Update AZ-POSTMAN deck: reframe as engagement timeline

- Vision slide: spatial proteomics from any H&E, signal was always there
- Problem: missed opportunity framing, almost none have spatial proteomics
- Solution: 183 channels, 3.45M patches
- Comparison table: added ROSIE (134M patches) and GigaTIME (49K) numbers
- Timeline: reframed as "how we'd work together" not internal status
- Next steps: reframed as "how we work"
- Changed biomarkers → channels/protein markers throughout

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: yuedai0531

slides/AZ-postman.md

@@ -12,36 +12,41 @@ footer: "Strand AI  ·  Confidential"
 ![w:400](../assets/logo-white.svg)
 
 ## POSTMAN
-<span style="color:#D9D1BB">In Silico spatial proteomics from H&E</span>
+<span style="color:#D9D1BB">In silico spatial proteomics from H&E</span>
 
 ---
 
 <!-- _class: dark -->
 <!-- _paginate: false -->
 
-## What if every H&E slide already contained spatial proteomics?
+## What if you could get spatial proteomics from any H&E?
 
-Every clinical trial generates **thousands of H&E slides**. The spatial protein information is encoded in the tissue morphology — we just couldn't read it. Until now.
+At a fraction of the cost. No limit on scale.
 
-POSTMAN unlocks **183 biomarkers** from routine histology — no wet lab, no extra tissue, no waiting.
+Run spatial proteomics on **1,000 H&E slides** for the cost of staining one.
+
+POSTMAN unlocks **183 spatially resolved protein biomarkers** from routine histology.
+
+<small style="margin-top:auto;color:#D9D1BB">POSTMAN is the first model in our spatial biology platform.</small>
 
 <!--
 - The hook: the data is already there, hidden in morphology
 - Reframe H&E from "basic stain" to "information-rich source"
 - This is the vision before we explain the problem/solution
+- The "first model" line plants the seed that there's more coming
 -->
 
 ---
 
 <!-- _class: dark -->
 <!-- _paginate: false -->
 
-## The problem: H&E archives are under-leveraged
+## The missed opportunity on H&E
 
-- You have **thousands of H&E slides** — but most lack paired spatial proteomics
-- Running mIF assays is **expensive** ($5-10K/slide) and slow (days per round)
-- Many archived samples are **degraded or exhausted** — you can't run new assays on them
-- Without spatial biomarker data, you're leaving **patient stratification insights** on the table
+- You have **thousands of H&E slides** — but you can't run spatial proteomics on all of them
+- Running mIF assays is **expensive** ($5-10K/slide) and takes days per round
+- Many samples are **degraded or hard to access** — you can't run new assays
+- **Thousands of patients. Decades of trials. Spatial proteomics on almost none of them.**
 
