Virtual Agentic Lab! (Slides)

# Virtual Agentic Lab!

<div style="margin-top: 1.2em; text-align: center; opacity: 0.9;">
Peng Chen, Aavash Shakya<br>
Mentor: Mohsen Hariri
</div>

<div style="margin-top: 2.1em; opacity: 0.75; font-size: 0.85em; text-align: center;">
SCIPE Workshop on Large Language Models • Final Presentaion • January 18, 2026
</div>

---

.reveal .c-grid { display: grid; gap: 1.2rem; }
  .reveal .c-grid-2 { grid-template-columns: 1fr 1fr; }
  .reveal .c-grid-3 { grid-template-columns: 1fr 1fr 1fr; }
  .reveal .c-card {
    background: var(--card-bg);
    border: 1px solid var(--card-border);
    border-radius: 18px;
    padding: 1.15rem 1.25rem;
    box-shadow: 0 16px 34px rgba(15,23,42,0.07);
  }
  .reveal .c-kicker {
    font-size: 0.55em;
    letter-spacing: 0.14em;
    text-transform: uppercase;
    opacity: 0.7;
    margin-bottom: 0.4rem;
  }
  .reveal .c-muted { color: var(--muted); }
  .reveal .c-titleRow { display: flex; align-items: center; gap: 0.6rem; margin-bottom: 0.65rem; }
  .reveal .c-titleRow strong { font-weight: 700; }

.reveal .c-pill {
    display: inline-flex;
    align-items: center;
    gap: 0.4rem;
    padding: 0.25rem 0.55rem;
    border-radius: 999px;
    font-size: 0.55em;
    font-weight: 700;
    border: 1px solid rgba(15,23,42,0.14);
    background: rgba(255,255,255,0.75);
    letter-spacing: 0.04em;
    text-transform: uppercase;
  }
  .reveal .c-pill.agent { border-color: rgba(2,132,199,0.24); background: var(--agent-soft); color: rgba(2,132,199,1); }
  .reveal .c-pill.bio   { border-color: rgba(249,115,22,0.26); background: var(--bio-soft);   color: rgba(194,65,12,1); }
  .reveal .c-pill.ok    { border-color: rgba(16,185,129,0.26); background: var(--ok-soft);    color: rgba(5,150,105,1); }

.reveal .c-list { margin: 0.65rem 0 0 0; padding-left: 1.15rem; font-size: 0.82em; }
  .reveal .c-list li { margin: 0.28rem 0; }

.reveal .c-callout {
    border-left: 6px solid rgba(16,185,129,0.70);
    background: rgba(16,185,129,0.08);
    padding: 0.9rem 1.0rem;
    border-radius: 16px;
    font-size: 0.82em;
  }

.reveal .c-pre {
    margin: 0.5rem 0 0 0;
    padding: 0.85rem 0.9rem;
    border-radius: 16px;
    background: rgba(15,23,42,0.05);
    border: 1px solid rgba(15,23,42,0.10);
    font-size: 0.68em;
    overflow: hidden;
  }
  .reveal .c-pre code { white-space: pre; }

.reveal .c-metrics {
    display: grid;
    grid-template-columns: 1fr 1fr 1fr;
    gap: 1.0rem;
    margin-top: 1.0rem;
  }
  .reveal .c-metric {
    background: rgba(255,255,255,0.70);
    border: 1px solid rgba(15,23,42,0.12);
    border-radius: 16px;
    padding: 0.9rem 1.0rem;
    font-size: 0.78em;
  }
  .reveal .c-metric strong { font-weight: 750; }
  .reveal .c-foot { font-size: 0.55em; opacity: 0.75; margin-top: 0.85rem; }

@media (max-width: 1100px) {
    .reveal .c-grid-3 { grid-template-columns: 1fr; }
    .reveal .c-grid-2 { grid-template-columns: 1fr; }
    .reveal .c-metrics { grid-template-columns: 1fr; }
  }
</style>

## The headline result

<div class="c-callout">
An AI–human “Virtual Lab” orchestrates multiple LLM agents to design and experimentally validate <strong>92</strong> new SARS‑CoV‑2 nanobody candidates, including <strong>two</strong> with improved binding to recent variants.
</div>

