Utku Umur Acikalin

Cornell Computer Science · Gomes Lab

AI for molecular
discovery.

I'm Utku, a fourth-year Ph.D. student at Cornell University, advised by Carla P. Gomes.

I work on large-scale scientific problems where standard deep learning struggles, and build methods that use the structure of the problem instead of ignoring it: chemical rules, graph structure, and combinatorial search. My main focus is identifying unknown molecules from tandem mass spectra.

Interpretability is central to this work. AIMe explains each predicted peak and its intensity with a mechanistic fragmentation pathway. GEMS links spectral evidence to specific atoms and bonds, and every step of its search is a valid molecule a chemist can inspect. I also want these methods to be usable by scientists without an AI background, which is why AIMe runs as a public search system over 105M molecules.

I also work on language models and their applications in financial economics, learned combinatorial optimization, and efficient deep learning.

Recent updates

Research

Publications and preprints

My main research question is how to combine learned models with scientific constraints to recover molecular structure from incomplete experimental evidence.

Molecular discovery · Neuro-symbolic AI

AIMe & MS²KOSMOS

AIMe combines learned spectral prediction with chemically constrained fragmentation, reaching 0.83 cosine similarity on NIST vs. 0.74 for the strongest baseline. MS²KOSMOS makes the resulting molecular space searchable, with peak-level structural annotations and fragmentation-pathway explanations.

I co-designed and implemented AIMe and co-developed its public interface. The collaborative study used the system to elucidate previously unreported microbiota-dependent polyamines, with validation by chemical synthesis.

Co-first author · bioRxiv 2026; under review at Nature

Read the preprint Try the system

Molecular structure · Learned graph search

GEMS

GEMS treats known-formula structure elucidation as learned search over isomer graphs. Starting from database candidates, it refines complete structures through bond edits that preserve the molecular formula and every atom's valence, so it can reach molecules outside the database.

A policy links fingerprints predicted from the spectrum by MIST to the candidate's atoms and bonds, and scores every valid edit in one forward pass. Shortest-path supervision allows pretraining on structures alone. With the true structure excluded from the starting set, GEMS reaches 30.2% Hit@1 on MassSpecGym and 29.6% on NPLIB1.

First author · under review at ICLR 2027 · earlier version: ICML 2026 workshops, GenBio spotlight and AI4Physics poster

Read the workshop paper

Deep learning · COLM 2026

MoRE: Mixture of Reused Experts

A middle ground between standard mixture-of-experts models and looped (Universal) Transformers: adjacent layers share one expert pool but keep their own routers, with depth embeddings so shared experts can tell layers apart. This expands routing choices about 18x without adding parameters, and beats a standard DeepSeek-v3-style MoE and MoEUT from 114M to 1.15B parameters at matched parameters and FLOPs.

Paper arXiv Code

Optimization · ICLR 2025

X²GNN: exploration and exploitation

Unsupervised graph-neural optimization for Maximum Independent Set, Maximum Clique, and Max-Cut, combining exploration with iterative solution improvement. Trained on a single refinement step, it can be iterated much longer at test time, trading inference compute for solution quality: within 1.2% of optimal on large Max Clique benchmarks, where prior neural methods stay above 22%, and ahead of Gurobi at similar runtime.

Paper Code

Financial economics · JFE 2026

Language models and their applications in financial economics

Fine-tuned long-context language models to score 643K patents for exposure to the Supreme Court's Alice decision. The measure shows large firms gaining and small firms losing after patent protection weakened.

Paper NBER

Selected publications

Google Scholar

* Equal contribution. Preprints and workshop papers are labeled explicitly.

