1. Transferable Mass Spectrum Prediction via Reference-Guided Test-time Specialization1
The unreviewed SPARC method adapts a mass-spectrum predictor using chemically related reference spectra while withholding the test compound’s spectrum. Tests across MassSpecGym, NPLIB1 and application-specific libraries report improved transfer between chemical and acquisition domains. This offers a way to specialize existing predictors for compound identification. Our account is limited to the original abstract.1
2. Correcting the Dropout-LayerNorm Expectation Gap Improves Protein Structure Models2
An unreviewed study identifies a systematic mismatch when dropout is followed by layer normalization and derives a small inference-time correction. Tests report accuracy improvements across nine protein-structure models and one protein–ligand docking model, with larger benefits in antibody-specific systems. The adjustment targets model evaluation, not experimental structure determination. Our account is limited to the original abstract.2
3. One Sequence, Many Decodings: CAGenMol-2 Recasts Drug Design as Masked Molecular Inference3
CAGenMol-2 represents molecules, properties and protein pockets within one masked sequence, allowing different design tasks through changes in what is hidden. Its unreviewed AdaFO search improves a reported pocket-design benchmark while largely preserving drug-likeness and diversity. These are computational design results, not measured binding or therapeutic activity. Our account is limited to the original abstract.3
4. Excited-State Photophysics Improves Scaffold-Robust Phototoxicity Prediction and Reveals Compound-Specific State Effects4
An unreviewed 450-compound study finds that adding excited-state descriptors improves phototoxicity prediction under repeated scaffold-grouped validation. Aggregating molecular microstates gives smaller gains, while extra conformers add little. The analysis helps distinguish useful physical information from representation complexity; it does not establish clinical photosafety. Our account is restricted to the repository-deposited abstract.4
5. De novo design of flexible protein interactions with GuideFlip5
GuideFlip jointly designs protein sequences and their evolving bound structures, targeting interactions too flexible for a fixed-backbone approach. The unreviewed study reports experimental binders to disordered alpha-synuclein and RBX1 regions, plus an active-state-selective receptor nanobody supported by cryo-EM. This extends binder design to conformationally flexible targets. Our account is limited to the repository abstract.5
References
Zhong, Yunhua; Li, Runting; Li, Yifan; Liu, Pan; Yang, Zhiwen; Wang, Zikun; Tang, Yixuan; Xia, Jun. Transferable Mass Spectrum Prediction via Reference-Guided Test-time Specialization. arXiv; 2026; Unreviewed preprint, version 1. Accessed 2026-09-29T16:12:31.974Z.
Source evidence and access
Abstract and version-1 submission history
Paraphrase of the inspected original abstract: The unreviewed SPARC method adapts a mass-spectrum predictor using chemically related reference spectra while withholding the test compound’s spectrum. Tests across MassSpecGym, NPLIB1 and application-specific libraries report improved transfer between chemical and acquisition domains. This offers a way to specialize existing predictors for compound identification.
Original arXiv abstract and submission history inspected; full manuscript not reviewed for this daily mention.
Ellmen, Isaac; Errington, David; Raybould, Matthew I. J.; Deane, Charlotte M. Correcting the Dropout-LayerNorm Expectation Gap Improves Protein Structure Models. arXiv; 2026; Unreviewed preprint, version 1. Accessed 2026-09-29T16:12:31.974Z.
Source evidence and access
Abstract and version-1 submission history
Paraphrase of the inspected original abstract: An unreviewed study identifies a systematic mismatch when dropout is followed by layer normalization and derives a small inference-time correction. Tests report accuracy improvements across nine protein-structure models and one protein–ligand docking model, with larger benefits in antibody-specific systems. The adjustment targets model evaluation, not experimental structure determination.
Original arXiv abstract and submission history inspected; full manuscript not reviewed for this daily mention.
Li, Yanting; Dai, Enyan; Wang, Lei; Ye, Wen-Cai; Liu, Li. One Sequence, Many Decodings: CAGenMol-2 Recasts Drug Design as Masked Molecular Inference. arXiv; 2026; Unreviewed preprint, version 1. Accessed 2026-09-29T16:12:31.974Z.
Source evidence and access
Abstract and version-1 submission history
Paraphrase of the inspected original abstract: CAGenMol-2 represents molecules, properties and protein pockets within one masked sequence, allowing different design tasks through changes in what is hidden. Its unreviewed AdaFO search improves a reported pocket-design benchmark while largely preserving drug-likeness and diversity. These are computational design results, not measured binding or therapeutic activity.
Original arXiv abstract and submission history inspected; full manuscript not reviewed for this daily mention.
Aris Tsai. Excited-State Photophysics Improves Scaffold-Robust Phototoxicity Prediction and Reveals Compound-Specific State Effects. ChemRxiv; 2026; Unreviewed preprint, version 1. DOI: 10.26434/chemrxiv.15009615/v1. Accessed 2026-09-29T16:12:31.974Z.
Source evidence and access
Original deposited abstract and posted/first-online metadata
Paraphrase of the inspected original abstract: An unreviewed 450-compound study finds that adding excited-state descriptors improves phototoxicity prediction under repeated scaffold-grouped validation. Aggregating molecular microstates gives smaller gains, while extra conformers add little. The analysis helps distinguish useful physical information from representation complexity; it does not establish clinical photosafety.
Restricted to original publisher/repository-deposited abstract and metadata in Crossref; landing page HTTP403 and full text not inspected.
Yi, K.; Chen, Q.; Zhang, D.; Tian, P.; Wagstaff, J. L.; McLaughlin, S. H.; Tate, C. G.; Jamali, K.; Scheres, S. H. W. De novo design of flexible protein interactions with GuideFlip. bioRxiv; 2026; Unreviewed preprint, version 1. DOI: 10.64898/2026.09.27.754145. Accessed 2026-09-29T16:12:31.974Z.
Source evidence and access
bioRxiv original abstract, experimental validation and version-1 release metadata
Paraphrase of the inspected original abstract: GuideFlip jointly designs protein sequences and their evolving bound structures, targeting interactions too flexible for a fixed-backbone approach. The unreviewed study reports experimental binders to disordered alpha-synuclein and RBX1 regions, plus an active-state-selective receptor nanobody supported by cryo-EM. This extends binder design to conformationally flexible targets.
Original bioRxiv API abstract and release metadata inspected; full manuscript not reviewed for this daily mention.
Publication record
Published 2026-09-29.
Sources, selection and claims were checked in an independent AI editorial review, followed by the AI editor's approval. This is not academic peer review.
- 2026-09-29 · Version 30deb4a0 · Viewing this version