FSscore: A Personalized Machine Learning-based Synthetic Feasibility Score - Focused Synthesizability score - Human-in-the-loop

Aug 5, 2026

Upload by Hazel

Assessing whether a molecule can be synthesized is central to chemistry and drug discovery, because it steers experimental prioritization and molecule ranking in de novo design. Existing synthetic-feasibility scorers generalize poorly to new chemical spaces and may miss fine-grained distinctions such as chirality. FSscore introduces a Focused Synthesizability score learned from machine learning, using a baseline on reactant–product pairs refined through expert human feedback per chemical space. This targeted fine-tuning better separates hard versus easy-to-synthesize molecules and demonstrates an effective human-in-the-loop framework for practical synthetic feasibility evaluation across applications.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages22

Reading Time55 min

AccessFree to download

Introduction

Importance of synthetic feasibility

Limitations of current scoring methods

Human feedback and human-in-the-loop learning

Related structure- and reaction-based scores

arXiv :2312 . 12737v2 [ cs .LG] 5 Oct 2024 FSscore: A Personalized Machine Learning-based Synthetic Feasibility Score Rebecca M. Neeser, ∗ ,†,‡ Bruno Correia,‡ and Philippe Schwaller†,¶ †Laboratory of Artificial Chemical Intelligence (LIAC), EPFL, Switzerland ‡Laboratory of Protein Design and Immunoengineering (LPDI), EPFL, Switzerland ¶National Centre of Competence in Research (NCCR) Catalysis, EPFL, Switzerland E-mail: {rebecca.neeser,bruno.correia,philippe.schwaller}@epfl.ch Abstract Determining whether a molecule can be synthesized is crucial in chemistry and drug discovery, as it guides experimental prioritization and molecule ranking in de novo design tasks. Existing scoring approaches to assess synthetic feasibility struggle to extrapolate to new chemical spaces or fail to discriminate based on subtle differences such as chirality. This work addresses these limitations by introducing the Focused Synthesizability score (FSscore), which uses machine learning to rank structures based on their relative ease of synthesis. First, a baseline trained on an extensive set of reactant-product pairs is established, which is then refined with expert human feedback tailored to specific chemical spaces. This targeted fine-tuning improves performance on these chemical scopes, enabling more accurate differentiation between molecules that are hard and easy to synthesize. The FSscore showcases how a human-in-the-loop framework can be utilized to optimize the assessment of synthetic feasibility for various chemical applications. 1 Introduction Assessing the synthetic feasibility of a small molecule is of great importance in many different areas of chemistry, notably in early drug discovery stages. Trained chemists traditionally perform this task through intuition or retrosynthetic analysis, allowing them to decide which molecules are likely possible to synthesize and prioritize based on synthetic complexity. However, the chemical space that might be accessible is massive, and only a small fraction has been explored. 1,2 Furthermore, computational approaches such as virtual screening (VS) 3 in drug discovery or the recent surge of generative methods for de novo molecular design 4–14 emphasize the requirement for suitable tools to score synthetic feasibility quickly in an automated fashion. 15,16 The current state-of-the-art methodologies perform well at discriminating feasible from unfeasible molecules in the data distribution they were designed for but often fail to generalize. This is especially true for machine learning (ML) predictors that cannot capture such an abstract concept as synthesizability and when applying these scores in the context of generative models. 15,17–19 However, exploring new chemical space is of great interest specifically in the context of de novo design or new drug modalities such as synthetic macrocycles or proteolysis targeting chimeras (PROTACs) . On the other hand, synthetic feasibility cannot be merely captured by the structure as it depends on a chemist’s available resources and expertise. 20 Thus, incorporating human preference would greatly improve the practical utility of such ascore.21 Various methods to capture synthetic accessibility or complexity have been previously proposed and include structure-based 22–24 or reaction-based 25–29 methods. The commonly used Synthetic Accessibility score (SA score) is rule-based and penalizes the occurrence of fragments rarely found in a reference dataset and the presence of specific structural features.22 Thus, it captures more synthetic complexity than accessibility and fails to identify big complex molecules with mostly reasonable fragments. 15,20 The SYBA score was trained to distinguish existing synthesizable molecules from artificial complex ones but the performance was found to be sub-optimal. 17,23 Many of these structure-/fragment-based approaches are unable to capture small structural differences due to low sensitivity. Yu et al. 24 a

FAQs

What problem does FSscore address in synthetic feasibility prediction?

It targets two limitations: poor extrapolation to new chemical spaces and insufficient discrimination of subtle differences such as chirality.

How does FSscore incorporate expert human knowledge?

It trains a baseline model on reactant–product pairs, then refines the model with human expert feedback tailored to specific chemical spaces.

What does FSscore aim to improve in practical molecule ranking?

It improves differentiation between molecules that are hard versus easy to synthesize, enabling more accurate prioritization in de novo design workflows.

Aug 5, 2026

Upload by Hazel