{"paper":{"title":"Rethinking Molecular OOD Generalization via Target-Aware Source Selection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A reinforcement learning policy selects source subsets to reduce extreme out-of-distribution errors in molecular property prediction by up to 11 percent.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Duanhua Cao, Jiajun Yu, Jiameng Chen, Kun Li, Wenbin Hu, Yizhen Zheng, Zhuohao Lin","submitted_at":"2026-05-13T16:09:46Z","abstract_excerpt":"Robust prediction of molecular properties under extreme out-of-distribution (OOD) scenarios is a pivotal bottleneck in AI-driven drug discovery. Current scaffold-splitting protocols fail to obstruct microscopic semantic overlap, predisposing models to shortcut learning and overestimating their true extrapolation capability; meanwhile, conventional domain adaptation paradigms suffer under extreme structural shifts, as blindly aligning heterogeneous source libraries injects topological noise and triggers negative transfer. To address these two challenges, scaffold-cluster out-of-distribution per"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Evaluations show that prediction errors of state-of-the-art 3D molecular models surge by up to 8.0x on SCOPE-BENCH with a mean of 5.9x, while POMA achieves up to an 11.2% reduction in mean absolute error with an average relative improvement of 6.2% across diverse backbone architectures.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The reinforcement-learning policy can reliably identify source subsets that avoid negative transfer under extreme structural shifts, and that cluster-level partitioning in physicochemical descriptor space fully eliminates microscopic semantic overlap between source and target.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SCOPE-BENCH shows state-of-the-art molecular models suffer up to 8x higher errors under extreme OOD, while POMA reduces mean absolute error by up to 11.2% via target-aware source selection and dual-scale adaptation.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A reinforcement learning policy selects source subsets to reduce extreme out-of-distribution errors in molecular property prediction by up to 11 percent.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c2e27df23a339679a1844568049e1ad8d2d3c0995809edf32b4ff555ae9aa2d0"},"source":{"id":"2605.13932","kind":"arxiv","version":1},"verdict":{"id":"0bfe1e92-65ce-48fa-b2f8-59d0aaacf456","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T05:04:32.992385Z","strongest_claim":"Evaluations show that prediction errors of state-of-the-art 3D molecular models surge by up to 8.0x on SCOPE-BENCH with a mean of 5.9x, while POMA achieves up to an 11.2% reduction in mean absolute error with an average relative improvement of 6.2% across diverse backbone architectures.","one_line_summary":"SCOPE-BENCH shows state-of-the-art molecular models suffer up to 8x higher errors under extreme OOD, while POMA reduces mean absolute error by up to 11.2% via target-aware source selection and dual-scale adaptation.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The reinforcement-learning policy can reliably identify source subsets that avoid negative transfer under extreme structural shifts, and that cluster-level partitioning in physicochemical descriptor space fully eliminates microscopic semantic overlap between source and target.","pith_extraction_headline":"A reinforcement learning policy selects source subsets to reduce extreme out-of-distribution errors in molecular property prediction by up to 11 percent."},"references":{"count":58,"sample":[{"doi":"","year":1907,"title":"Invariant Risk Minimization","work_id":"d76c6842-b84d-44ec-bcea-b80cd8d07981","ref_index":1,"cited_arxiv_id":"1907.02893","is_internal_anchor":true},{"doi":"","year":2007,"title":"k-means++: The advantages of careful seeding","work_id":"32919b45-82ef-4e70-91a0-0b7e30449a4f","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Gotennet: Rethinking efficient 3d equivariant graph neural networks","work_id":"273fa05e-a9e1-41fa-a1e6-ac6b53a9d56a","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2015,"title":"Why is tanimoto index an appropriate choice for fingerprint-based similarity calculations?Journal of cheminformatics, 7(1):20","work_id":"262f996a-8a63-47b1-8b7c-c507b17f40ad","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"E (n) equivariant topological neural networks.arXiv preprint arXiv:2405.15429, 2024","work_id":"222767a7-1bf4-4fbd-8e4d-01509a5de661","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":58,"snapshot_sha256":"d76bd538ba707300cd3a94232ec144511fedbea8b2c6afb31f2f9e26341b60ec","internal_anchors":5},"formal_canon":{"evidence_count":2,"snapshot_sha256":"f7d03053790ee27caad9f4fcab258cb581e9a872eaf1b32ca3858a54ec406552"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}