Bio papers — 2026-09-23

Using reinforcement learning to handle the massive amount of genomic data being generated is a key area of focus because manual analysis cannot keep up with the exponential growth of raw genomic information. Researchers are exploring how reinforcement learning algorithms can learn from experience with little human guidance, which is important because this significantly cuts down on the need for expensive labeled training data compared to traditional supervised learning methods.

One area of research involves applying these reinforcement learning techniques across different genomics research fields, such as gene regulatory networks, genome assembly, and sequence alignment. This overview brings together existing studies on using RL in these areas while also suggesting future directions like creating better reward functions and combining RL with other machine learning methods.

Another piece of research looks at decoding visual imagery from fMRI data by modifying the state-of-the-art perception decoder called DynaDiff to reconstruct imagined content. This is significant because it suggests that semantic structures learned from perception can stabilize visual imagery decoding even when the data is out of distribution. This approach uses a latent functional alignment technique to map activity into the model's semantic space, and they found this method consistently improved high-level semantic reconstruction metrics across four subjects.

Finally, there is work on creating pathology foundation models called WILSON, which represent whole-slide images as single composite images trained on nearly two hundred thousand slides from Mayo Clinic. This framework demonstrates that WILSON can outperform dedicated case-level models in performance and requires much less computing power than larger slide-level models while still improving histologic subtyping when fine-tuned on specific cancer cases.

Today's papers

• Revolutionizing Genomics with Reinforcement Learning Techniques. [paper]
• Seeing the imagined: latent functional alignment in visual imagery decoding from fMRI data. [paper]
• WILSON - a pathology foundation model framework for patient-level analysis and diagnostic text generation. [paper]

The papers

• Revolutionizing Genomics with Reinforcement Learning Techniques —
• Seeing the imagined: latent functional alignment in visual imagery decoding from fMRI data —
• WILSON - a pathology foundation model framework for patient-level analysis and diagnostic text generation —

Important terms

Reinforcement Learning (RL)

An AI technique where an agent learns to make decisions by interacting with an environment and receiving rewards, allowing it to learn from experience without needing extensive human-labeled data.

Genomic Data Analysis

The process of analyzing massive amounts of genetic information, which is growing exponentially. RL is being used here because manual analysis cannot keep up with the sheer volume of raw genomic data.

DynaDiff

A state-of-the-art perception decoder that has been modified to decode visual imagery from fMRI data. It uses latent functional alignment to reconstruct imagined content.

Pathology Foundation Models (WILSON)

Models like WILSON that represent whole-slide images as single composite images. They are trained on large datasets and can outperform case-level models while requiring less computing power.