Active Learning on Synthons for Molecular Design

TL;DR

This work addresses the combinatorial bottleneck of exhaustive virtual screening in drug discovery, particularly under multi-vector expansion. The authors introduce SALSA, a synthon-based active-learning framework that factorizes learning and acquisition across independent synthons, reducing inference from a prohibitive $O(\prod_i |\mathcal{S}_i|)$ to a manageable $O(\sum_i |\mathcal{S}_i|)$. SALSA employs per-vector surrogate models (MPNNs with mean-variance estimation) and approximate Thompson sampling to iteratively propose high-scoring synthons, assemble molecules, and update data across $N$ rounds with a per-round budget $K$. Experiments show SALSA is highly sample-efficient, scalable to spaces up to ~2 trillion compounds, and capable of optimizing multi-parameter objectives across protein targets, achieving competitive MPO scores and superior scaffold diversity compared to LibINVENT. The approach provides explicit synthesis-aware control, embedded synthetic routes, and practical advantages for real-world medicinal chemistry workflows, representing a scalable alternative to generative methods in constrained, synthesizable design tasks.

Abstract

Exhaustive virtual screening is highly informative but often intractable against the expensive objective functions involved in modern drug discovery. This problem is exacerbated in combinatorial contexts such as multi-vector expansion, where molecular spaces can quickly become ultra-large. Here, we introduce Scalable Active Learning via Synthon Acquisition (SALSA): a simple algorithm applicable to multi-vector expansion which extends pool-based active learning to non-enumerable spaces by factoring modeling and acquisition over synthon or fragment choices. Through experiments on ligand- and structure-based objectives, we highlight SALSA's sample efficiency, and its ability to scale to spaces of trillions of compounds. Further, we demonstrate application toward multi-parameter objective design tasks on three protein targets - finding SALSA-generated molecules have comparable chemical property profiles to known bioactives, and exhibit greater diversity and higher scores over an industry-leading generative approach.

Figures (10)

• Figure 1: A Construction of a 2-vector synthon space. B AL loop against scoring function $f$.
• Figure 2: Recall of top-1K compounds in the 1M target space for ROCS-TC (top) and docking (bottom) as a function of molecules acquired, smoothed over 5 trials. The heatmaps show the enumerated target space decomposed across synthon axes and coloured by score. For a given SALSA round, synthons are ordered by Monte Carlo-estimated acquisition probability, i.e. the likelihood of sampling increases moving up and right. Top-1K ground truth molecules are highlighted in red.
• Figure 3: The large violin plots show the min, max, mean, and estimated score density for the top-1K molecules identified by SALSA as space size increases for shape- (left) and structure-based (right) objectives, smoothed over 3 trials. Subplots on the right show the evolution of the top-1K distribution over AL rounds – the final top-1K molecules are marked in red at their sample index.
• Figure 4: A. Top-1K molecules identified across three targets for SALSA, random acquisition, and LibINVENT using QED Bickerton plus ROCS-TC and Hybrid Docking objectives. Mean scores and top-20 pareto optimal molecules are denoted by triangles and stars, respectively. B. Number of unique Bemis-Murcko scaffolds above a given score value for molecules identified by SALSA and LibINVENT. SALSA compounds show substantially greater diversity.
• Figure 5: Core scaffolds mapped to synthetic intermediates with functionalized reaction handles.