Bio papers — 2026-09-15

The AbFlow framework introduces a one-step flow-matching method for designing full-atom antibodies end-to-end by using paratope restrictions to guide the process. This approach refines antibody structures, particularly the CDRH3 region, by using an extended velocity field network enhanced by an equivariant surface multi channel encoder that takes surface interaction data as input. This method produces superior antigen antibody complexes and markedly improves binding affinity compared to previous designs.

This success builds on earlier work that focused on optimizing binding affinity through different structural modeling techniques. Stable-Shift addresses the challenge of predicting how genes will respond to genetic changes in ways not seen before, which is important for reducing the need for expensive functional genomics experiments. It achieves this by aggregating single cell measurements into perturbation level expression shifts and fitting a low rank response basis using only training perturbations.

This prediction method achieved a cosine similarity of 0.592 on the K562 Perturb-seq benchmark, outperforming GEARS at 0.569. Furthermore, it showed better correlation with biological context inputs like STRING interactions and Gene Ontology annotations. This suggests that structuring the response prediction based on known biological networks helps capture unseen gene behavior more accurately than methods relying solely on raw data.

We also looked at how models can communicate complex perceptual experiences without changing their core meaning using the Perceptual Reality Transformer, which keeps source accounts alongside generated illustrations. Studies showed that while frozen neural representations can predict complex content well with an AUROC of 0.887 to 0.894, recovering subjective qualities like vividness and persistence is more difficult than predicting the content itself.

Finally, we explored how recurrent neuronal networks in the hippocampus achieve optimal goal-directed navigation through relaxation dynamics instead of explicit planning algorithms. This network learns transition probabilities between spatial locations where obstacles are represented by vanishing connection weights. When a goal is presented, the dynamics relax into a field equivalent to a Linearly solvable Markov Decision Process.

Today's papers

• AbFlow End-to-end Paratope-Centric Antibody Design by Interaction Enhanced Flow Matching This framework designs full-atom antibodies by focusing on antigen binding interfaces using flow matching. Stable-Shift Biologically Structured Prediction of Transcriptional Responses to Unseen Gene Perturbations This method predicts how a gene will respond to a change even if it was never tested before by using biological context. Perceptual Reality Transformer What Must an Illustration Preserve? This system helps communicate unusual perceptual experiences by keeping the original account alongside generated images. Random matrix theory of sparse neuronal networks with heterogeneous timescales This paper develops a mathematical theory describing how recurrent neural networks compute working memory using different timescales for inhibitory and excitatory units. URCHIN A Horizontal Spiking Language Model for Data-Constrained Pretraining This model creates a biologically plausible language model using spiking neurons that can be trained efficiently and deployed on edge devices. Chemical and geometric representation fidelity improves drug--target affinity prediction This work shows that keeping chemical and structural details intact during molecular representation is crucial for accurate drug binding predictions. Planning as Dynamics Relaxation Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation This study demonstrates that the dynamics of a hippocampal network naturally lead to optimal navigation by relaxing into a goal-directed activity field. From objective discovery to prediction of global ocean eco-provinces: A pathway for trustworthy learning This research uses machine learning to find and predict meaningful marine ecosystems in the ocean based on satellite data. [paper] [episode]
• The paper's title is AbFlow End-to-end Paratope-Centric Antibody Design by Interaction Enhanced Flow Matching and its core idea is this framework designs full-atom antibodies by focusing on antigen binding interfaces using flow matching.
• The paper's title is Stable-Shift Biologically Structured Prediction of Transcriptional Responses to Unseen Gene Perturbations and its core idea is this method predicts how a gene will respond to a change even if it was never tested before by using biological context.
• The paper's title is Perceptual Reality Transformer What Must an Illustration Preserve? and its core idea is this system helps communicate unusual perceptual experiences by keeping the original account alongside generated images.
• The paper's title is Random matrix theory of sparse neuronal networks with heterogeneous timescales and its core idea is this paper develops a mathematical theory describing how recurrent neural networks compute working memory using different timescales for inhibitory and excitatory units.
• The paper's title is URCHIN A Horizontal Spiking Language Model for Data-Constrained Pretraining and its core idea is this model creates a biologically plausible language model using spiking neurons that can be trained efficiently and deployed on edge devices.
• The paper's title is Chemical and geometric representation fidelity improves drug--target affinity prediction and its core idea is this work shows that keeping chemical and structural details intact during molecular representation is crucial for accurate drug binding predictions.
• The paper's title is Planning as Dynamics Relaxation Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation and its core idea is this study demonstrates that the dynamics of a hippocampal network naturally lead to optimal navigation by relaxing into a goal-directed activity field.

The papers

• AbFlow: End-to-end Paratope-Centric Antibody Design by Interaction Enhanced Flow Matching — AbFlow is a novel generative framework for "end-to-end full-atom antibody design" that addresses the limitations of existing models in modeling structural information flow and utilizing fine-grained antigen geometry. [episode]
• Stable-Shift: Biologically Structured Prediction of Transcriptional Responses to Unseen Gene Perturbations — This paper presents Stable-Shift, a "structured method for estimating unseen-gene responses" in the context of functional genomics. [episode]
• From objective discovery to prediction of global ocean eco-provinces: A pathway for trustworthy learning —
• Planning as Dynamics Relaxation: Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation —
• Chemical and geometric representation fidelity improves drug--target affinity prediction —
• Perceptual Reality Transformer: What Must an Illustration Preserve? —
• URCHIN: A Horizontal Spiking Language Model for Data-Constrained Pretraining —
• Random matrix theory of sparse neuronal networks with heterogeneous timescales —

Important terms

AbFlow

This framework is a one-step flow-matching method for designing complete antibodies from start to finish by using paratope restrictions to steer the process. It refines antibody structures, especially the CDRH3 region, leading to better antigen antibody complexes and stronger binding affinity.

Stable-Shift

This method predicts how genes will react to genetic changes in novel ways without needing costly functional genomics experiments. It does this by grouping single cell data into expression shifts and fitting a low-rank response basis.

Perceptual Reality Transformer

This model communicates complex experiences, like illustrations, while preserving the original meaning of the source accounts. It shows that predicting content is possible, but capturing subjective qualities like vividness is harder.

Recurrent neuronal networks

These networks in the hippocampus achieve goal-directed navigation by relaxing into specific dynamics rather than using explicit planning. They learn movement probabilities between locations where obstacles are represented by weak connections.