ParamAI is developing a model for drug developers that estimates hard-to-measure drug parameters from fragmented evidence to predict exposure across patient groups and support defensible dosing decisions for regulatory review.
How much gets absorbed? How quickly is it cleared? Unknown drug behavior can change which dose looks right.
Estimate the missing parameters, simulate drug levels, and identify what you need to learn before choosing a dose.
FDA's Project Optimus seeks doses that best balance benefit, safety, and tolerability. A higher dose can add side effects without adding benefit.
The opportunity is to compare dosing options earlier, using models alongside experimental and clinical evidence. Explaining the choice matters. Making a better choice comes first.
The law recognizes non-animal approaches, including computer models, within nonclinical testing. It does not make every model an accepted substitute for a study.
Enacted law, section 3209 ↗︎FDA's oncology guidance recommends evaluating dosage during development, drawing on exposure, activity, safety, and tolerability evidence.
Oncology dosage guidance ↗︎M15 sets out common principles for planning, evaluating, and documenting model-informed evidence for regulatory review.
M15 final guidance ↗︎Estimate how a drug behaves when key measurements are missing. Predict what that means for drug levels in the body. Help teams compare doses and see which uncertainties could change the choice.
Exposure is one input to dose optimization, alongside evidence of benefit and harm. These milestones do not establish regulatory acceptance of ParamAI's outputs.
Thousands of parameter sets can explain a drug differently. The loop predicts what each would imply, estimates which observation would reduce uncertainty most, then updates against the result. The next choice depends on what remains unknown.
Every point is a candidate model of how a drug moves through the body. Each one predicts a different exposure.
The working engine holds 4,000 candidate parameter sets. This simplified illustration shows 180 candidates.
01 / 05The working engine samples hypothetical results for each available action, trials the corresponding belief updates, and estimates the average reduction in uncertainty. It then selects the available action with the highest estimated information gain.
The three examples are adapted from its action menu: compare total exposure (AUC), compare terminal half-life, and run an internal repeated-dose simulation. This visual uses a simplified two-parameter model, seven previous doses, and a score based on log-prediction variance relative to assumed observation noise. A simulated result updates weights through a log-normal likelihood. Running candidate simulations alone supplies no independent evidence.
The point cloud shows weighted resampling. Small, fixed screen offsets separate overlapping points without changing the parameters or exposure predictions.
A controlled example, not patient data. We generate a measurement from a known model, then ask the inference loop to recover its behavior without seeing the answer. The dashed line reveals that known answer at the end. In a real application, predictions would need to be checked against independently measured drug levels. The working gefitinib engine also uses simulated observations.
We are exploring how ParamAI could support drug development scientists who need to connect incomplete measurements with predictions of drug exposure.
Oncology is one compelling application. The broader question is where better estimates of hard-to-measure drug behavior could most improve a dosing decision.
Established tools already simulate how drugs move through the body. ParamAI focuses on the uncertain parameters those simulations depend on: estimating them, testing their implications, and identifying which observation would help most.
The goal is to give teams a prediction they can examine, with its assumptions, supporting evidence, and remaining uncertainty connected.
A working inference engine provides a starting point. We are developing the science, shaping potential applications, and seeking support to advance the work.
Strengthen the model, learn from drug development teams, and evaluate where it can be useful. Broader predictive capabilities and validation remain part of the work ahead.
Max has first-authored numerous computational models and brings a background in computer science. He is an MD candidate at WashU School of Medicine and a recipient of Fulbright and Goldwater scholarships.
Building ParamAI at the intersection of computational modeling and drug development. Open to conversations with scientists, collaborators, and early supporters.