 <!--
 - This is the pain point: H&E is cheap and abundant, proteomics is expensive and scarce
@@ -57,7 +62,7 @@ POSTMAN unlocks **183 biomarkers** from routine histology — no wet lab, no ext
 
 ```mermaid
 graph LR
-    A["H&E image"] --> B["POSTMAN"] --> C["Predicted mIF<br/>183 biomarkers"]
+    A["H&E image"] --> B["POSTMAN"] --> C["Predicted mIF<br/>183 channels"]
     style A fill:#F2F1ED,color:#00120A,stroke:#D9D1BB
     style B fill:#00120A,color:#F2F1ED,stroke:#00120A
     style C fill:#F2F1ED,color:#00120A,stroke:#D9D1BB
@@ -66,7 +71,7 @@ graph LR
 We trained on **7.5 TB of paired H&E + mIF data** — the largest dataset in the field.
 
 - **3.45M patches** across **18,573 regions**
-- Covers **183 unique biomarkers** — immune, structural, functional
+- Covers **183 unique protein markers** — immune, structural, functional
 - Pan-cancer, multi-site training data
 
 <!--
@@ -86,19 +91,19 @@ We trained on **7.5 TB of paired H&E + mIF data** — the largest dataset in the
 
 ### Training scale
 
-| Method | Patches | Biomarkers |
+| Method | Patches | Channels |
 |--------|---------|------------|
-| HEX (2026) | 755K | 40 |
-| ROSIE (2025) | — | 50 |
-| GigaTIME (2026) | — | 21 |
+| HEX (2025) | 755K | 40 |
+| ROSIE (2025) | 134M | 50 |
+| GigaTIME (2025) | 49K | 21 |
 | **POSTMAN** | **3.45M** | **183** |
 
 </div>
 <div style="flex:1">
 
 ### Architecture advantages
 
-- **VAE + Flow Matching** — generates realistic spatial distributions, not blurry averages
+- **State-of-the-art generative modeling** — produces realistic spatial distributions, not blurry averages
 - Predicts **spatially resolved** expression, not slide-level labels
 - Fine-tunable to your indications and target panels
 
@@ -114,74 +119,48 @@ We trained on **7.5 TB of paired H&E + mIF data** — the largest dataset in the
 
 ---
 
-<!-- _class: dark -->
 <!-- _paginate: false -->
 
-## The cost equation changes entirely
-
-**Today**: spatial proteomics costs **$5-10K/slide**, limits you to hundreds of patients.
-
-**With POSTMAN**: run the assay on a small subset, predict the rest.
-
-<div style="display:flex;gap:60px;margin-top:30px">
-<div style="flex:1;text-align:center">
-
-### Traditional approach
-100 slides × $7K = **$700K**
-(and you can't re-run degraded samples)
-
-</div>
-<div style="flex:1;text-align:center">
-
-### POSTMAN approach
-10 slides × $7K + inference = **$70K**
-(predict the other 90 from H&E)
+## How we'd work together
 
+<div style="position:relative;margin:40px 0 30px 0;padding:0 20px;height:20px;display:flex;align-items:center">
+<div style="position:absolute;left:20px;right:20px;height:4px;background:linear-gradient(to right,#004D3B 30%,#D9D1BB 30%);border-radius:2px"></div>
+<div style="position:absolute;left:20px;width:20px;height:20px;background:#004D3B;border-radius:50%;border:3px solid #fff"></div>
+<div style="position:absolute;left:30%;width:20px;height:20px;background:#D9D1BB;border-radius:50%;border:3px solid #fff"></div>
+<div style="position:absolute;left:42%;width:20px;height:20px;background:#D9D1BB;border-radius:50%;border:3px solid #fff"></div>
+<div style="position:absolute;right:20px;font-size:1.2em;color:#D9D1BB">→</div>
 </div>
+<div style="display:flex;justify-content:space-between;padding:0 20px;font-size:0.7em">
+<div style="text-align:left"><strong>Now</strong><br/>Align on validation criteria</div>
+<div style="text-align:center;margin-left:5%"><strong>March 1</strong><br/>We deliver model</div>
+<div style="text-align:center"><strong>You validate</strong><br/>Run on your cohort</div>
+<div style="text-align:right"><strong>Ongoing</strong><br/>Fine-tune & expand</div>
 </div>
 
-<!--
-- 10x cost reduction is conservative
-- Real value: you can analyze samples you physically can't re-stain
-- Enables retrospective analysis of trial archives
--->
-
----
-
-<!-- _paginate: false -->
-
-## Status and timeline
-
 <div style="display:flex;gap:60px;margin-top:20px">
 <div style="flex:1">
 
-### Current status
+### Validation plan
 
-- Model: **Training**
-- Dataset: 7.5 TB indexed and preprocessed
-- Early reconstructions showing strong PCC/SSIM
+- **Reconstruction metrics** — PCC, SSIM per marker
+- **Cox regression** — overall survival prediction
+- **Open to additional endpoints** you want validated
 
 </div>
 <div style="flex:1">
 
-### Validation plan
+### What's next
 
-- **Reconstruction metrics** — PCC, SSIM per biomarker
-- **Cox regression** — overall survival prediction
-- Open to **additional validations** AZ wants to see
+- Fine-tuning on your indications and panels
+- **In-silico perturbations** — responder prediction
 
 </div>
 </div>
 
-### Future work
-
-- Perturbation prediction — how would biomarker expression change under treatment?
-- Fine-tuning on AZ-specific indications and panels
-
 <!--
-- Training expected complete by mid-Feb
-- Happy to discuss what validation endpoints matter most to AZ
-- Perturbation is the longer-term vision: predict counterfactuals
+- Reframed as engagement timeline, not internal dev status
+- March 1st is a delivery commitment to them
+- "You validate" puts them in control
 -->
 
 ---
@@ -190,13 +169,12 @@ We trained on **7.5 TB of paired H&E + mIF data** — the largest dataset in the
 <!-- _footer: "" -->
 <!-- _paginate: false -->
 
-### Next steps
+### How we work
 
-1. **Define validation criteria** — which biomarkers, which indications, what correlation thresholds?
+1. **Align on success criteria** — which biomarkers, which indications, what validation thresholds
 
-2. **Share paired data** (optional) — fine-tuning on AZ data will outperform generic model
+2. **Scope the engagement** — define a pilot cohort, timelines, and collaboration terms
 
-3. **Pilot on your H&E archive** — pick a cohort, we run inference, you validate
 
 <div style="display:flex;align-items:center;justify-content:center;gap:30px;margin-top:auto">
 <img src="../assets/logo-pthalo.svg" width="220" />