Thao Nguyen

Research profile

Research summary

Thao Nguyen is a researcher at the University of Illinois Urbana-Champaign working at the intersection of machine learning and molecular science. The visible record consists of three first-authored 2026 arXiv preprints reviewed by Evolve, spanning the three fields that make up this profile evenly (~33% each): generative molecular design, AI for drug discovery / computational chemistry, and molecular property prediction. The through-line is a coherent 'ML for molecules' agenda executed at three levels: Fraglingo generates molecules through attachment-aware autoregressive fragment generation with retrieval-augmented decoding; SpecOpt optimizes existing drugs for binding specificity using an LLM agent with contact-diff reasoning against AutoDock Vina docking; and EnSol predicts solubility with an environment-aware graph neural network featuring cross-attention, FiLM temperature conditioning, and mixture-density-network uncertainty. Evolve's reproducibility assessment is the strongest measured signal — a mean of 85 across the three papers, with high readiness for SpecOpt and EnSol and medium for Fraglingo — while the same reviews flag specific weaknesses: a weakly validated docking proxy in SpecOpt, marginal Leeds gains and thin statistics in EnSol, and NJD losses to MolRWKV in Fraglingo. Collaboration is moderate (score 44, 8 distinct coauthors), anchored by recurring co-authors including Heng Ji, Jeonghwan Kim, Zhenhailong Wang, Saman Shafaei, Zhengyi Zhang, and Huimin Zhao. Output, impact, and activity metrics are not yet available and are left unscored.

Research keywords

Highlights

• SpecOpt: LLM-agent editing of drugs toward binding specificity

  First-author 2026 arXiv preprint. An LLM agent applies contact-diff reasoning to modify existing drug molecules and improve the gap between target and off-target docking scores computed with AutoDock Vina. Evolve's verdict: rigorous and reproducible, but resting on a weakly validated docking proxy; reproducibility readiness is rated high.

• EnSol: environment-aware graph neural network for solubility prediction

  First-author 2026 arXiv preprint that augments solubility GNNs with cross-attention, FiLM temperature conditioning, and mixture-density-network uncertainty, targeting solvent-screening use cases. Evolve finds the results credible, but notes that Leeds gains are marginal and the statistics thin; reproducibility readiness is rated high.

• Fraglingo: attachment-aware autoregressive fragment generation

  First-author 2026 arXiv preprint combining latent retrieval of fragments with attachment-aware autoregressive decoding for fragment-based molecular design. Evolve judges the work well-ablated but notes that plain BRICS trails MolRWKV on NJD and that headline wins rest on the BFE metric; reproducibility readiness is rated medium.

Timeline

1. 2026First-authored SpecOpt appears on arXiv: contact-diff reasoning for agentic molecule optimization toward binding specificity; Evolve rates it rigorous and reproducible with high reproducibility readiness.
2. 2026First-authored EnSol appears on arXiv: an environment-aware graph neural network for solubility prediction with cross-attention, FiLM temperature conditioning, and MDN uncertainty.
3. 2026First-authored Fraglingo appears on arXiv: attachment-aware autoregressive fragment generation for molecular design; Evolve calls it well-ablated, with headline wins resting on BFE.

Papers in the Evolve catalog

• 2026-09-18first author

  SpecOpt: Contact-Diff Reasoning for Agentic Molecule Optimization Toward Binding Specificity

• 2026-09-17first author

  EnSol: an environment-aware graph neural network for molecular solubility prediction

• 2026-09-11first author

  Fraglingo: Molecular Design via Attachment-Aware Autoregressive Fragment Generation

Frequent co-authors