arXiv Science⌕ Search

arXiv · 2610.03992

MathAgent: Multi-Agent Optimization of Mathematical Invariants for Molecular Property Prediction

Abstract

Mathematical invariants provide complementary representations of molecular structure, but their predictive utility can vary across datasets and prediction tasks. For a new dataset, appropriate construction of invariant representations from algebraic topology, differential geometry, topological spectral theory, and commutative algebra is a challenging task, particularly when evaluating multiple representations is computationally expensive. We develop a Mathematical Agentic Model (MathAgent) to automatically optimize invariant representations in molecular property prediction. The framework combines dataset-aware representation design, low-cost candidate screening, and progressive evaluation. Language-model reasoning supports task interpretation, strategy proposal, and workflow coordination, while scientific computations and numerical decisions are carried out by deterministic tools. We evaluate MathAgent on protein--ligand binding-affinity and quantitative toxicity prediction tasks. Different datasets favor different combinations of mathematical invariants, demonstrating the need for dataset-adaptive representation selection rather than a fixed universal representation. Objective-conditioned executions further show that the framework can adjust its evaluation strategy according to user-specified goals. MathAgent therefore provides a systematic, evidence-driven, and traceable approach for optimizing mathematical invariant representations and reducing reliance on manual configuration and/or experimentation in molecular property prediction.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Yiming Ren, Josh V. Vermaas, Guo-Wei Wei. 2026-10-02. MathAgent: Multi-Agent Optimization of Mathematical Invariants for Molecular Property Prediction. https://arxiv.science/abs/2610.03992

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Discovery of new antiviral agents through artificial intelligence: in vitro and in vivo results

In this research, we employed a deep reinforcement learning (RL)-based molecule design platform to generate a diverse set of compounds targeting the neuraminidase (NA) of influenza A and B viruses. A total of 60291 compounds were generated, of which 86.5% displayed superior physicochemical properties compared to oseltamivir. After narrowing down the selection through computational filters, nine compounds with non-sialic acid-like structures were selected for in vitro experiments. We identified two compounds, DS-22-inf-009 and DS-22-inf-021 that effectively inhibited the NAs of both influenza A and B viruses (IAV and IBV), including H275Y mutant strains at low micromolar concentrations. Molecular dynamics simulations revealed a similar pattern of interaction with amino acid residues as oseltamivir. In cell-based assays, DS-22-inf-009 and DS-22-inf-021 inhibited IAV and IBV in a dose-dependent manner with EC50 values ranging from 0.29 uM to 2.31 uM. Furthermore, animal experiments showed that both DS-22-inf-009 and DS-22-inf-021 exerted antiviral activity in mice, conferring 65% and 85% protection from IAV (H1N1 pdm09), and 65% and 100% protection from IBV (Yamagata lineage), respectively. Thus, these findings demonstrate the potential of RL to generate compounds with promising antiviral properties.

q-bio.BM↗

Agent-guided epitope excision enables economical de novo binder design

De novo protein design campaigns remain bottlenecked by the inference compute time required to generate and rank candidates, limiting wider functional exploration. Here, we introduce and evaluate an epitope excision approach that truncates targets to local interface windows to accelerate both backbone diffusion and structure prediction. Across 15 targets and 3,666 designs, excised windows closely reproduced full-target binding interfaces (median CαRMSD of 0.37 Å) and made it practical to design binders against the 719-residue extracellular domain of fibroblast activation protein alpha (FAP) by cutting RFdiffusion and AlphaFold3 GPU time by ~ 88% in one of the epitopes. Pairing this workflow with an agentic optimization loop allowed the system to autonomously identify catalytic-site targeting strategies. Under a fixed compute budget, a human-in-the-loop (HITL) workflow raised the computational hit rate from 5.06% to 9.33% while further cutting GPU hours by ~ 42%. We also found that omitting target multiple sequence alignments during refolding roughly substantially increased the apparent yield. Wet-lab validation with yeast display identified nanomolar binders from the generated designs with the HITL and the fully autonomous agent (human-out-of-the-loop, HOOTL) arms recovering more enriched designs than the PI-alone baseline. Together, target excision and agent-guided search offer a practical way to scale de novo binder design to large, therapeutically relevant proteins.

q-bio.BM↗

Cross-species digital xenocortication model delineates molecular correlates of human brain connectome

Cross-species digital xenocortication (CSDX) model based on pairing of a human brain structure with one or more mouse brain structures was constructed and implemented to dissect the molecular correlates of the human brain connectome. Design of the CSDX model is based on the assumption that a human brain region and its named mouse counterpart occupy corresponding positions in an evolutionary conserved mammalian brain connectome plan, so that a projection strength measured between two mouse structures is informative about the relation between the two human structures paired with them. In the CSDX model, gene expression for the node was measured in the human brain, while connection strength between two nodes was measured in the mouse brain. CSDX models of different parcellations were employed to explore the potential regulatory associations of brain connectome patterns and constituents of protein-protein interaction (PPI) cascades of hybrid assembly pathways of endogenous multiprotein complexes. Metrics of continuous engagement of the PPI cascade nucleated protein dimers of CORUM multiprotein complexes have been identified as statistical components of the multivariate transcriptome-connectome associations. It accounts for additional connectome associated variations beyond the five measured axes in the 82 nodes CSDX model. Comparisons of the seven arm models of transcriptome-connectome associations demonstrate that the mouse derived model generally fits the mouse brain connectome better, while the human derived model assigns a significantly larger fraction of its transcriptome-connectome association to unique rather than net shared contributions. This study nominates elite panels of top scoring candidate protein dimers and CORUM multiprotein complexes for follow up mechanistic studies of their potential roles in structural functional features governing mammalian brain connectome.

q-bio.BM↗