Research

Publications

Publications in journals with peer-reviewing

J5. Definition and Exploration of Realistic Chemical Spaces Using the Connectivity and Cyclic Features of ChEMBL and ZINC

Thomas Cauchy, Jules Leguy, Benoit Da Mota
2023 Digital Discovery
DOi : 10.1039/D2DD00092J   scimago : Rank Q1 (chemistry)   github : BenoitDamota/gcf   figshare : data

J4. Scalable estimator of the diversity for de novo molecular generation resulting in a more robust QM dataset (OD9) and a more efficient molecular optimization

Jules Leguy, Marta Glavatskikh, Thomas Cauchy, Benoit Da Mota
2021 Journal of Cheminformatics
DOi : 10.1186/s13321-021-00554-8   scimago : Rank Q1 (CS applications)   github : jules-leguy/EvoMol   figshare : data

J3. EvoMol: a flexible and interpretable evolutionary algorithm for unbiased de novo molecular generation

Jules Leguy, Thomas Cauchy, Marta Glavatskikh, Béatrice Duval, Benoit Da Mota
2020 Journal of Cheminformatics
DOi : 10.1186/s13321-020-00458-z   scimago : Rank Q1 (CS applications)   github : jules-leguy/EvoMol

J2. MixONat, a Software for the Dereplication of Mixtures Based on 13C NMR Spectroscopy

Antoine Bruguière, Séverine Derbré, Joël Dietsch, Jules Leguy, Valentine Rahier, Quentin Pottier, Dimitri Bréard, Sorphon Suor‑Cherer, Guillaume Viault, Anne‑Marie Le Ray, Frédéric Saubion, Pascal Richomme
2020 Analytical Chemistry
DOi : 10.1021/acs.analchem.0c00193   scimago : Rank Q1 (analytical chemistry)   sourceforge : mixonat

J1. Dataset's chemical diversity limits the generalizability of machine learning prediction

Marta Glavatskikh, Jules Leguy, Gilles Hunault, Thomas Cauchy, Benoit Da Mota
2019 Journal of Cheminformatics
DOi : 10.1186/s13321-019-0391-2   scimago : Rank Q1 (CS applications)   figshare : data

Publications in conferences with peer-reviewing

C1. Surrogate‑Based Black‑Box Optimization Method for Costly Molecular Properties

Jules Leguy, Béatrice Duval, Benoit Da Mota, Thomas Cauchy
2021 IEEE 33rd International Conference on Tools with Artificial Intelligence (ICTAI)
DOi : 10.1109/ICTAI52525.2021.00124   ↕️ Conference Ranks : Rank A2 Qualis/B ERA   github : jules-leguy/BBOMol

Publications in national conferences with peer-reviewing

CN1. Des réseaux de neurones pour prédire des distances interatomiques extraites d'une base de données ouverte de calculs en chimie quantique

Publications of book chapters with peer-reviewing

B2. Goal‑directed generation of new molecules by AI methods (Chapter 2)

Jules Leguy, Thomas Cauchy, Béatrice Duval, Benoit Da Mota
2022 Computational and Data‑Driven Chemistry Using Artificial Intelligence
DOi : 10.1016/B978-0-12-822249-2.00004-9

B1. Predicting Interatomic Distances of Molecular Quantum Chemistry Calculations (long version of [CN1])

Jules Leguy, Thomas Cauchy, Béatrice Duval, Benoit Da Mota
2022 Advances in Knowledge Discovery and Management: Volume 9. Studies in Computational Intelligence. Springer International Publishing
DOi : 10.1007/978-3-030-90287-2_8

Publications in conference workshops with light peer-reviewing

W1. Génération d'explications contre‑factuelles pour la chimie moléculaire

Jules Leguy, Bryan Garreau, Thomas Cauchy, Benoit Da Mota, Béatrice Duval
2022 Workshop EXPLAIN'AI hosted at EGC 2022, Blois, France (National-level conference)
HAL open archive: hal-04034150   ↕️ Conference Ranks : EGC Rank C ERA

Posters

Po1. Surrogate-based black-box framework to optimize electronic properties for de novo organic molecular materials

Preprints

P1. WebXAII: An Open-Source Web Framework to Study Human-XAI Interaction

Jules Leguy, Pierre-Antoine Jean, Felipe Torres Figueroa, Sébastien Harispe
2025 arXiv:2506.14777 [cs.HC]
DOi : 10.48550/arXiv.2506.14777   github : PAJEAN/WebXAII

PhD thesis

My Ph.D thesis manuscript entitled "Combinatorial search guided by machine learning for molecular chemistry" (2022) is available on HAL (in french). The defense slides can be found here.