  1. Intellectual property protection lost and competition: An examination using large language models.U. U. Acikalin, T. Caskurlu, G. Hoberg, G. M. Phillips.Journal of Financial Economics 182:104306, 2026.
  2. Charting the small-molecule universe from mass spectra with neuro-symbolic AI.U. U. Acikalin*, D. Feng*, A. M. Ferber*, G. J. Gouveia*, et al.bioRxiv 2026; under review at Nature.
  3. GEMS: Molecular Structure Elucidation as Learned Search over Isomer Graphs.U. U. Acikalin, Y. Wang, G. J. Gouveia, T. Schwertfeger, F. C. Schroeder, C. P. Gomes.Under review at ICLR 2027. Earlier version: ICML 2026 workshops, GenBio spotlight; AI4Physics poster.
  4. MoRE: Mixture of Reused Experts.E. S. Qiu*, U. U. Acikalin*, J. Lovelace, C. Belardi, A. B. Mulchandani, C. P. Gomes, K. Q. Weinberger.COLM 2026, accepted.
  5. Learning to Explore and Exploit with GNNs for Unsupervised Combinatorial Optimization.U. U. Acikalin, A. M. Ferber, C. P. Gomes.ICLR 2025.
  6. Unsupervised Combinatorial Probabilistic Reasoning: Probabilistic Coin Change Problem.Z. Qu, Y. Wang, U. U. Acikalin, et al.AAAI 2026.
Earlier publications
  1. Does Public Procurement Promote Markets for New Technologies? Implications for Competition and Small Firms.U. U. Acikalin, M. Kutlu, R. Palalic, B. K. O. Tas.Strategic Change, 2025.
  2. Models for test cost minimization in database migration.B. Caskurlu, K. Subramani, U. U. Acikalin, A. Velasquez, P. Wojciechowski.INFORMS Journal on Computing 37(2):249-269, 2025.
  3. Solving Secluded Path Problem by Heuristics and Local Search.U. U. Acikalin, B. Caskurlu, O. Abul.ACM AICONF 2024.
  4. Security-aware database migration planning.U. U. Acikalin, B. Caskurlu, K. Subramani.Constraints 28(3):472-505, 2023.
  5. How you describe procurement calls matters: Predicting outcome of public procurement using call descriptions.U. U. Acikalin, M. K. Gorgun, M. Kutlu, B. K. O. Tas.Natural Language Engineering, 2023.
  6. On approximate Nash equilibria of the two-source connection game.B. Caskurlu, U. U. Acikalin, F. E. Kizilkaya, O. Ekici.Turkish Journal of Electrical Engineering and Computer Sciences 30(6):2206-2220, 2022.
  7. Patent Search Using Triplet Networks Based Fine-Tuned SciBERT.U. U. Acikalin, M. Kutlu.ACM SIGIR Patent Text Mining and Semantic Technologies Workshop, 2022.
  8. Multilevel memetic hypergraph partitioning with greedy recombination.U. U. Acikalin, B. Caskurlu.ACM GECCO Companion, 2022.
  9. New Results on Test-Cost Minimization in Database Migration.U. U. Acikalin, B. Caskurlu, P. Wojciechowski, K. Subramani.ALGOCLOUD 2021.
  10. Turkish Sentiment Analysis Using BERT.U. U. Acikalin, B. Bardak, M. Kutlu.IEEE SIU, 2020.
  11. Security-Aware Database Migration Planning.K. Subramani, B. Caskurlu, U. U. Acikalin.ALGOCLOUD 2019.

Background

At Cornell, I work with Carla P. Gomes and collaborate with Frank Schroeder's chemistry group, so our methods are applied to real experimental data, not only benchmarks. In the AIMe study, this led to previously unreported microbiota-dependent polyamines, confirmed by chemical synthesis.

Before Cornell, I earned my M.S. and B.S. in Computer Engineering at TOBB University of Economics and Technology in Ankara, Turkey, working on combinatorial optimization and on NLP for patents and public procurement.

Both threads carried forward. Optimization turned into learned search, first with X²GNN and then with GEMS over molecular graphs. Patent NLP turned into our Journal of Financial Economics paper, where a fine-tuned language model became a patent-exposure measure that economists could use directly.

Research résumé (PDF) · HTML version

Education

  • Cornell UniversityPh.D. in Computer ScienceAdvisor: Carla P. Gomes2023 - present
  • TOBB University of Economics and TechnologyM.S. in Computer Engineering · GPA 4.00/4.002019 - 2022
  • TOBB University of Economics and TechnologyB.S. in Computer Engineering · GPA 3.74/4.002015 - 2019

Research & collaboration

Interested in molecular machine learning, AI4Science, and language models.

uuacikalin@gmail.com