{"paper":{"title":"Chem-GMNet: A Sphere-Native Geometric Transformer for Molecular Property Prediction","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Chem-GMNet, a sphere-native geometric transformer, outperforms same-sized ChemBERTa-2 on 7 of 10 MoleculeNet endpoints with about 35 percent fewer parameters.","cross_cats":["q-bio.QM"],"primary_cat":"cs.LG","authors_text":"Deepak Warrier, Raja Sekhar Pappala","submitted_at":"2026-05-13T09:43:55Z","abstract_excerpt":"Modern SMILES-based chemical language models obtain strong MoleculeNet performance by treating SMILES as generic text and compensating with multi-million-molecule self-supervised pretraining. We ask: when a domain carries structural priors as rich as chemistry's, does it warrant a domain-native transformer rather than a generic one rescued by scale? We answer affirmatively with \\textbf{GM-Net} (Geometric Measure Network), a transformer family in which every module is replaced by a sphere-native counterpart, and instantiate it as \\textbf{Chem-GMNet}. Three blocks follow: SH-Embedding (tokens as"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On canonical DeepChem scaffold splits, random-initialised Chem-GMNet wins on 7 of 10 MoleculeNet endpoints at ~35% fewer parameters than same-shape ChemBERTa-2 baselines; pretrained on the same 10M-SMILES ZINC corpus it matches or beats the public release on 6 of 8 shared endpoints.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the reported performance gains arise from the sphere-native inductive biases rather than from differences in training protocol, hyperparameter tuning, or data handling that are not fully detailed in the abstract.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Chem-GMNet uses sphere-native embeddings, DualSKA attention, and SH-FFN layers to match or beat ChemBERTa-2 on MoleculeNet tasks with fewer parameters and sometimes no pretraining.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Chem-GMNet, a sphere-native geometric transformer, outperforms same-sized ChemBERTa-2 on 7 of 10 MoleculeNet endpoints with about 35 percent fewer parameters.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"ab747a7f607569ac27790cb5fd86aef9a800cc076d95cdd23f5a6f62d7df4618"},"source":{"id":"2605.13262","kind":"arxiv","version":1},"verdict":{"id":"91a5a470-4dda-4f38-ad18-b0b104e36d88","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:48:15.988785Z","strongest_claim":"On canonical DeepChem scaffold splits, random-initialised Chem-GMNet wins on 7 of 10 MoleculeNet endpoints at ~35% fewer parameters than same-shape ChemBERTa-2 baselines; pretrained on the same 10M-SMILES ZINC corpus it matches or beats the public release on 6 of 8 shared endpoints.","one_line_summary":"Chem-GMNet uses sphere-native embeddings, DualSKA attention, and SH-FFN layers to match or beat ChemBERTa-2 on MoleculeNet tasks with fewer parameters and sometimes no pretraining.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the reported performance gains arise from the sphere-native inductive biases rather than from differences in training protocol, hyperparameter tuning, or data handling that are not fully detailed in the abstract.","pith_extraction_headline":"Chem-GMNet, a sphere-native geometric transformer, outperforms same-sized ChemBERTa-2 on 7 of 10 MoleculeNet endpoints with about 35 percent fewer parameters."},"references":{"count":37,"sample":[{"doi":"","year":2024,"title":"Kendall Atkinson and Weimin Han.Spherical Harmonics and Approximations on the Unit Sphere: An Introduc- tion, volume 2044 ofLecture Notes in Mathematics","work_id":"7f7dbda8-3cb4-472c-a440-a82227746c52","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1007/97","year":null,"title":"A Practical Guide to TPM 2.0","work_id":"0ea0856b-5e33-4766-9b0a-4041f79be461","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1038/s41467-022-29939-5","year":null,"title":"E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials","work_id":"f52ce8d5-e753-4360-a60f-60c8faae7c16","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1038/s41570","year":null,"title":"Boris Bonev, Thorsten Kurth, Tom Kölbl, et al","work_id":"9ddf75c9-5940-43de-abbb-133dd54e4e26","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2010,"title":"Seyone Chithrananda, Gabriel Grand, and Bharath Ram- sundar","work_id":"2e8f3af2-95cf-4889-9301-a4b279aea96c","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":37,"snapshot_sha256":"70e2cf5e55b29d5061326ae56c1fee57c56bb7b7a3d57b1a6edc41248f4fe48a","internal_anchors":5},"formal_canon":{"evidence_count":2,"snapshot_sha256":"1a35cc4079fbbc24f3620203bca31fee01a1d4f6c15b91130af121718aa28977"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}