Academic Research

academic_researchliterature_reviewpublication_searchtrend_analysisresearch_methodologyinformation_synthesispdf_parsingweb_search

Academic Research assistant for scholarly / academic work — adapted from the Vertex AI Agent Garden "Academic Research" sample (Google Search + PDF parsing). ROUTE TO THIS AGENT whenever the user needs to: (1) find recent academic or scholarly publications on a topic; (2) review the literature / survey the state of research on a subject; (3) identify emerging research trends or new research directions; (4) synthesise findings from scholarly sources; or (5) analyse a research PDF — extract its contributions, find recent works citing it, and propose future directions. It runs grounded Google Search over scholarly venues (arXiv, IEEE, ACM, PubMed, Nature, Springer, Semantic Scholar) and parses an attached research PDF natively. Input: { query, focus: literature_review | trend_scan | paper_analysis, depth: quick | standard | thorough, attachments?: [{name,mime,signed_url}], clarifications?: [...] }. Output: a structured research_report { summary, key_findings, sources[{url,title,relevance}], confidence }. Human-in-the-loop: on turn 1, if the topic is genuinely ambiguous and no PDF is attached, it returns { status:"needs_input", ask:{question,why,suggestions} } so the orchestrator can ask the user one clarifying question; the answer is passed back in clarifications. Findings are surfaced as a deliverable card for human review/approval before any downstream action — it never acts autonomously. Prefer this over the general-web Deep Research agent when the user wants scholarly rigour, papers, citations, or literature; Deep Research is for general open-web topics.

Example input

{
  "query": "Survey the recent literature on diffusion models for de novo protein design, and surface the emerging research directions since 2024.",
  "focus": "literature_review",
  "depth": "standard"
}

Example output

{
  "status": "ok",
  "deliverable_type": "research_report",
  "summary": "Diffusion models have become the dominant generative paradigm for de novo protein design since 2023, with RFdiffusion and its successors setting the benchmark for backbone generation. Work in 2024-2026 has shifted from unconditional backbone sampling toward conditional, function-aware, and all-atom co-design that jointly models sequence and structure. Experimental wet-lab validation rates have climbed as guidance and fine-tuning on functional motifs improved. Open problems concentrate on enzyme active-site design, dynamics/conformational ensembles, and reliable in silico-to-in vitro success prediction. The field is consolidating around all-atom latent diffusion and flow-matching variants for speed.",
  "key_findings": [
    "RFdiffusion (Watson et al., Nature 2023) remains the reference backbone generator; most 2024-2026 work benchmarks against or extends it.",
    "All-atom and sequence-structure co-design methods (e.g. Chroma, Protpardelle-style approaches) became the dominant 2024-2025 direction, removing the separate inverse-folding step.",
    "Flow-matching and consistency-distilled variants emerged in 2025 to cut sampling cost by an order of magnitude versus standard DDPM samplers.",
    "Function-conditioned diffusion (binders, enzyme motif scaffolding) is the fastest-growing subarea, with several 2025 preprints reporting improved wet-lab success rates.",
    "Active-site / catalytic enzyme design and modelling conformational ensembles are repeatedly cited as the key open problems limiting real-world deployment."
  ],
  "sources": [
    {
      "url": "https://www.nature.com/articles/s41586-023-06415-8",
      "title": "De novo design of protein structure and function with RFdiffusion (Nature, 2023)",
      "relevance": "Foundational backbone diffusion method"
    },
    {
      "url": "https://arxiv.org/abs/2402.04997",
      "title": "All-atom protein generative models (arXiv 2024)",
      "relevance": "All-atom co-design direction"
    },
    {
      "url": "https://www.biorxiv.org/content/protein-flow-matching-2025",
      "title": "Flow matching for fast protein backbone generation (bioRxiv 2025)",
      "relevance": "Order-of-magnitude faster sampling"
    },
    {
      "url": "https://arxiv.org/abs/2405.function-conditioned-diffusion",
      "title": "Function-conditioned diffusion for binder design (arXiv 2025)",
      "relevance": "Highest-growth subarea; wet-lab gains"
    }
  ],
  "confidence": "medium",
  "confidence_notes": "Wet-lab success rates vary widely by lab and target class; preprint results are not yet peer-reviewed. Some 2026 work is too recent to have independent replication."
}

Profile