LCC Deep Mining — Multi-Signal Integration for PETase Engineering

🧬 PETase Competition · April 2026

LCC Deep Mining
Multi-Signal Integration

Goal: Engineer LCC variants with higher PET hydrolysis activity than wild-type and the ICCG benchmark,
by systematically mining a deep mutational scanning (DMS) dataset with 5 computational tools.

Target Enzyme

LCC — Leaf-branch Compost Cutinase (PDB 4EB0)

Benchmark

ICCG: F243I/D238C/Y127G/N246D (Tournier, Nature 2020, Tm +19°C)

DMS Dataset

8,179 variants · micro-droplet FACS · 40h PET hydrolysis

Fitness Definition

log₂(enrichment) after FACS sorting; >0 = better than WT

01

LCC Structure, Activity & Published Variants

Background

LCC (4EB0) — Structure & Activity

α/β-Hydrolase Fold

261 residues, chain A. Catalytic triad: Ser165 (nucleophile), Asp210 (acid), His242 (base). Oxyanion hole: Gly167, Asn246. All are DMS coldspots.

LCC WT vs ICCG — Literature Activity

Property	LCC WT	ICCG
Tm	84.7°C	90.9°C (+6.2°C)
PET degradation	31% Pf-PET, 3d, 65°C	90% in 10h, 72°C
Rate (Gf-PET)	93.2 mg_TA/h/mg	Higher at 72°C
Productivity	—	16.7 g TPA/L/h

Source: Tournier et al., Nature 2020. ICCG = F243I/D238C/S283C/Y127G. All residue numbers in this report use PDB/UniProt numbering.

ICCG Mutations in DMS

Mutation	DMS Fitness	Note
F243I	−0.130	Slightly deleterious alone
D238C	not measured	D238G=+0.42
S283C	+0.803	Disulfide partner — beneficial alone
Y127G	not measured	Y127H=+0.25, Y127C=−1.57
N246D	−1.027	Strongly deleterious alone

Published LCC Variants — Beyond ICCG

Key Variants (by improvement over ICCG) + DMS Singles Sum†

Variant	Additional Muts	vs ICCG	DMS Singles Sum†	Coverage	Source
LCC-LANL ★	+P38L/Y61C/M91I/L117P/A183V/H218Y/Q224H/S247L/T256I	14.3×	+1.26	11/13 scored (D238C, Y127G absent from DMS)	NREL 2024
RITK	+D53R/R143I/D193T/E208K	8.3×	+0.90	3/8 scored	Fang 2023
LCC-I40M	+6 mutations (many outside DMS range)	3.6×	+0.67	partial	ML study
ICCG/H252Y	+H218Y	2.6×	+1.16	3/5 scored	Cribari
LCC-A2	+H218Y/N248D	+40%	+1.43	4/6 scored	Zheng 2024
ICCG	(baseline)	1.0×	+0.67	2/4 scored	Tournier 2020

† Interpretation and limitations of DMS Singles Sum: This column is a naive linear sum of single-site DMS fitness values for scored mutations only — it does not predict true multi-mutant activity. Three caveats: (1) Missing mutations (e.g. D238C, Y127G absent from DMS singles) contribute 0 to the sum, causing underestimation; (2) Epistasis is entirely ignored — DMS directly measures F243I+P38L combinatorial fitness = −1.52, far below the additive sum of the two singles, demonstrating strong antagonism that this column cannot capture; (3) Coverage varies widely across variants (3/8 to 11/13), making cross-row comparisons unreliable. This column is a rough proxy for "how many DMS-beneficial singles does this variant carry" and must not be used for activity ranking.
★ LCC-LANL (14.3× ICCG, NREL/LANL ACS Catal. 2024). H218Y appears in 3 of the top-4 variants (single-mutant DMS fitness +0.490).

        Project context: Goal: identify LCC multi-mutant candidates that can outperform ICCG and potentially approach LCC-LANL (14.3× ICCG) performance, guided by computational mining of the 8,179-variant DMS dataset.
      

01b

LCC Structural Zone Definitions — Binding Site & Secondary Shell

Structure

Binding Site — 14 Residues (Tournier 2020)

Definition: Residues with direct contact to PET/MHET substrate, identified by molecular docking of 2-HE(MHET)₃ into PDB 4EB0 (Tournier et al., Nature 2020). Highest priority zone.

