Utku Umur Acikalin

AIMe & MS²KOSMOS: neuro-symbolic mass spectrometry at database scale

Co-first author · bioRxiv 2026, under review at Nature · Preprint · Public system

Outcome: synthesis-validated discovery of previously unreported microbiota-dependent polyamines.

• Co-designed and implemented AIMe, a neuro-symbolic MS/MS predictor that explains peaks through chemically constrained fragmentation pathways; 0.83 cosine on NIST vs. 0.74 for the strongest baseline.
• Built CPU/GPU retrieval over 105M structures and 800M+ predicted spectra (MS²KOSMOS), with interactive queries in a few seconds and a repository-scale search of 7.14M GNPS consensus spectra.
• Co-developed the public web system for spectral search, pathway inspection, and on-demand prediction.

GEMS: structure elucidation as learned search over isomer graphs

First author · under review at ICLR 2027 · ICML 2026 workshops: GenBio (spotlight), AI4Physics · Workshop paper

• Developed a learned search that refines candidates through formula- and valence-preserving bond edits; a policy maps MIST-predicted fingerprints to atoms and bonds and scores all edits in one pass.
• Designed shortest-path supervision, enabling pretraining on PubChem structures alone and fine-tuning on spectrum-predicted fingerprints.
• Hit@1 of 29.6% on NPLIB1 and 30.2% on MassSpecGym with the target excluded from initialization, above published baselines in comparable settings; higher accuracy than FRIGID on NPLIB1 in ~2.6 min vs. 26 h.

MoRE: Mixture of Reused Experts

Co-first author · COLM 2026 · arXiv · Code

• Co-developed a middle ground between standard MoE and looped (Universal) Transformers: adjacent layers share an expert pool with per-layer routers, giving ~18x more routing choices at no parameter cost.
• Beats a DeepSeek-v3-style MoE and MoEUT at 114M-1.15B parameters with matched parameters and FLOPs.

X²GNN: learning to explore and exploit

First author · ICLR 2025 · Code

• Unsupervised GNN solver for MIS, Max Clique, Max‑Cut; trained on one refinement step, it scales with test‑time compute: 1.2% from optimal on large Max Clique graphs (prior neural: >22%), beating Gurobi at similar runtime.