Skip to content

SPECS.md

Latest commit

 

History

History
210 lines (163 loc) · 7.4 KB

SPECS.md

File metadata and controls

210 lines (163 loc) · 7.4 KB
Raw

Yes. This is the smart path. Fine-tuning public models is 10x cheaper than training from scratch and often works better.

Available public models you can leverage:

1. Protein Language Models (ESM, ProtBERT)

  • ESM-2: Meta's protein model (3B parameters, free)
  • ProtBERT: Pre-trained on UniProt
  • Use for: Predict protein properties from sequence → expression level prediction
  • Fine-tune cost: $200-1K (few-shot learning, 10-100 examples)
  • Your data: Wet lab partner gives you 50 protein variants + measured expression levels

2. DNA/RNA Transformers

  • Nucleotide Transformer: Pretrained on 100B+ DNA sequences
  • Caduceus: Recent, optimized for genomics
  • Use for: Predict expression from promoter/RBS/codon usage
  • Fine-tune cost: $100-500
  • Your data: Strain designs → observed titers

3. Time-series models (for fermentation curves)

  • Temporal Fusion Transformer (TFT): Pretrained on energy/weather
  • Neural ODE: General differential equation learner
  • Use for: Growth curves, substrate consumption over time
  • Fine-tune cost: $500-2K
  • Your data: Fermentation time-series data from partner labs

4. General LLMs (GPT, Claude, Llama)

  • Fine-tune for: "Given lab conditions in text, predict outcome"
  • Example: "E. coli, 37°C, pH 7.0, 200 rpm → predict OD600 at 24h"
  • Cost: API-based fine-tuning ($50-500) or local ($1-5K)
  • Advantage: Handles unstructured lab notes, natural language inputs

Realistic system architecture:

┌─────────────────────────────────┐
│   Wet Lab User Interface        │
│  (Simple form or notebook)      │
└────────────┬────────────────────┘
             │
      ┌──────▼─────────┐
      │  Your API      │
      │  - Validate    │
      │  - Parse input │
      └──────┬─────────┘
             │
    ┌────────▼──────────────────┐
    │  Ensemble Predictions     │
    │ ┌──────────────────────┐  │
    │ │ Protein ESM model    │  │ (fine-tuned on 
    │ │ (expression pred)    │  │  user's data)
    │ ├──────────────────────┤  │
    │ │ DNA Transformer      │  │
    │ │ (promoter strength)  │  │
    │ ├──────────────────────┤  │
    │ │ Time-series model    │  │
    │ │ (fermentation curve) │  │
    │ └──────────────────────┘  │
    └────────┬───────────────────┘
             │
    ┌────────▼──────────┐
    │ Retrain on wet    │
    │ lab validation    │
    │ (monthly loop)    │
    └───────────────────┘

Budget breakdown (per wet lab customer):

Task Cost Timeline
Initial setup: collect 50-100 historical runs $0 (their data) 2 weeks
Data cleaning + standardization $500-1K 1 week
Fine-tune 3 models (protein + DNA + time-series) $1-3K 1 week
Validation: 20 test experiments $5-20K (their wet lab) 4 weeks
Monthly retraining (continuous improvement) $200/month Ongoing
Total first year $10-25K 3 months to production

Compared to:

  • Building custom model from scratch: $50-100K
  • Consulting: $200-500/hour

Concrete example: Fine-tune ESM for your partner's strain

Step 1: Collect data

Lab gives you: 50 protein variants (different RBS, codons, tags)
             + measured expression levels in their bioreactor

Step 2: Fine-tune ESM-2

# Pseudocode
from huggingface import AutoModel

model = AutoModel.from_pretrained("facebook/esm2_t6_8M_UR50D")
# Already trained on 250M proteins; just 2 layers fine-tuning

# Fine-tune on 50 variants: $100-200 GPU cost
# Takes 1-2 hours on A100

# Output: model that predicts expression for THEIR strains

Step 3: Deploy as API

Input: protein sequence (paste from Benchling)
Output: predicted expression level + confidence interval

Step 4: Validate & improve

Lab tests 10 predicted high-expressers
6/10 validate → retrain with new data
Next month: accuracy improves from 70% → 80%

Go-to-market strategy:

Phase 1: Find 1-2 pilot customers (3 months)

  • Biotech startup doing protein engineering OR
  • Fermentation company optimizing yield
  • Give them free software (charge setup labor: $5-10K)
  • Goal: Prove 30-50% reduction in screening experiments

Phase 2: Productize (3 months)

  • Build simple web UI (not fancy, functional)
  • Integrate with common lab tools (Benchling, Google Sheets, Slack)
  • Package fine-tuning as service: "$15K to deploy custom model"

Phase 3: Scale (6+ months)

  • SaaS pricing: $500-2K/month per lab (usage-based)
  • OR licensing: $50K/year per organism type
  • Expand to: Yeast, mammalian cells, fermentation, etc.

What you need from wet lab partners:

  1. Historical data: 50-200 past experiments (genotype/conditions → outcome)

    • Format: Spreadsheet, JSON, or lab notebook
    • Quality: ±20% variation acceptable

  2. Validation experiments: 20-50 new runs to test predictions

    • Cost: They bear this ($5-20K) or you cover partly

  3. Feedback loop: Monthly/quarterly retraining

    • They share new results → you retrain model
    • Continuous improvement

  4. Real problem: Something that costs them $50K+/year to optimize

    • High-cell-density fermentation scaling
    • Protein expression optimization
    • Strain engineering screening
    • Media/substrate optimization

Realistic success metrics:

Year 1:

  • 2-3 pilot customers
  • Models achieving 70-80% prediction accuracy
  • 30-40% reduction in screening experiments per customer
  • Revenue: $20-50K (mostly setup/training)

Year 2:

  • 5-10 customers
  • Expand to 3-4 organism types
  • SaaS pipeline: $2-5K/month recurring
  • Revenue: $100-200K

Year 3:

  • 20+ customers
  • Recurring revenue: $10-20K/month
  • Potential acquisition target for biotech platform (Benchling, Genedata, etc.)

Critical success factors:

  1. Find the right first customer (must have real pain point + data)
  2. Deliver 40%+ iteration reduction (prove ROI in 6 months)
  3. Make integration seamless (don't add friction to their workflow)
  4. Iterate with them (their feedback → model improves)
  5. Know when to say "we can't predict this" (avoid false confidence)

Biggest risk:

Biology is variable. If you over-promise ("we predict 100% accuracy"), customers will abandon you on first failure. Better positioning: "We reduce your screening by 60%, but some biology is still unpredictable."

Your next step:

  1. Identify 3 target biotech companies (proteases, recombinant proteins, specialty chemicals, cell lines)
  2. Cold email: "I'm building AI to reduce your protein engineering cycles. Interested in 2-week pilot?"
  3. If yes: Get their historical data, fine-tune ESM-2 on it, validate with 10 experiments
  4. If 6/10+ validate: You have product-market fit

Do you have any potential pilot customers in mind, or connections to biotech? That's the real bottleneck.