Skip to content
Approach

Mechanism-first drug design

Many drugs reach clinical trials with the wrong mechanism. Before any animal study, InterAx Deep Signal™ determines what a drug must trigger inside the cell for efficacy and tolerability, its target signaling profile. We then design the chemistry to trigger it.

The challenge

Traditional drug discovery misses what really drives disease.

It focuses on structure and chemistry.

It refines the molecule: how it binds its target and its drug-like properties.

It underrates the biology.

Conventional screening reduces the cell's response complexity to potency and efficacy (EC50 and Emax) at a single time point.

It leads to unexpected biological effects.

Candidates turn out less effective or less safe than intended, and many fail in clinical trials.

Our target class

GPCRs are the most tunable drug targets, and the easiest to get wrong.

GPCRs are the sensory organs of our cells and the largest family of proteins in humans. Roughly 36% of approved drugs act on them, and about 280 GPCR targets remain undrugged (Hauser et al., 2017). Yet many GPCR drugs carry sub-optimal efficacy or tolerability.

A kinase behaves like an on/off switch: a drug turns it off. A GPCR behaves like a dimmer. A drug can push its activity up or down from baseline, by any amount, and it does so on several pathways at once, each with its own kinetics. Each of those settings is a way to get the mechanism wrong, and the error shows up only in clinical trials (Bender, 2026).

Kinase: a switch

One response: the drug turns it off.

GPCR: a dimmer, on several pathways

G protein (cAMP)
β-arrestin
Other pathways
full agonistpartial agonistneutralinverse agonist

Up or down from baseline, by any amount, and differently on each pathway.

InterAx approach

We extend rational drug design to the cellular mechanism, where disease outcome is decided.

Conventional design builds models of the ligand and the target. We add the third model, the cell's response, and optimize all three together in every design cycle.

1930s to 1980s
Ligand-based
Model of the ligand
1990s to 2020s
Ligand-based + Structure-based
+ Model of the target
2020s onwards
Ligand-based + Structure-based + Mechanism-based
+ Model of the system

Each era of drug discovery has turned more guesswork into rational design. Ligand-based models forecast a compound's properties, such as bioavailability. Structure-based models compute its binding and affinity. Mechanism-based models now predict efficacy and in vivo behavior.

How we design a drug candidate

Mechanistic signaling parameters link chemistry to clinical outcome.

A molecule and a patient are nine orders of magnitude apart, nanometers to meters, too far to map chemistry straight to outcome. Deep Signal bridges them with a compact set of mechanistic parameters, the InterAx signaling fingerprint, measured and modeled in the cell. It serves the two steps below: it sets the cellular profile to aim for, and it extends structure-activity relationships from receptor to downstream signaling. InterAx does both: the signaling assays and models, and the chemistry.

Patient
efficacy and tolerability
1 Identify the optimal cellular profile
Cell
InterAx Deep Signal™
Mechanistic signaling fingerprint
2 Build the chemistry that triggers it
Molecule
InterAx chemistry and structure

1 Identify the optimal cellular profile

What the drug must trigger inside the cell to achieve therapeutic efficacy and tolerability.

  • Time-resolved, multi-pathway signaling assays: cAMP, calcium, β-arrestin and receptor internalization, in human-lineage cell lines at varying receptor densities.
  • A mathematical model of the receptor's signaling, fitted to those time courses, generates the InterAx signaling fingerprint. It extends the classical parameters of pharmacology, potency and efficacy (EC50 and Emax), with mechanistic receptor activation parameters and the kinetics of each pathway.
  • The signaling fingerprint separates the measured response into the biological processes behind it, each described by parameters that do not depend on the assay conditions or the cellular background. These parameters predict the response in primary cells and in vivo better than EC50 and Emax alone. Linked to in vivo or clinical outcomes, they reveal which mechanism drives therapeutic efficacy and tolerability: what an optimal drug must do.
  • Beyond a certain complexity, intuition cannot keep up; the mathematical model can. In GPCR signaling, what sounds right is often wrong.
Endpoint readout: all drugs look the same
Over time: different dynamics

2 Build the chemistry that triggers it

Structure-activity relationships connect drug chemistry with the optimal cellular profile, built with molecular dynamics, quantum chemistry and machine learning, so every compound is designed toward that profile.