<div class="c-grid c-grid-3" style="margin-top: 1.1rem;">
  <div class="c-card">
    <div class="c-kicker">Problem</div>
    <div class="c-titleRow">
      <span class="c-pill bio">Biology</span>
      <strong>Variants outpace binders</strong>
    </div>
    <div class="c-muted" style="font-size: 0.78em;">Need new binders fast as SARS‑CoV‑2 evolves (KP.3 / JN.1 era).</div>
  </div>
  <div class="c-card">
    <div class="c-kicker">Method</div>
    <div class="c-titleRow">
      <span class="c-pill agent">Agents</span>
      <strong>Meeting‑driven research</strong>
    </div>
    <div class="c-muted" style="font-size: 0.78em;">PI + specialist agents + critic + human guidance → decisions + code.</div>
  </div>
  <div class="c-card">
    <div class="c-kicker">Outcome</div>
    <div class="c-titleRow">
      <span class="c-pill ok">Validated</span>
      <strong>Pipeline + wet lab</strong>
    </div>
    <div class="c-muted" style="font-size: 0.78em;">ESM + AlphaFold‑Multimer + Rosetta → designs → expression + ELISA.</div>
  </div>
</div>

---

## Agentic Virtual Lab

---

## Virtual Lab architecture (agents + meetings)

<pre class="c-pre" style="margin-top: 1.0rem;"><code class="language-mermaid">%%{init: { 'theme': 'default', 'themeVariables': { 'fontSize': '18px' }, 'flowchart': { 'nodeSpacing': 45, 'rankSpacing': 55 }, 'config': { 'scale': 1 } }}%%
flowchart TD
  subgraph S0[Setup]
    H((Human researcher))
    PI[PI agent]
    C[Scientific Critic]
    S[Scientist agents]
    H -->|define roles| PI
    H -->|define roles| C
    PI -->|create| S
  end

subgraph M0[Meeting loop]
    A[Agenda] --> G[PI guide + questions]
    G --> R[Scientist responses]
    R --> K[Critic feedback]
    K --> X[PI synthesis + follow-ups]
    X -->|repeat N rounds| R
    X --> O[Final answer]
  end

H -->|writes| A
  S --> R
  C --> K
  PI --> G
  PI --> X</code></pre>

<div class="c-foot">Team meetings tackle broad decisions; individual meetings tackle focused tasks (e.g., writing code) with critique loops.</div>

---

## From idea → working workflow (5 phases)

<pre class="c-pre" style="margin-top: 1.0rem;"><code class="language-mermaid">%%{init: { 'theme': 'default', 'themeVariables': { 'fontSize': '20px' }, 'flowchart': { 'nodeSpacing': 55, 'rankSpacing': 60 }, 'config': { 'scale': 2 } }}%%
flowchart LR
  P1[1 Team selection] --> P2[2 Project specification] --> P3[3 Tool selection] --> P4[4 Tool implementation] --> P5[5 Workflow design]

style P1 fill:#e0f2fe,stroke:#0891b2,stroke-width:1px
  style P3 fill:#ffedd5,stroke:#ea580c,stroke-width:1px
  style P5 fill:#dcfce7,stroke:#16a34a,stroke-width:1px</code></pre>

<div class="c-metrics">
  <div class="c-metric"><strong>LLM used:</strong> GPT‑4o (gpt‑4o‑2024‑08‑06)</div>
  <div class="c-metric"><strong>Speed:</strong> ~5–10 min per meeting; ~1–2 h total</div>
  <div class="c-metric"><strong>Cost (reported):</strong> ~$10–$20 in GPT‑4o tokens</div>
</div>

---

## The scientific task: redesign nanobody binders

<div class="c-grid c-grid-2" style="margin-top: 1.1rem;">
  <div class="c-card">
    <div class="c-kicker">Goal</div>
    <div style="font-size: 0.92em; line-height: 1.35;">
      Start from known nanobodies that bind <strong>Wuhan</strong> SARS‑CoV‑2 RBD and mutate them to bind a newer variant (<strong>KP.3</strong>, closely related to <strong>JN.1</strong>).
    </div>
    <div class="c-foot">Starting nanobodies: Ty1, H11‑D4, Nb21, VHH‑72.</div>
  </div>
  <div class="c-card">
    <div class="c-kicker">Assay panel</div>
    <div class="c-muted" style="font-size: 0.82em;">Binding tested by ELISA against:</div>
    <ul class="c-list">
      <li>Wuhan RBD, BA.2 RBD</li>
      <li>JN.1 RBD, KP.3 RBD, KP.2.3 RBD</li>
      <li>Controls: BSA (and MERS‑CoV RBD in follow‑ups)</li>
    </ul>
  </div>
</div>

