Fulcrum Platform

AI is an amazing tool for answering well-framed questions. Fulcrum’s platform is uniquely capable of identifying the right questions to ask and providing the appropriate frame in which to contextualize them.  

Fulcrum models disease as a departure from system-level homeostasis. The platform incorporates human variability and structured uncertainty, generating a virtual population of tens of millions of discrete trajectories.  Therapeutic candidates, corresponding biomarkers, appropriate patients, and time-of-intervention are all identified within the same framework based on expected improvement over forecast.

The platform’s core is a world model representing neurobiological feedback loops as nonlinear differential equations. Critical homeostatic processes such as mass-energy balance are explicitly incorporated, traversing scales from the system to the cell, protein, and gene, as necessary to faithfully reproduce clinically meaningful biological and phenotypic trajectories. AI enters the picture by dramatically accelerating our ability to conduct perturbation and restabilization studies in silico. The latest AI tools also greatly speed the incorporation of additional data, extending and deepening model scope.

Fulcrum is Mechanism-First

The conventional drug discovery paradigm today is find-then-rationalize. A statistical method ranks candidates; the top-ranked candidate becomes a target; a hypothesis is then built to explain how the target drives the disease, with supporting data recruited along the way. Programs are scored relativistically, with no shared model beneath them, and each is advanced as an independent shot on goal.

A line drawn from a target to an endpoint is a thin thing. Very little of what we know about a disease can press on that line. Said another way: we use almost none of the available data about a disease to test our hypotheses before progressing into drug development

Perhaps this is why the failure rate for disease-altering neurodegeneration drugs is nearly one hundred percent.

Fulcrum reverses the conventional approach. For us, the hypothesis comes first, and we make it wide. It is a model of the regulated system: the buffering capacity that holds physiology steady, the feedback that maintains that capacity, and the way the whole arrangement loses stability and decompensates into disease. All available data is incorporated into this unified model. We then perturb the system at thousands of candidate mechanisms, watch which perturbations restore homeostasis or blunt the forces driving its collapse, and only then ask which protein can best carry out that perturbation. Function first, entity second. In this manner, targets emerge from the model.

Because every candidate lives inside a single platform, the entire portfolio is weighed in a single framework, and each candidate is ranked along the same critical axes such as the phase of the disease, the patient subpopulation, and the degree of perturbation.

This changes the role of the data. The conventional approach uses data to support an already-selected target. Fulcrum uses data to reconstruct causality across scales and across the differences between patients. The organizing principle is homeostasis, not the weight of evidence stacked along a one-directional graph, and so a prediction is shaped by the whole reconstructed system rather than by the facts that happen to land on the line. Return to the thin line problem. A whole-disease hypothesis presents an enormous surface, and one body of evidence presses on it from every side at once. The more data a hypothesis can explain, the more likely it is to be correct.

Fulcrum is Unique and Positioned for Rapid Progress

Nobody else builds the whole-disease model the way Fulcrum does. This is extremely demanding work: reconciling nonlinear feedback loops across brain regions, time scales, and levels of abstraction from gene to system, and reconciling the findings of over 1,800 independent papers.  

Fortunately, the advent of advanced AI tools in the past year has transformed our ability to interrogate and extend the model. AI agentic assistance is invaluable in interrogating and extending the model. Our speed of execution is increasing dramatically. What took a year or more now can be accomplished in weeks. We expect this pace to accelerate further in fall 2026 as we release version 4 of our platform and hire additional personnel to run in silico studies.