Sign in Subscribe
5 min read

An AI agent for omics, end-to-end

Share
An AI agent for omics, end-to-end
Nº 01 · The Lede bioRxiv Computational biology

CARIBOU runs bioinformatics end-to-end

CARIBOU runs bioinformatics end-to-end
Fig. IbioRxiv · Filed 29 May 2026.

CARIBOU autonomously runs omics analyses end-to-end, chaining tool selection, parameter setting, and result interpretation across bulk RNA-seq, scRNA-seq, and variant-calling workflows in a single agent loop. The system pairs a reasoning LLM with a curated skill library of callable bioinformatics routines, the same pattern that has been working for software-engineering agents. Reported runs cover full pipelines from raw FASTQ through differential expression without human handholding between steps.

Read the source →

Graph LLM tackles drug-synergy generalization
Fig. IIarXiv · Filed 29 May 2026.
Nº 02 arXiv Drug discovery · Computational

Graph LLM tackles drug-synergy generalization

OOD-GraphLLM predicts drug synergy on cell lines and combinations the model never saw in training, fusing molecular graphs with an LLM backbone to push out-of-distribution accuracy past graph-only baselines. Generalization to unseen contexts is the failure mode that has kept synergy predictors out of real triage — this anchors a new reference point for what graph-LLM hybrids can claim on the hardest split.

Read more →
Bayesian LoRA for microbiome diagnosis
Fig. IIIarXiv · Filed 29 May 2026.
Nº 03 arXiv Computational biology

Bayesian LoRA for microbiome diagnosis

iLoRA adapts foundation models to microbiome-based disease classification using Bayesian low-rank adaptation (LoRA — a lightweight fine-tuning method that updates a small slice of weights) plus latent interaction graphs over taxa. The combination tracks uncertainty alongside predictions, narrowing the gap between the microbiome foundation models we've been tracking and clinical-grade diagnostics where calibrated confidence is non-negotiable.

Read more →
Also Filed · Three Briefs from the queue
Nº 04 bioRxiv Field report

DNA foundation model ranks CRC variants

A DNA foundation model prioritizes promoter regulatory variants in colorectal cancer directly from sequence, skipping the hand-engineered features that have dominated non-coding variant scoring. Raises the floor for what sequence-only models can do on regulatory regions, the hardest part of the cancer-variant interpretation stack.

Read →
Nº 05 Anthropic Agents · Infrastructure

Social scientists adopt coding agents

Anthropic surveyed 1,260 social scientists on AI and coding-agent use, finding broad uptake for data cleaning, analysis scripting, and literature review. Adjacent to the bioinformatics-agent push in #1 — the same agent patterns are quietly becoming default infrastructure across data-heavy research fields, not just CS-adjacent ones.

Read →
Nº 06 OpenAI Field report

OpenAI model disproves geometry conjecture

An OpenAI model disproved an 80-year-old central conjecture in discrete geometry, cracking the unit distance problem and producing a counterexample mathematicians had missed. First-of-kind result for AI-generated original mathematics on a long-standing open problem — resets what counts as a credible claim when models say they've found something new.

Read →

Reply with your discoveries. A human reads them. Forward freely.

Agentic Discovery  ·  Nº 26  ·  29 May 2026

Editor's Note

Today: a bioinformatics agent that runs your pipeline, two graph-LLM swings at hard generalization, and an OpenAI model that just cracked an 80-year-old conjecture.

 

Nº 01 · The Lede  —  bioRxiv  —  Computational biology

CARIBOU runs bioinformatics end-to-end

CARIBOU runs bioinformatics end-to-end

Fig. I  bioRxiv · Filed 29 May 2026.

CARIBOU autonomously runs omics analyses end-to-end, chaining tool selection, parameter setting, and result interpretation across bulk RNA-seq, scRNA-seq, and variant-calling workflows in a single agent loop. The system pairs a reasoning LLM with a curated skill library of callable bioinformatics routines, the same pattern that has been working for software-engineering agents. Reported runs cover full pipelines from raw FASTQ through differential expression without human handholding between steps.

Read the source →

Why it matters

First general-purpose omics agent demonstrated end-to-end on the workflows that consume the bulk of bench-adjacent compute budgets — moves autonomous bioinformatics from narrow per-tool demos to a deployable reference point every platform now has to match.

 

Nº 02  —  arXiv  —  Drug discovery · Computational

Graph LLM tackles drug-synergy generalization

Fig. II  arXiv · Filed 29 May 2026.

Graph LLM tackles drug-synergy generalization

OOD-GraphLLM predicts drug synergy on cell lines and combinations the model never saw in training, fusing molecular graphs with an LLM backbone to push out-of-distribution accuracy past graph-only baselines. Generalization to unseen contexts is the failure mode that has kept synergy predictors out of real triage — this anchors a new reference point for what graph-LLM hybrids can claim on the hardest split.

Read more →

 

Nº 03  —  arXiv  —  Computational biology

Bayesian LoRA for microbiome diagnosis

Fig. III  arXiv · Filed 29 May 2026.

Bayesian LoRA for microbiome diagnosis

iLoRA adapts foundation models to microbiome-based disease classification using Bayesian low-rank adaptation (LoRA — a lightweight fine-tuning method that updates a small slice of weights) plus latent interaction graphs over taxa. The combination tracks uncertainty alongside predictions, narrowing the gap between the microbiome foundation models we've been tracking and clinical-grade diagnostics where calibrated confidence is non-negotiable.

Read more →

 

Also Filed  ·  Three Briefs from the queue

Nº 04  —  bioRxiv  —  Field report

DNA foundation model ranks CRC variants

A DNA foundation model prioritizes promoter regulatory variants in colorectal cancer directly from sequence, skipping the hand-engineered features that have dominated non-coding variant scoring. Raises the floor for what sequence-only models can do on regulatory regions, the hardest part of the cancer-variant interpretation stack.

Read →

Nº 05  —  Anthropic  —  Agents · Infrastructure

Social scientists adopt coding agents

Anthropic surveyed 1,260 social scientists on AI and coding-agent use, finding broad uptake for data cleaning, analysis scripting, and literature review. Adjacent to the bioinformatics-agent push in #1 — the same agent patterns are quietly becoming default infrastructure across data-heavy research fields, not just CS-adjacent ones.

Read →

Nº 06  —  OpenAI  —  Field report

OpenAI model disproves geometry conjecture

An OpenAI model disproved an 80-year-old central conjecture in discrete geometry, cracking the unit distance problem and producing a counterexample mathematicians had missed. First-of-kind result for AI-generated original mathematics on a long-standing open problem — resets what counts as a credible claim when models say they've found something new.

Read →

 

· · ·

Reply with your discoveries. A human reads them. Forward freely.

Published by:

ARC