PhD Scientific Days 2026

Poster Session 2.G - Pharmaceutical Sciences and Health Technologies

Machine Learning Methods for Predicting Chiral Separations

Előadó neve

Mr. Dombi, Gergely

Neptune code

X6EXGC

Előadó munkahelye

Semmelweis University Department of Pharmaceutical Chemistry

Előadó telefonszáma

+36300104450

Előadó e-mail címe

dombi.gergely@semmelweis.hu

Az előadás címe

Machine Learning Methods for Predicting Chiral Separations

Szerző(k) neve és munkahelye

Gergely Dombi¹, Attila Imre², Ali Mhammad¹, Arash Mirzahosseini¹, Anita Rácz³, Gergő Tóth¹

1: Semmelweis University Department of Pharmaceutical Chemistry
2: Semmelweis University Center for Health Technology Assessment
3: HUN-REN Institute of Materials and Environmental Chemistry

Bemutatás módja

Poszter

Szekció

Poster Session 2.G - Pharmaceutical Sciences and Health Technologies

Language of the presentation

English

Preferred session

Pharmaceutical Sciences and Health Technologies

Összefoglaló szövege

Introduction: Method development for enantioseparation in HPLC remains a difficult task because it is still based on extensive trial-and-error screening.
Aim: Our work focuses on predicting separations on polysaccharide-based chiral stationary phases (CSPs) using machine learning based methods.
Methods: Using a Lux Cellulose-1 column in polar organic mode we generated a homogenous in-house dataset which contained 535 structurally diverse compounds yielding 1,414 retention time measurement under four mobile phases: acidic and basic methanol, acidic and basic acetonitrile. On this pilot dataset two modelling approaches were investigated. A consensus model, which combined partial least squares regression with neural networks; also, a newer graph neural network (GNN) was built. Broadening the applicability, a new dataset was also constructed using Lux Amylose-1 and i-Amylose-1 columns using the same eluents. This dataset consisted of 2063 unique retention time data from 625 structurally diverse molecules on this dataset only a GNN was used.
Results: On the original dataset the consensus model performed best when each solvent was considered separately, reaching R2 values above 0.70. The GNN model performed better with the unified dataset achieving an R2 of 0.85. Using the GNN model it provided correct predictions of stereoisomer elution orders for approximately 69-82% of experimentally separated pairs. An open access web server, chiralscreen.com was built where users can input SMILES codes and get retention times. The GNN model was also validated using stereoisomers that were not in the learning dataset, providing a 75% prediction accuracy. On the new dataset R2 values of 0.64 could be achieved.
Conclusion: This study shows that prediction of retention times and separation of analytes on polysaccharide based CSPs still has difficulties, however, machine learning can substantially reduce the time and effort required for trial-and-error screening and provide a useful tool for early analytical decision making.
Funding: Supported by the 2025-2.1.1-EKÖP-2025-00014 University Research Scholarship Programme of the Ministry for Culture and Innovation from the Source of the National Research, Development and Innovation Fund (G.D.). This work was funded by the National Research, Development, and Innovation Office, Hungary (grant: NKFIH FK 146930) (G.T).

University

Semmelweis University

Supervisor

Dr. Gergő Tóth

Publication of my abstract

I do not give consent to the publication of my abstract on the website of the congress.

phd.section.field

in doctoral studies after complex exam (PhD)

Kind

Szabad

Status

elfogadva

Accepted presentation method

poszter

Előadás fájl jóváhagyás

nem rendelkezett róla

Előadó

8931

Start

18:06

End

18:09