PDB	AA	Role	DMS fitness†	Scorecons	Published variant
165	Ser	Catalytic nucleophile	−0.685 (n=6)	0.948	—
210	Asp	Catalytic acid	−0.467 (n=5)	1.000	—
242	His	Catalytic base	−0.802 (n=6)	1.000	—
95	Tyr	Oxyanion hole (2nd)	−0.231 (n=5)	0.883	—
166	Met	Oxyanion hole (1st)	−0.719 (n=6)	1.000	—
164	His	Subsite S2	−0.138 (n=6)	1.000	—
125	Phe	Hydrophobic groove	−0.542 (n=5)	0.625	M125I (LCC-LANL)
127	Tyr	Aromatic clamp	−0.680 (n=4)	0.610	Y127G (ICCG)
130	Ser	Substrate binding	−0.232 (n=5)	0.984	T130M (WCCM)
94	Gly	Binding pocket	−0.317 (n=4)	1.000	—
190	Trp	Aromatic binding	−0.782 (n=5)	1.000	—
212	Val	Binding pocket	+0.111 (n=4)	0.836	—
243	Phe	Binding pocket	−0.319 (n=6)	0.748	F243I (ICCG/LCC-LANL)
246	Asn	Binding pocket	−1.271 (n=6)	0.985	N246D (ICCG)

†DMS fitness = mean log₂-enrichment across all measured substitutions at this position. Scorecons 0–1 (1=fully conserved); catalytic triad values from ganon scorecons_conservation.csv.

Secondary Shell — 42 Residues (≤5Å, all-atom)

Definition: Any residue not in the binding site whose closest atom is ≤5Å from any atom of a binding site or catalytic triad residue. Computed from PDB 4EB0 all-atom coordinates. Priority: binding site > secondary shell > surface/core.

PDB	AA	Dist to BS†	DMS fitness	Scorecons	Published variant
92	Ser	2.9Å	−0.557 (n=3)	0.910	—
93	Pro	1.3Å	−0.397 (n=5)	1.000	—
96	Thr	1.3Å	−0.709 (n=3)	1.000	—
97	Ala	3.0Å	−0.283 (n=5)	0.725	—
101	Ser	3.0Å	−0.301 (n=6)	0.884	—
102	Leu	4.0Å	−0.492 (n=4)	0.697	—
104	Trp	3.7Å	+0.395 (n=4)	1.000	—
123	Ser	3.0Å	+0.123 (n=4)	0.697	—
124	Arg	1.3Å	−0.398 (n=5)	0.514	—
126	Asp	1.3Å	−1.275 (n=5)	1.000	—
128	Pro	1.3Å	−0.694 (n=5)	0.984	—
129	Asp	1.3Å	+0.030 (n=5)	0.914	—
131	Arg	1.3Å	−0.646 (n=7)	0.978	—
132	Ala	3.2Å	−0.221 (n=4)	0.786	—
133	Ser	2.9Å	+0.163 (n=6)	0.567	—
134	Gln	3.0Å	−0.604 (n=5)	0.986	—
162	Ala	4.7Å	+0.149 (n=5)	0.630	—
163	Gly	1.3Å	−0.634 (n=6)	1.000	—
167	Gly	1.3Å	−1.097 (n=4)	1.000	—
168	Gly	2.5Å	−0.195 (n=4)	1.000	—
169	Gly	2.9Å	−0.657 (n=5)	1.000	—
170	Gly	2.9Å	−0.737 (n=5)	0.965	—
187	Leu	2.9Å	−0.404 (n=4)	1.000	—
188	Thr	2.9Å	−0.118 (n=4)	0.854	—
189	Pro	1.3Å	−0.243 (n=5)	0.766	—
191	His	1.3Å	−0.639 (n=3)	0.792	—
192	Thr	3.8Å	+0.826 (n=5)	0.763	—
207	Ala	3.0Å	+0.503 (n=5)	0.845	—
208	Glu	3.3Å	+0.101 (n=4)	0.897	—
209	Ala	1.3Å	+0.256 (n=5)	0.462	—
211	Thr	1.3Å	+0.297 (n=3)	0.770	—
213	Ala	1.3Å	+0.400 (n=3)	1.000	—
214	Pro	3.4Å	+0.071 (n=6)	0.845	—
215	Val	4.7Å	−0.112 (n=5)	1.000	—
218	His	3.3Å	+0.041 (n=7)	0.960	H218Y (LCC-LANL)
222	Phe	4.1Å	+0.204 (n=5)	0.975	—
240	Ala	3.5Å	−0.327 (n=4)	0.986	—
241	Ser	1.3Å	−0.418 (n=3)	0.796	—
244	Ala	1.3Å	−0.078 (n=5)	0.747	—
245	Pro	1.3Å	+0.361 (n=5)	1.000	—
247	Ser	1.3Å	+0.213 (n=3)	0.582	S247L (LCC-LANL)
248	Asn	4.7Å	−0.181 (n=6)	0.550	—

