutical ingredient synthesis route.
SELECTION AT SCALE↗LUCAI built a predictive formulation engine that maps molecular and process interactions to rank promising formulations before running a single laboratory assay.
PREDICTION BEFORE THE BENCH↗LUCAI built a model that reads clinical records and ranks the genes most likely responsible for a patient's symptoms, improving diagnostic success from 40% to 80%.
SIGNAL FROM THE RECORD↗GOT A HARD MOLECULAR PROBLEM?
START HERE.
Tell us what you need to understand, predict or design. We start from the question—not a predefined service.
We evaluate the scientific question, available information and whether our capabilities fit the problem.
Scope, methods and collaboration model are shaped around the question.