Contributions to research projects

XAI and Human-XAI interaction (postdoctoral work, current)

AI-assisted decision-making systems are increasingly deployed in high-stakes domains such as medicine and military applications. In these contexts, the appropriate use of algorithmic assistance requires adequate calibration of human operators' trust and reliance. Reliance refers to the observed behavior of agreeing with the predictions of AI models. Inadequate calibration can lead to under-use of the assistance, or conversely to over-reliance, where the user delegates decisions to the system without critical evaluation.

Explainable AI (XAI) can be seen as an intermediary between the often opaque functioning of predictive models and human reasoning and decision-making processes. In the context of AI-assisted decision making, XAI is generally expected to support appropriate calibration of trust and reliance toward predictive models. Contextual factors such as time pressure and task difficulty can further affect this calibration.

I propose a large-scale study (600 human subjects) of human-computer interaction in this AI-assisted decision-making context, for the resolution of a pattern detection task. The protocol is designed to evaluate the impact of several explainability techniques and paradigms (SHAP, counterfactual explanations, GradCAM, multimodal explanations generated by LLMs), and to assess the influence of time pressure and task difficulty on human behavior.

This study was approved by the ethics committee of the University of Montpellier and pre-registered before data collection (https://osf.io/56wnj/). I am currently working on the presentation of these results in 2 scientific articles.

The experiments described above were implemented in a web platform named WebXAII, specialized for the study of human-XAI interaction, whose development I led and which enables strong reproducibility of experiments [P1].

Optimization and machine learning for molecular discovery (PhD work, 2019-2022)

The main goal of this PhD in Computer Science was to develop methods for the automatic generation of molecules satisfying target properties, with a focus on the domain of organic molecular materials chemistry. This domain is characterized by the fact that it is less explored than other areas of chemistry, and that many molecular properties of interest depend on quantum chemistry computations that can be extremely costly (from minutes to hundreds of hours).

Combinatorial search. I proposed an evolutionary algorithm named EvoMol for exploring this search space. It operates on molecular graphs through a set of simple mutation operators, in order to limit and control assumptions about the search space [J3]. EvoMol achieves results comparable to the best state-of-the-art methods. I also proposed a tree-based visualization of the search exploration, enabling a degree of interpretability of the search process. EvoMol has become a reference algorithm in the literature for optimizing molecular properties.

Diversity optimization. Most target properties in our application domain depend on costly quantum chemistry evaluations, which motivates the use of machine learning models as low-cost estimators. Working with a postdoctoral researcher, we showed that a reference molecular dataset (QM9) poses generalization issues when used to train a predictive model [J1]. I proposed a molecular diversity optimization algorithm based on EvoMol and an efficient approximation of Shannon entropy applied to a set of molecular descriptors [J4].

Surrogate-based black-box optimization. I proposed a black-box optimization approach based on a surrogate model, combining an optimization method with a machine learning model of molecular properties. The surrogate model predicts values of the costly target property to select promising candidates in the search space. I showed that this approach can be more efficient than an evolutionary algorithm for the optimization of an electronic property, in terms of calls to the objective function [C1, Po1].

Explainability. The use of machine learning models for chemistry raises questions about their interpretability. I proposed an approach based on EvoMol to generate counterfactual explanations for any binary classification model of a molecular property [W1].

Realistic molecular generation. Local search methods for molecular optimization tend to generate molecules considered unrealistic by chemists. However, defining a measure of molecular "realism" is very difficult as it is a subjective notion based on expert knowledge. We proposed a filtering approach based on whitelists of molecular features to generate more realistic molecules [J5].

I also published a state-of-the-art review of the field of automatic molecular generation as a book chapter [B2].

Prediction of molecular geometry (Master's internship work, 2018-2019)

During my Master's research internships (M1 and M2), I worked on problems related to predicting the results of costly quantum chemistry computations for estimating molecular properties. These computations are closely related to computing a "converged molecular geometry", which can be represented as the relative positioning of atoms in 3D Cartesian space. We hypothesized that obtaining a good estimate of this geometry would reduce the total computation cost. I worked on defining a predictive machine learning model for this geometry. We observed that the global geometry prediction problem was very complex, and we proposed to solve a sub-problem consisting in predicting distances between pairs of atoms. I proposed a predictive model that achieves good results for these interatomic distances [CN1, B1].

Dereplication in vegetal-based chemistry (Master's project work, 2019)

I worked with a group of scientists in vegetal chemistry during my MSc studies, on a project aiming to improve a software tool (then at the prototype stage) that uses NMR spectra for the identification of compounds in a mixture. I reorganized the source code, formalized the main algorithm and improved its algorithmic efficiency. The tool was subsequently publicly distributed and has become a reference in the literature [J2].