Saturday, October 24, 2026

Preprints

• Sat, Oct 24AI drug discoveryProtein design

  Language model conditions antibody loop design on target sequences without structural data

  A masked language model called LangAAI designs antibody binding loops using only linear target and antibody sequences.

  Faisal Bin Ashraf, Stefano Lonardi

  University of California, Riverside

  bioRxiv
• Sat, Oct 24Foundation model

  Physics-informed neural networks calculate circadian clock sensitivity curves from standard light-pulse data

  A machine-learning tool called PINN-iPRC infers infinitesimal phase response curves for biological clocks by embedding the underlying oscillator equations directly into a neural network's loss function.

  Dongwoo Won, Gyuyoung Hwang, Jae Kyoung Kim

  KAIST; Institute for Basic Science

  bioRxiv
• Sat, Oct 24Foundation modelSingle cell biology

  Fine-tuning vision models to human brain cells uncovers Alzheimer's pathology links

  Adapting the Cellpose-SAM segmentation model on 5,832 annotated human prefrontal cortex cells boosted segmentation accuracy from an F1 score of 0.37 to 0.84, according to a benchmark of eight imaging methods.

  Liang You, Jamie L. Durbin, Guanqun Meng, and 6 othersLiang You, Jamie L. Durbin, Guanqun Meng, Wayne Stallaert, Guang Li, Shikhar Uttam, Hao Feng, Jill R. Glausier, Jiebiao Wang fewer authors

  University of Pittsburgh; University of Texas

  bioRxiv
• Sat, Oct 24Single cell biologyVirtual cell

  Confidence scores in histology-based virtual spatial transcriptomics fail to isolate model errors

  A systematic benchmark across melanoma, lung, and kidney tumors shows that uncertainty estimates generated by deep-learning models predicting gene expression from histology slides largely mirror predicted mean expression levels rather than true error.

  Thomas Batique, Kalin Nonchev, Viktor H Koelzer, Gunnar Ratsch

  ETH Zurich; University of Bern

  bioRxiv