1. Scalable molecular representations enabled by multimodal fusion and sequence distillation1
Chemical Dice Integrator combines six molecular representations, then distils them into an embedding generated from molecular strings. The accepted paper reports useful prediction across scaffold-based and limited-data evaluations, followed by yeast assays for prioritised compounds. This abstract-only account highlights a route to cheaper representations; the experimental endpoint is a yeast DNA-damage response, without evidence here of human therapeutic benefit.1
2. Procedural Pretraining for Molecular Property Prediction2
Training on abstract tasks before molecular data can improve subsequent property prediction, according to this unreviewed preprint. A sequence-reversal task helped a molecular-language model on lipophilicity and a small quantum-property dataset, suggesting an additional way to learn when chemical labels are scarce. Benefits depended on the task and training duration; hydration-free-energy results were inconclusive rather than uniformly improved.2
3. Molecular Geometry Understanding Has Unintendedly Emerged in Frontier Large Language Models3
Can a language model rank molecular conformations from atomic coordinates? An unreviewed benchmark tests 810 geometries across 27 molecules against quantum-chemical energy rankings. Some recent models approach a modern force-field baseline, with performance varying by molecule and interaction type. The result probes structural reasoning relevant to molecular design, while leaving broader chemical-space generalisation and experimentally useful design unestablished.3
4. AnewDDE: An Agentic Drug Discovery Engine for Biomolecular Interaction Modelling and Closed-Loop Design4
Anew Labs describes an engine linking structure prediction, affinity estimation and design to experimental feedback. In its unreviewed technical report, a nanobody campaign produced measured binders and improved affinity through a second design round. The demonstration gives concrete evidence for integrating computational tools with assays, but uses an undisclosed target and leaves functional activity and further developability evaluation outstanding.4
5. Exploring Optimal Parameters for Ligand-Based Virtual Screening in Early Drug Discovery5
Fingerprint settings can change which compounds a similarity search retrieves, even at the same search depth. This unreviewed study compares four reference ligands and blinded chemist assessments, finding that useful similarity thresholds and fingerprint choices depend on the ligand. It supports calibrated screening rather than a universal cutoff; the expert ratings and computational accessibility scores are not synthesis or activity measurements.5
References
Suvendu Kumar; Saveena Solanki; Mudit Gupta; Sonam Chauhan; Sanjay Kumar Mohanty; Shiva Satija; Abhinav Kumar Sharma; Subhadeep Duari; Arushi Sharma; Raidhani Shome; Vishakha Gautam; Sakshi Arora; Syed Yasser Ali; Adnan Raza; Sourav Sinha; Aayushi Mittal; Debarka Sengupta; Natarajan Arul Murugan; Gaurav Ahuja. Scalable molecular representations enabled by multimodal fusion and sequence distillation. Nature Communications; 2026; Peer-reviewed accepted manuscript, early online publication; not yet final Version of Record. DOI: 10.1038/s41467-026-77700-z. Accessed 2026-09-19T16:15:00.442Z.
Source evidence and access
Publisher abstract, accepted-manuscript notice, author list and published date; Crossref verifies metadata.
Evidence paraphrase from publisher abstract: six views merge into a latent representation distilled into a sequence model. Scaffold/low-data tests and yeast damage-response assays are reported. Quote: "experimental validation of isoeugenol and eugenyl acetate in a yeast damage-response assay."
Publisher abstract and metadata inspected; full accepted manuscript and supplements not inspected for this daily mention.
Moritz Friedemann; Zachary Shinnick; Philip Torr; Bruno Andreis. Procedural Pretraining for Molecular Property Prediction. arXiv; 2026; Unreviewed preprint, version 1. DOI: 10.48550/arXiv.2609.17831. Accessed 2026-09-19T16:15:00.442Z.
Source evidence and access
arXiv v1 Table 1, sections 4.1, 4.2.2, 4.2.3 and 5; repository v1 submission history verified.
Evidence paraphrase: Table 1 tests procedural then PubChem250k pretraining; Reverse improves Lipophilicity and QM9 gap. Sections 4.2/5 examine low-data benefit and excessive training. Quote: "We do not observe a clear benefit on FreeSolv."
Open full HTML inspected for this daily mention; relevant results, methods and discussion checked. No experiments or code independently rerun.
Gregorii A. Semakin; Timofey V. Losev; Ilya V. Prolomov; Stepan N. Ostarkov; Igor V. Alabugin; Michael G. Medvedev. Molecular Geometry Understanding Has Unintendedly Emerged in Frontier Large Language Models. arXiv; 2026; Unreviewed preprint, version 1. DOI: 10.48550/arXiv.2609.20666. Accessed 2026-09-19T16:15:00.442Z.
Source evidence and access
arXiv v1 benchmark design, Figure 1 and Table 1, Methods; full HTML; repository v1 submission history verified.
Evidence paraphrase: benchmark uses 25 GEOM-QM9 molecules plus two control molecules, thirty conformers each, common r2SCAN-3c reference rankings, and three prompt presentations. Table1 and per-molecule results compare force fields. Quote: "810 geometries in total".
Open full HTML inspected for this daily mention; relevant results, methods and discussion checked. No experiments or code independently rerun.
Anew Labs Team. AnewDDE: An Agentic Drug Discovery Engine for Biomolecular Interaction Modelling and Closed-Loop Design. Anew Labs technical report; 2026; Unreviewed company technical report; manuscript dated 16 September2026, publicly listed17September2026. Accessed 2026-09-19T16:15:00.442Z.
Source evidence and access
Company homepage Sep17 Tech Report entry; PDF pp1,18–23(section4, Figures10–12),36–37(Supplementary Methods A.3–A.4).
Evidence paraphrase: section4 reports seven of fifty initial VHH-Fc clones below400nM, followed by sixteen clones below10nM after100 additional designs. Target is masked. Quote: "These clones are ready for subsequent developability and functional evaluation."
Full52-page PDF downloaded from original company link; abstract, section4, Figures10–12 and Supplementary Methods A.3–A.4 inspected. No assay/data rerun.
Temitope Sobodu; Victor Chibuzor Johnson; Ryan Kern; Peter Oni. Exploring Optimal Parameters for Ligand-Based Virtual Screening in Early Drug Discovery. arXiv; 2026; Unreviewed preprint, version 1. DOI: 10.48550/arXiv.2609.16356. Accessed 2026-09-19T16:24:02.379Z.
Source evidence and access
arXiv v1 abstract and full HTML methods/results for matched-depth fingerprints and blinded chemist validation; repository v1 posting14September2026 21:22:51UTC.
Evidence paraphrase: four aminergic ligands queried against Enamine; matched retrieval depth comparisons of ECFP variants and feature-class fingerprints; 240 records representing229unique structures independently scored by3blinded chemists. No universal threshold; expert agreement moderate and SCScore association varied by ligand. No experimental synthesis or binding validation reported.
Open full HTML inspected for actual methods, results and limitations. No calculations or ratings independently reproduced.
Publication record
Published 2026-09-19.
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.
- 2026-09-19 · Version 22d6678d · Viewing this version