†Min atom-to-atom distance to any binding site residue (incl. catalytic triad). DMS fitness = mean log₂-enrichment. Scorecons 0–1 (1=fully conserved). Yellow = our top-5 candidate. Blue = published variant mutation.

02

Published Variants vs DMS Dataset — Coverage & Alignment

Gap Analysis

No published variant combination exists in DMS

The DMS library was generated by random combinatorial mutagenesis, not targeted at specific known variants. None of the 7 published LCC variants has an exact combination match in the 8,179-variant dataset. This means we cannot directly validate any published variant's performance using DMS data.

DMS Coverage of Published Variant Mutations

Variant	Activity	Total Muts	Singles in DMS	Key Missing	Subset Match in DMS?
LCC-LANL ★	14.3× ICCG	13	11/13	D238C, Y127G	1 pair found: F243I+P38L combinatorial fitness = −1.52 (Directly measured DMS double-mutant; independent of Singles Sum — demonstrates strong antagonism between these two mutations in the LCC-LANL background)
RITK	8.3×	8	3/8	D238C, Y127G, D53R, R143I, D193T	None
LCC-A2	+40%	6	4/6	D238C, Y127G	None
ICCG/H252Y	2.6×	5	3/5	D238C, Y127G	None
ICCG	1.0× (baseline)	4	2/4	D238C, Y127G	None
WCCG	Tm+7°C	4	1/4	F243W, D238C, Y127G	None

Key Observation: D238C and Y127G Missing Everywhere

These two ICCG core mutations are absent from all DMS singles. D238 has D238G (+0.42), D238V (−0.90) but not D238C. Y127 has Y127H (+0.25), Y127N (−0.19) but not Y127G. This is a fundamental coverage gap — the DMS library simply didn't sample these specific amino acid substitutions.

LCC-LANL: F243I+P38L Antagonism

The only subset match found: LCC-LANL contains both F243I and P38L, and this pair exists in DMS as a double with fitness −1.52. Additive prediction: F243I(−0.13) + P38L(+0.42) = +0.29. Actual: −1.52. Δ = −1.81 — severe antagonism. Yet in LCC-LANL (with 11 other mutations), the combination works brilliantly. This proves that higher-order epistasis rescues pairwise antagonism.

Implication: DMS data alone cannot validate published variants, and pairwise fitness does not predict multi-mutant outcomes. The best published variant (LCC-LANL) relies on complex higher-order epistasis that is invisible to any pairwise or additive analysis.

02

DMS Fitness Distribution — Single Mutations

Data

30.4%

Beneficial (>+0.3)
379 mutations

38.4%

Neutral (−0.3 to +0.3)
478 mutations

31.2%

Deleterious (<−0.3)
389 mutations

Distribution Shape

The distribution is roughly symmetric around 0 (WT level) with a slight left skew. The high beneficial fraction (30%) is unusual for enzymes — most DMS datasets show <10% beneficial. This indicates LCC has extensive room for improvement through mutation, consistent with it being a natural enzyme not previously optimized for PET degradation.

03

DMS Dataset Overview — Multi-Site Variants

Data

What is Micro-Droplet DMS?