  • GPCR molecular dynamics, covering nearly every GPCR target, and quantum chemistry describe how each ligand binds and how the receptor responds to it, so a design idea is tested on the structure before anything is synthesized.
  • Ultra-large virtual screens, physics-based affinity prediction and generative AI design propose and rank new compounds against the optimal cellular profile.
  • Machine-learning property prediction and medicinal chemistry turn those candidates into oral, drug-like molecules, with potency, selectivity, permeability and metabolic stability tuned together, and synthesis at contract research organizations.
  • Every design-make-test-analyze cycle runs through the signaling assays, so each compound's effect on the cellular profile feeds the next design and biology and chemistry are optimized together.
AI

We generate cell signaling data that AI can learn from.

Most AI effort in drug discovery goes into the molecule, and most drugs fail on the biology. The effort goes to chemistry because that is where the training data is, and that data means the same thing everywhere, whichever laboratory measured it. Cell signaling does not: the same drug gives a different potency, efficacy or bias depending on the assay, the cell line and the time point, so a model trained on such readouts learns the assay conditions as much as the drug. InterAx produces signaling data that holds across assays.

AI needs data with low conditionality: a value that means the same thing wherever it was measured.

Images
a cat is a cat, however it's photographed
low
Chemistry
structure, affinity
medium
Signaling readouts
potency, efficacy, bias
high
Deep Signal™ mechanistic parameters
the biology itself
low

Conditionality after Bender et al., Drug Discovery Today, 2021.

Deep Signal™ removes that conditionality. The mechanistic model separates the readout into parameters of the biology itself. They stay the same whatever the assay, so they can be learned from and compared across receptors, chemical series and tissues. It also needs far less data: the model already holds the biological mechanisms, so AI does not have to learn them from scratch. Each experiment leaves only a few parameters to fit, and a handful of time courses is enough.

AI then works on both sides of the loop: on the chemistry, where generative design, property prediction and ultra-large screens propose compounds for the target profile, and on the biology, where our AI pharmacologist, built on the same mechanistic models, analyzes and interprets every experiment.

What this changes

Four differences from conventional design.

  • We read the response over time, not an endpoint. Conventional assays summarize a drug's effect as potency and efficacy, EC50 and Emax; we measure the cell's response as it unfolds, over seconds to minutes, on several pathways at once. That readout is the design objective for the chemistry.
  • The parameters hold outside the assay. They quantify the biological processes themselves, from drug binding and receptor conformation to signaling partners and downstream pathway activation, so they carry over to other cells and in vivo.
  • Bias is measured apart from kinetics. A drug's bias between pathways, as usually estimated, changes with the time point of the assay, because the estimate mixes how strongly the drug drives each pathway with how long it stays on the receptor (Klein Herenbrink et al., 2016). The InterAx signaling fingerprint separates the two, so the bias we design to is the drug's own, and its effect on efficacy and tolerability can be read.
  • Clinical outcomes become predictable early. For six GLP-1 compounds with clinical data, parameters measured in cells explain over 80% of the variation in their gastrointestinal side effects (R² ≥ 0.8).
Ten years of evidence behind the engine

What the engine has already measured and modeled.

Mathematical models, experimental assays and a tight biology-chemistry loop took years to get right and are not quickly rebuilt.

15+
GPCRs with kinetic signaling models, across classes A, B and C
3,000+
time-resolved assays performed
1,000+
drug candidates characterized
1,000+
GPCR-ligand molecular dynamics simulations, covering nearly every GPCR target

Built on internal programs and research collaborations, the engine has characterized small molecules, peptides and antibodies, agonists and antagonists, orthosteric and allosteric mechanisms, from hit identification to lead optimization, in metabolic, CNS, oncology, inflammation and cardiovascular indications.

More drugs enter clinical trials with the right mechanism.