Protein-function prediction must handle a vast sequence space with limited experimental observations. Aghazadeh’s primer approaches this problem through Fourier representations, sparsity and epistasis: the effects of sequence changes can depend on their combinations. Sequence variation becomes more useful when interactions remain visible. This is a perspective on representing and sampling function, not evidence that every protein can be predicted from a small dataset.
Key statement · From the executive summary
Sequence variation becomes more useful when interactions remain visible.
Multimodal models of protein functionBoth examine protein function through representations of sequence and related information; they are distinct modelling approaches.
Protein design using deep learningFunction prediction can inform which designed candidates to examine; prediction and experimental validation remain separate.
For further reading and Research Journal access, contact [email protected].
Public background · External work
Protein design using deep learning
CALYR 80% · LITHIOS 20%
David Baker
Executive summary
Computational protein design connects proposed sequences with structural and functional objectives. Baker’s Broad/MIT colloquium presents this research direction through deep learning. Design creates candidates. Measurement tests their properties. The presentation is a public research reference; this summary does not establish independent experimental validation of any CALYR design.
Key statement · From the executive summary
Design creates candidates. Measurement tests their properties.
Protein function, sparsity and epistasisFunction prediction can inform which designed candidates to examine; prediction and experimental validation remain separate.
Multimodal models of protein functionStructure and function representations connect these topics; generating a candidate and predicting its properties are different tasks.
For further reading and Research Journal access, contact [email protected].
Public background · External work
Scientific agents for chemistry
CALYR 80% · LITHIOS 20%
Sam Cox; primer by Andrew White
Executive summary
Scientific agents extend language models by connecting them to specialist tools. The Cox and White session discusses literature work, chemistry and molecular-dynamics workflows, including ChemCrow and MDCrow. A scientific agent is only as useful as the tools and checks around it. A tool-using workflow still requires task-specific evaluation; successful execution alone does not establish a correct scientific conclusion.
Key statement · From the executive summary
A scientific agent is only as useful as the tools and checks around it.
Biomedical research agentsBoth discuss language-model agents using scientific tools; domain-specific tasks and evaluations remain distinct.
Biomni: public project referenceBoth connect language-model reasoning to specialist tools and execution; their tool environments and scientific scope differ.
Text2BIM: language-assisted building modelsShared workflow: language-model agents call domain tools or APIs and iteratively check outputs. Chemical and building-model validity do not transfer between domains.
For further reading and Research Journal access, contact [email protected].
Public background · External work
Multimodal models of protein function
CALYR 90% · LITHIOS 10%
Marinka Zitnik, Owen Queen and Yepeng Huang
Executive summary
Protein-function modelling can combine more than amino-acid sequences alone. The Zitnik, Queen and Huang presentation describes a multimodal approach connecting sequence and structure models with textual protein descriptions. Sequence, structure and descriptions offer complementary views of protein function. Such representations support research questions; a model response remains distinct from an independently measured property.
Key statement · From the executive summary
Sequence, structure and descriptions offer complementary views of protein function.
Protein function, sparsity and epistasisBoth examine protein function through representations of sequence and related information; they are distinct modelling approaches.
Protein design using deep learningStructure and function representations connect these topics; generating a candidate and predicting its properties are different tasks.
For further reading and Research Journal access, contact [email protected].
Public background · External work
Biomedical research agents
CALYR 90% · LITHIOS 10%
Kexin Huang; primer by Hanchen Wang
Executive summary
Biomedical research often spans literature, databases, analysis tools and experimental protocols. Huang’s Biomni presentation describes an agent architecture that combines planning, retrieval and executable workflows; Wang’s primer discusses approaches to biomedical discovery agents. Connect tools into a workflow, and keep the evidence traceable. Reported demonstrations belong to the cited work and are not independent CALYR benchmark results.
Key statement · From the executive summary
Connect tools into a workflow, and keep the evidence traceable.
Biomni: public project referenceThe talk and repository describe the same Biomni project. They provide complementary context, not independent validation.
Scientific agents for chemistryBoth discuss language-model agents using scientific tools; domain-specific tasks and evaluations remain distinct.
For further reading and Research Journal access, contact [email protected].
Public background · External work
Biomni: public project reference
CALYR 80% · LITHIOS 20%
Stanford SNAP / Biomni contributors
Executive summary
The public Biomni project provides a concrete reference for connecting biomedical reasoning with retrieval, planning and code execution. It offers a way to examine how a research question becomes a sequence of tool-assisted operations. Executable research needs sources, tools and a reproducible record. The public repository documents the project; this entry does not claim that CALYR has reproduced its benchmarks or validated it for a specific experimental task.
Key statement · From the executive summary
Executable research needs sources, tools and a reproducible record.
Biomedical research agentsThe talk and repository describe the same Biomni project. They provide complementary context, not independent validation.
Scientific agents for chemistryBoth connect language-model reasoning to specialist tools and execution; their tool environments and scientific scope differ.
For further reading and Research Journal access, contact [email protected].
Public background · External work
Text2BIM: language-assisted building models
CALYR 30% · LITHIOS 70%
Changyu Du, Sebastian Esser, Stavros Nousias and André Borrmann
Executive summary
Text2BIM studies how language-model agents can generate editable building models through specialist software. It connects an expressed design intention to model construction and iterative checking. An editable model is a starting point for checking a design. The reported research demonstrations do not establish general structural safety, building-physics validity or regulatory compliance.
Key statement · From the executive summary
An editable model is a starting point for checking a design.
Scientific agents for chemistryShared workflow: language-model agents call domain tools or APIs and iteratively check outputs. Chemical and building-model validity do not transfer between domains.
For further reading and Research Journal access, contact [email protected].