Speaker
Description
Dose selection in early-stage clinical trials often relies on pharmacokinetic–pharmacodynamic (PK/PD) models, where limited data, inter-subject variability and safety constraints make optimisation challenging. We introduce \textbf{QuSOP}, a hybrid quantum-classical framework for safe dose selection in Phase 1 trials. QuSOP combines a dual long short-term memory (LSTM) pipeline with Monte Carlo estimation to identify the minimal dose (d) whose clinical success probability (p(d)) exceeds a pre-specified safety threshold (\pi). The first LSTM predicts plasma concentration (PK) trajectories from time-dependent dosing, while the second predicts biomarker (PD) trajectories from the inferred PK profile. Synthetic data expansion is used to augment pre-trained models, enabling extrapolation to higher doses, longer dosing intervals and different body-weight ranges.
We extend QuSOP in two quantum directions. First, we investigate variational quantum circuits and quantum kernels as quantum-augmented components within the PK/PD pipeline. Second, we present a proof-of-concept quantum Monte Carlo extension using quantum amplitude estimation to compute (p(d)), offering a theoretical quadratic improvement in sample complexity over classical Monte Carlo.
We evaluate QuSOP on two Quantum Innovation Challenge datasets: a biomarker-suppression task and a biomarker-elevation task. For once-daily dosing, the classical dual-LSTM pipeline recommends (d^{}_{\mathrm{DS1,QD}}(90%) \approx 8.5) mg and (d^{}{\mathrm{DS2,QD}}(90%) \approx 0.46) mg. Once-weekly recommendations are (d^{}{\mathrm{DS1,QW}}(90%) \approx 85) mg and (d^{}_{\mathrm{DS2,QW}}(90%) > 10) mg. Stress tests identify COMED as the dominant Dataset 1 covariate.
| Topic | Quantum algorithms for machine learning applications |
|---|---|
| Additional Topic | Hybrid quantum-classical approaches for learning and optimization |