MSc Thesis Defense: Ekin Başar Gökçe, GENERATING SYNERGISTIC DRUG PARTNERS USING CHEMICAL LANGUAGE MODELS, Date & Time: 17 July, 2026 – 3:00 PM, Place: FENS L027

GENERATING SYNERGISTIC DRUG PARTNERS USING CHEMICAL

LANGUAGE MODELS

Ekin Başar Gökçe
Data Science, MSc Thesis, 2026

Thesis Jury

     Assoc. Prof. Öznur TAŞTAN OKAN (Thesis Advisor)

   Asst. Prof. Dilara KEKÜLLÜOĞLU

  Prof. Arzucan ÖZGÜR

Date & Time: July 17th, 2026 – 3.00 PM

Place: FENS L027

Keywords : Chemical Language Models, Conditional Molecular Generation, Drug Synergy, Molecular Generation, Drug Combination Therapy

Abstract

Experimental identification of synergistic drug combinations is limited by the vast number of possible drug pairs, making exhaustive screening impractical. Existing computational methods predict the synergy of predefined drug pairs that are already available in screening libraries. This thesis focuses on synergistic partner discovery and formulates it as a conditional molecular generation problem. Given an anchor drug, a target cancer cell line, and a target synergy score, the objective is to generate a molecular structure predicted to act synergistically with the anchor drug. To address this problem, we adapt two pretrained chemical language models, GP-MoLFormer and LLamoL, for conditional molecular generation using model-specific conditioning strategies. We systematically evaluate multiple and single conditioning mechanisms and fine-tuning protocols on the DrugComb and NCI-ALMANAC datasets. We evaluate the generated molecules using chemical validity, uniqueness, novelty, structural similarity, and distributional fidelity. We also assess their predicted synergy using an independent supervised drug synergy prediction model. Both models generate chemically valid molecules while responding to the conditioning signals. GP-MoLFormer achieves higher validity and molecular novelty, whereas LLamoL more closely reproduces the distribution of experimentally observed synergistic partner molecules. Comparisons with unconditional generation show that conditioning guides generation toward molecules that are more closely aligned with the target drug combination. We demonstrate that pretrained chemical language models can be adapted for conditional molecular generation of synergistic drug partners. Our approach extends computational drug synergy research beyond predicting existing drug pairs by generating new candidate partner molecules for combination therapy.