Lookahead Bayesian Optimisation for Molecular Growth in Self-Driving Laboratories
- 1Faculty of Engineering
- 2University Park Nottingham
Abstract
Self-driving laboratories can autonomously execute iterative design-make-test-analyse cycles, but molecular optimisation presents a distinctive challenge: selecting which compound to elaborate determines not only the next experiment, but also the chemical space accessible in subsequent rounds. Conventional molecular Bayesian optimisation typically evaluates compounds without considering how their selection changes what can subsequently be synthesised. Here, we introduce Lookahead Bayesian Optimisation (LABO), which instead values candidate molecules according to the predicted value of the synthetically accessible chemical space they unlock, enabling optimisation to account for future experimental opportunities. We evaluate LABO in simulated five-round molecular optimisation campaigns across 30 lead-like starting compounds and five kinase targets. LABO outperformed established molecular optimisation strategies from the literature across all 30 optimisation landscapes, as well as random, greedy and Bayesian optimisation controls. LABO was shown to improve ligand-lipophilicity efficiency by 1.25 units over random selection, equivalent to approximately 18-fold greater potency at equal lipophilicity, and by 0.34 units (2.2-fold) over greedy elaboration. These results demonstrate a general principle for autonomous molecular optimisation: under synthetic constraints, the value of a molecule depends not only on its predicted properties, but on the future experiments that selecting it makes possible. By explicitly incorporating this future experimental value into molecular selection, LABO provides a framework for planning successive rounds of molecular growth in self-driving laboratories.
Related
- What Can the Records of Published Self-Driving Laboratories Predict About a New Platform? A Step-Level Database and Explorer
- Autonomous Robotic Workflow for Solution-Based Synthesis and High-Temperature Thermal Processing toward Self-Driving Laboratories
- La Agente Óptima: Towards Agentic Self-Driving Laboratories
- Talking to instruments: device-integration standards for self-driving laboratories
- Two-stage cooperative optimisation of signal timing and vehicle trajectories for eco-driving at signalised intersections under HDV/CAV/CCAV mixed traffic
BibTeX
@misc{taylor2026lookahead,
title = {{Lookahead Bayesian Optimisation for Molecular Growth in Self-Driving Laboratories}},
author = {Connor J. Taylor},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15009896/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009896/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Lookahead Bayesian Optimisation for Molecular Growth in Self-Driving Laboratories
AU - Connor J. Taylor
PY - 2026
DA - 2026/10/05
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009896/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009896/v1
AB - Self-driving laboratories can autonomously execute iterative design-make-test-analyse cycles, but molecular optimisation presents a distinctive challenge: selecting which compound to elaborate determines not only the next experiment, but also the chemical space accessible in subsequent rounds. Conventional molecular Bayesian optimisation typically evaluates compounds without considering how their selection changes what can subsequently be synthesised. Here, we introduce Lookahead Bayesian Optimisation (LABO), which instead values candidate molecules according to the predicted value of the synthetically accessible chemical space they unlock, enabling optimisation to account for future experimental opportunities. We evaluate LABO in simulated five-round molecular optimisation campaigns across 30 lead-like starting compounds and five kinase targets. LABO outperformed established molecular optimisation strategies from the literature across all 30 optimisation landscapes, as well as random, greedy and Bayesian optimisation controls. LABO was shown to improve ligand-lipophilicity efficiency by 1.25 units over random selection, equivalent to approximately 18-fold greater potency at equal lipophilicity, and by 0.34 units (2.2-fold) over greedy elaboration. These results demonstrate a general principle for autonomous molecular optimisation: under synthetic constraints, the value of a molecule depends not only on its predicted properties, but on the future experiments that selecting it makes possible. By explicitly incorporating this future experimental value into molecular selection, LABO provides a framework for planning successive rounds of molecular growth in self-driving laboratories.
ER -