Weekly AI/ML & Biotech Digest — Oct 19 to Oct 25, 2026

This week we witnessed the release of RFD4-Proteina, a flow-matching foundation model for biomolecular design — Baker Lab × NVIDIA GenAIR. There is no publication out yet. But their gitrepo commit suggests a preprint is coming soon. However, the supported hardware list is very short. You need to have a server grade Nvidia GPU or at least a DGX Spark to run it.

Other than RFD4-Proteina, there are also quite a few other protein design related models published. It seems research and publication is this area is accelerating (perhaps with the help of agentic research tools and interests from private sector?).

Another trend is the interpretation of reasoning tokens. This is particularly interesting since OpenAI reported that their 6.1 model showed deceptive behavior in the reasoning tokens. In other words, what the model wants you to believe it is doing/thinking, is not what it is actually doing. OpenAI decided not to release their model partly because of this deceptive behavior. Understanding model behavior has really important implications in monitoring reasoning models, security, and harnessing agentic tools. This is why Levy et al suggested the State over Token (SoT) conceptual framework. I remember suggested SoT in an earlier preprint. If you have read the preprint, you can find the final version published in Nature Machine Intelligence below.

Articles selected for the past week:

• ★ Protein Deimmunization with a Generative Diffusion Model — Protein immunogenicity is a major cause of biologic drug failure, leading to the production of anti-drug antibodies (ADA) that accelerate drug clearance, rejection of repeated doses, abolished efficacy, and sometimes life threatening anaphylaxis in patients.
• ★ DeepD enables efficient genome modeling and controllable long-context sequence generation across domains of life — Genome foundation models hold great potential for advancing biological research, but modeling diverse complex genomes and generating controllable genome-scale sequences remain challenging.
• ★ De novo design of metalloenzymes using Metal Mer — De novo enzyme design provides a route to create biocatalysts for target reactions, yet achieving high activity and selectivity directly from computation remains challenging.
• ★ Deep dynamical models of single-cell multiomic velocities predict loss-of-function and rescue perturbations in B cells — DynaVelo integrates single-cell omics data to infer cell state dynamics and regulatory interactions.
• ★ Characterizing the role of reasoning tokens in large language models as state over tokens
• ★ Towards understanding and monitorability of reasoning LLMs — New study examines how reasoning tokens in LLMs relate to internal computations, highlighting challenges in monitoring and explaining how models reach step-by-step conclusions.
• ★ Guidelines for evaluating structure-based methods for protein–ligand binding prediction — Nature Machine Intelligence publishes guidelines advocating transparency in data, code and experimental procedures to enable fair, reproducible comparisons across scoring methods.
• ★ Efficient equivariant framework for universal charge density prediction — Introduces EdenGNN-Uni, a universal charge density model over an order of magnitude faster than existing methods for efficient electronic structure calculations.
• ★ PaperBanana: Automating Academic Illustration for AI Scientists — Despite rapid advances in autonomous AI scientists powered by language models, generating publication-ready illustrations remains a labor-intensive bottleneck in the research workflow.
• ★ Proteome scale drug prioritization from AlphaFold 3 interaction confidence — AlphaFold 3 (AF3) predicts protein ligand complexes, but the utility of its confidence scores for drug prioritization remains unclear.
• ★ recAMPLIFY: Large-Scale Pretraining of Recursive Protein Language Models — Protein language models (pLMs) are powerful tools for protein annotation and mutation-effect prediction, but progress has largely depended on scaling model size and data, making them expensive and prone to memorization.
• ★ Reliable and Efficient Agentic Protein Design with Toolset Harnessing — Computational protein design workflows require the orchestration of multiple tools, data formats, and analysis steps that are often fragmented.
• ★ De novo design of enantioselective artificial metalloenzyme photocatalysts
• ★ Programmable design of functional proteins from natural language
• ★ Biology+AI Daily (@BiologyAIDaily) on X — Deep: Extracting latent energetic landscapes from protein structure predictors 1.
• ★ Reinforcement learning with evolutionary priors unlocks multi-site enzyme optimization from sparse experimental feedback — Industrial enzymes underpin sustainable manufacturing, yet natural proteins rarely combine the activity, robustness and application compatibility required for practical use.
• ★ Trion-1: Unifying Language, Protein Sequence, and Structure with Diffusion Language Models — We envision a general-purpose protein foundation model as a unified, interactive model of sequence, structure, and function, capable of understanding, generation, and cross-modal translation while retaining the instruction-following and reasoning capabilities of modern LLMs.