Daily Paper Brief - 2026-06-26

Summary

• Top 5 papers selected
• Molecular / Protein ML: 3
• Pure-AI primitive: 2
• Notable signal: Unified multimodal molecular foundation model (Molexar) reaches 100% validity + strong drug-likeness; equivariant transformer backbone (MEET) achieves linear memory scaling for atomistic peptide design
• Public status: selected molecular/protein items were promoted to reading notes; pure-AI items remain in Curation queue

Papers

Molexar: A Unified Multimodal Molecular Foundation Model for Drug Design

• Track: Structure-based modeling
• Paper note: Molexar
• Source metadata: arXiv:2606.25865, submitted 2026-06-24
• Related concepts:
  • Molecular generation
  • SMILES
  • Fragment-SELFIES
• Priority: Follow-up
• Score: 5.7 / 10
• Follow-up:
  • Verify MolGenBench benchmark details
  • Check cross-docked2020 baseline comparisons
  • Reproducibility: code release to verify

Scalable Peptide Design via Memory-Efficient Equivariant Transformer (MEET)

• Track: Protein modeling
• Paper note: MEET
• Source metadata: arXiv:2606.25006, submitted 2026-06-23
• Related concepts:
  • Equivariant GNN
  • Protein design
• Priority: Follow-up
• Score: 5.5 / 10
• Follow-up:
  • Verify AFDB dataset scale and benchmark results
  • Check linear memory scaling claims with atom count
  • Reproducibility: preprint or code to verify

Multi-Scale Machine Learning for Antibody-Antigen Binding Affinity Prediction

• Track: Structure-based modeling
• Paper note: Multi-scale ML for Antibody-Antigen Binding
• Source metadata: bioRxiv preprint, posted 2026-06-23
• Related concepts:
  • Binding affinity
  • Protein representation
  • Boltzmann ceiling analysis
• Priority: Hold
• Score: 5.34 / 10
• Follow-up:
  • Verify LOCO-DMS cross-validation methodology
  • Check Boltzmann ceiling analysis details
  • Cross-pathogen transfer failure: follow up for generalization research

The Structural Origin of Attention Sink

• Track: AI
• Related concepts:
  • Attention
  • Transformer
• Priority: Hold
• Score: 4.54 / 10
• Follow-up:
  • Verify venue and publication status
  • Check super neuron analysis details

RelayCaching: Accelerating LLM Collaboration via Decoding KV Cache Reuse

• Track: AI systems
• Related concepts:
  • Inference
  • Context window
• Priority: Skip
• Score: 4.0 / 10
• Follow-up:
  • Verify 4.7x TTFT speedup claims
  • Check KV cache reuse rate benchmarks