---

## The design pipeline (ESM → AlphaFold‑Multimer → Rosetta)

<pre class="c-pre" style="margin-top: 1.0rem;"><code class="language-mermaid">%%{init: { 'theme': 'default', 'themeVariables': { 'fontSize': '22px' }, 'flowchart': { 'nodeSpacing': 50, 'rankSpacing': 60 }, 'config': { 'scale': 2 } }}%%
flowchart LR
  WT[Starting nanobody] --> ESM[ESM LLR]
  ESM -->|top20| AF[AlphaFold Multimer ipLDDT]
  RBD[KP.3 RBD] --> AF
  AF --> RS[Rosetta dG]
  RS --> WS[Weighted score]
  WS --> OUT[Final selection 92]</code></pre>

---

## Scoring + iteration (the “optimization loop”)

$$
WS = 0.2\cdot\text{ESM LLR} + 0.5\cdot\text{AF ipLDDT} - 0.3\cdot\text{RS dG}
$$

<div class="c-grid c-grid-2" style="margin-top: 0.9rem;">
  <div class="c-card">
    <div class="c-kicker">Per round</div>
    <ul class="c-list">
      <li>ESM scores all single mutations → keep <strong>top 20</strong>.</li>
      <li>AlphaFold‑Multimer + Rosetta score binding → rank by <strong>WS</strong>.</li>
      <li>Select <strong>top 5</strong> → repeat (4 rounds; up to 4 mutations).</li>
    </ul>
    <div class="c-foot">Final selection uses a WS variant (WS<sub>WT</sub>) comparing to wild-type.</div>
  </div>
  <div class="c-card">
    <div class="c-kicker">Aggregate quality (92 mutants)</div>
    <ul class="c-list">
      <li><strong>100%</strong> had positive ESM LLR (mutant preferred over WT)</li>
      <li><strong>85%</strong> improved AF ipLDDT vs WT; <strong>35%</strong> had AF ipLDDT ≥ 80</li>
      <li><strong>65%</strong> improved RS dG vs WT; <strong>25%</strong> had RS dG ≤ −50</li>
    </ul>
  </div>
</div>

---

## Experimental validation highlights

<div class="c-metrics">
  <div class="c-metric"><strong>Expression:</strong> 38% (35/92) &gt; 25 mg/L; 6.5% (6/92) &lt; 5 mg/L</div>
  <div class="c-metric"><strong>Specificity:</strong> Most H11‑D4/Nb21 mutants retained Wuhan binding; some Ty1 mutants lost it</div>
  <div class="c-metric"><strong>Hits:</strong> 2 mutants gained moderate JN.1 binding (and one enriched KP.3 binding)</div>
</div>

<div class="c-grid c-grid-2" style="margin-top: 1.0rem;">
  <div class="c-card">
    <div class="c-kicker">Nb21 mutant</div>
    <div style="font-size: 0.84em; line-height: 1.35;">
      <strong>Nb21(I77V/L59E/Q87A/R37Q)</strong><br/>
      Gains ELISA binding to <strong>JN.1</strong> (EC<sub>50</sub> ≈ 2.0 ng/mL; Wuhan EC<sub>50</sub> ≈ 0.2 ng/mL) and shows enriched KP.3 binding signal.
    </div>
  </div>
  <div class="c-card">
    <div class="c-kicker">Ty1 mutant</div>
    <div style="font-size: 0.84em; line-height: 1.35;">
      <strong>Ty1(V32F/G59D/N54S/F32S)</strong><br/>
      Improves Wuhan binding and gains moderate ELISA binding to <strong>JN.1</strong>.
    </div>
  </div>
</div>

---

## Takeaways + limitations

<div class="c-grid c-grid-2" style="margin-top: 1.1rem;">
  <div class="c-card">
    <div class="c-kicker">Takeaways</div>
    <ul class="c-list">
      <li>Multi‑agent structure + critique can turn LLMs into an <strong>execution engine</strong> (not just Q&amp;A).</li>
      <li>Agents can stitch together domain tools into a usable pipeline, then hand off to wet lab validation.</li>
      <li>Framework is domain‑agnostic: swap agents, agendas, tools.</li>
    </ul>
  </div>
  <div class="c-card">
    <div class="c-kicker">Limitations (explicitly discussed)</div>
    <ul class="c-list">
      <li><strong>Knowledge cutoff</strong> → may miss newest tools/papers (mitigate with RAG / finetuning).</li>
      <li><strong>Prompt sensitivity</strong> → agenda iteration needed for specificity.</li>
      <li><strong>Hallucinations</strong> → human verification remains essential.</li>
    </ul>
    <div class="c-foot">Proposed next steps include tool sandboxes and better grounding/evaluation.</div>
  </div>
</div>

<div class="c-foot">Paper: https://doi.org/10.1038/s41586-025-09442-9</div>

Navigation