How I Built the Peppers

Step 1

Starting with the Transcriptome

Before engineering a pepper to be hotter, you need to understand how it makes heat in the first place. Capsaicinoids are synthesized in the placental tissue of the fruit — the white pith that holds the seeds. The biosynthesis genes activate in a precise developmental sequence, turning on and off across fruit maturation.

We profiled this using RNA-seq: extracted RNA from Capsicum chinense (habanero-type) placental tissue at five developmental time points — 4, 24, 34, 44, and 54 days post-anthesis (DPA), five biological replicates per time point, 38–42 million reads each. Salmon 2.8.0 quantified transcript abundance against the C. annuum UCD10Xv1.1 reference genome (GCF_002878395.1, 51,555 predicted proteins). Mapping rates: 80.8–84.0%.

PyDESeq2 0.5.4 ran the differential expression analysis — Wald test, Benjamini-Hochberg correction, α = 0.05 — across 28,074 expressed genes. Result: 4,365 genes significantly differentially expressed between peak capsaicinoid accumulation (34–44 DPA) and early development (4 DPA). Every gene in the capsaicinoid pathway was in that set, each with its own expression trajectory.

But knowing which genes are active doesn't tell you which one limits production. For that, you need to model the entire pathway as a system.

Step 2

Flux Balance Analysis

Capsaicinoid biosynthesis is a convergent pathway. Two branches feed into a final condensation step:

• The vanillylamine branch (from phenylalanine): PAL → C4H → 4CL → HCT → COMT → pAMT → vanillylamine
• The acyl-CoA branch (from valine/leucine): BCAT → KAS → FAR → branched-chain acyl-CoA
• Condensation: Pun1/AT3 joins vanillylamine + acyl-CoA → capsaicin

A constraint-based metabolic model of the full system was built in COBRApy 0.32.1: 37 metabolites, 44 reactions. Flux balance analysis (FBA) finds the steady-state flux distribution that maximizes capsaicinoid output subject to stoichiometric and thermodynamic constraints, then computes how much control each reaction exerts over that output — the flux control coefficient (FCC).

The Discovery

The Vanillylamine Finding

For decades, capsaicin synthase (Pun1/AT3) was assumed to be the rate-limiting step. It's the last enzyme in the pathway, the one that literally makes capsaicin. Papers had focused on it as the key engineering target for higher heat.

Pun1/AT3 has zero control over total pathway output. The real bottleneck is upstream — in the supply of vanillylamine, produced by the four-gene phenylpropanoid branch (PAL, C4H, 4CL, pAMT). This single finding changed the entire engineering strategy.

Instead of trying to make Pun1/AT3 run faster (which the model showed would accomplish nothing), the constructs overexpress the vanillylamine branch genes and eliminate the metabolic drain — the peroxidase enzymes that degrade capsaicinoids after they're made.

Step 3

AI-Guided Enzyme Design

Pun1/AT3 doesn't control flux, but it still performs the condensation. For higher SHU cultivars, we wanted variants that handle increased substrate volume without aggregating or losing activity.

ESM2 — Meta's 650M-parameter protein language model — scored every possible single-amino-acid substitution across the full 396-residue Pun1 sequence: 7,920 candidates total. Each was scored by log-likelihood ratio (LLR) relative to the wild-type. Higher LLR means the model predicts the mutation is more compatible with the BAHD acyltransferase fold. Candidates with LLR > 2.0 near the catalytic residues (H169, D297) were validated with AlphaFold2 structural prediction to confirm fold stability.

These three mutations are used together in the highest-tier products. S39L stabilizes the N-terminal domain. L345G opens the substrate channel. C175S reduces disulfide-mediated aggregation near the catalytic H169 residue. All three together form the Triple variant, used in the Heat Ladder SS-0087 through SS-0096.

Step 4

CRISPR Peroxidase Knockout

Capsaicinoids are actively degraded by class III peroxidase enzymes in the placental tissue. Without addressing this drain, any gains from boosting the vanillylamine branch get partially erased. Five CRISPR-SpCas9 guide RNAs were designed against the peroxidase loci (CaPOX1–CaPOX5), with filtering for GC content (40–70%), PAM site (NGG), and off-target homology against the full pepper genome.

Peroxidase knockout is the baseline modification in all 10 cultivar specifications — the first engineering layer before any supply-side changes.

Step 5

Codon Optimization & Construct Assembly

All 8 pathway genes were codon-optimized for C. chinense using a custom harmonization algorithm: replace each codon with the most-used synonymous codon for the target organism; scan for cryptic splice signals, premature polyadenylation, and homopolymeric runs; verify 5′ mRNA secondary structure (ΔG > −30 kcal/mol for ribosome loading). The optimized sequences achieve CAI 0.82–0.91 versus 0.45–0.55 for native sequences.

Final constructs are T-DNA binary vectors carrying 8 expression cassettes in this order: nptII (selection) → pAMT → Pun1/CS (Triple) → BCAT → KAS → FAR → PAL → C4H → 4CL. Total length: 13,345 bp per cultivar, synthesis-ready for direct ordering.

Beyond Peppers

Why It Works in Any Vegetable

The FBA finding had a second implication that only became clear when looking at the pathway architecture across plant families. The vanillylamine branch — the part that controls the Scoville ceiling — is made of enzymes that exist in every vascular plant on Earth. PAL, C4H, 4CL, HCT, COMT, BCAT, KAS, FAR: these feed lignin, flavonoids, and anthocyanins. No plant survives without them.

Only two enzymes are strictly pepper-specific:

• pAMT — converts vanillin to vanillylamine (the last step before condensation)
• Pun1/AT3 — condenses vanillylamine with the fatty acyl chain to make capsaicin

A tomato already has PAL at 78% protein identity, C4H at 82%, COMT at 74%, and all the rest. Add two genes, get a spicy tomato. The cross-species paper ran ortholog searches and FBA predictions for 17 plant species across 4 engineering tiers — from tomato (2 genes needed, >1M SHU predicted) to rice (full synthetic pathway). The ESM2-engineered Pun1 Triple variant transfers unchanged across all hosts — protein-level improvements are organism-independent.

The Result

Ten engineered Capsicum chinense cultivar specifications spanning roughly 0.9 million to nearly 3 million SHU. Twenty-five DNA constructs. Five CRISPR guide RNAs per cultivar. Every sequence synthesis-ready. A platform that can now extend the same pathway into 17 other species.

It started with RNA-seq. It ended with synthesis-ready DNA. The key that unlocked it was a metabolic model proving the bottleneck wasn't where anyone thought it was.

Read about the person who built it →