CITATION.cff

cff-version: 1.2.0
message: "If you refer to this work, please cite it as below."
title: "TopoResponse: does persistent homology help an E(3)-equivariant model predict molecular dipole moments and polarizability tensors?"
type: software
authors:
  - name: "Mindvisio"
license: MIT
repository-code: "https://github.com/Mindvisio/topo-response"
url: "https://mindvisio.github.io/topo-response/"
date-released: "2026-07-23"
keywords:
  - equivariant neural networks
  - persistent homology
  - topological data analysis
  - molecular property prediction
  - negative results
abstract: >-
  A controlled study of topological conditioning for molecular response properties.
  Across five seeds on a topology-out-of-distribution split, persistent-homology
  conditioning showed no advantage over the plain equivariant baseline or over
  matched-capacity random control; linear and small nonlinear residual
  probes on the frozen baseline likewise found no beneficial correction. A qualitative
  negative, specific to this descriptor, conditioning scheme, dataset and split.