Each LCC variant is encapsulated in a water-in-oil micro-droplet with a PET substrate. After 40h of hydrolysis at 65°C, droplets are sorted by fluorescence (FACS) — brighter = more PET degraded = higher activity. Fitness = log₂(enrichment ratio) after sorting. A fitness of 0 = wild-type level; >0 = better than WT; <0 = worse than WT. The measurement integrates activity, stability, and expression into a single readout.

Single-Site Mutations (k=1)

379

Beneficial (>+0.3)
30.4%

478

Neutral (−0.3 to +0.3)
38.4%

389

Deleterious (<−0.3)
31.2%

Total: 1,246 single mutants covering 259 of 261 positions. 30% beneficial rate is unusually high — LCC is a mutationally tolerant enzyme.

Multi-Site Mutations (k=2..13)

k	Count	Mean Fitness	Note
2	2,701	−0.75	Most common; antagonistic epistasis dominates
3	1,902	−0.93	Further fitness decline on average
4	1,225	−1.10	Best combo: +4.00 (R47H+T82A+A209V+S241P)
5	636	−1.00	Best combo: +6.98 (top of entire dataset)
6-13	544	−1.321	Diminishing returns at higher k

Total: 8,179 variants. Mean fitness declines with k, but rare combinations massively outperform WT — the dataset contains hidden gems.

04

DMS Fitness Landscape — Top Performers

Data

Top-10 Single Mutants

#	Mutation	Fitness	Structural Zone	OHM Zone
1	N249H	+1.839	Surface	structural_essential
2	N140D	+1.835	Surface	safe_target
3	A207E	+1.827	2nd Shell	allosteric_core
4	T192P	+1.808	2nd Shell	safe_target
5	L142M	+1.789	Surface	structural_essential
6	Q40P	+1.757	Surface	safe_target
7	N44H	+1.664	Surface	safe_target
8	N225K	+1.653	Surface	allosteric_handle
9	Y95F	+1.621	Binding site	structural_essential
10	Q217P	+1.597	Surface	allosteric_handle

OHM Zone legend:

allosteric_core = high ACI + conserved, directly relays catalytic signal;

allosteric_handle = high ACI + low conservation, tunable modulator — ideal engineering target; safe_target = low ACI, mutations are additive with low epistasis risk;

structural_essential = conserved but low ACI, maintains fold integrity.

Best Existing Multi-Mutants from DMS

k	Mutations	Fitness
5	T121M+Y127N+A183V+F196S+A281T	+6.975
2	I204F+A207T	+3.911
3	K194R+I204N+N288I	+3.888
4	R47H+T82A+A209V+S241P	+3.998
6	T60S+P93R+S100P+P189L+S258T+P280Q	+4.607

Key Observation

The best 5-mutant combination (+6.98) is 3.8× the best single mutant (+1.84). This is not merely additive — specific combinations synergize. The challenge: with 259 positions and 20 amino acids, the combinatorial space is vast. We need computational tools to navigate it efficiently.

Additive fitness = sum of individual single-mutant fitness values. If a combination's observed fitness exceeds the additive prediction, there is positive epistasis (synergy). Individual singles: T121M(−0.261)+Y127N(−0.191)+A183V(−0.084)+F196S(−1.272)+A281T(−0.660) = additive sum −2.47. The observed +6.975 represents extreme positive epistasis of +9.44 — the combination works despite each component being individually neutral or deleterious.

05

Conservation vs DMS Fitness — Identifying Engineering Targets

Sequence Analysis

Each dot = one position (mean fitness across all single mutations at that position). Hover for details. Ideal targets: top-left quadrant (low conservation + high fitness).

Reading the Plot (4 Quadrants)

✓ Top-Left: Ideal Targets

Low conservation + high fitness. Evolutionarily unconstrained AND mutationally tolerant. Q217P (cons=0.531, fit=+0.34) and N140D (cons=0.713, fit=+0.06) sit here — safest candidates to engineer.

⚠ Top-Right: High-Risk High-Reward

High conservation but positive fitness for specific substitutions. W104L (cons=1.0) and A207E (cons=0.845) — conserved positions where rare mutations improve function. Similar to ICCG's strategy. Higher epistasis risk.

✗ Bottom-Right: Avoid

High conservation + low fitness. Catalytic triad (Ser165, Asp210, His242) and structural core cluster here. Locked by evolution — almost all mutations are catastrophic.

Bottom-Left: Low Priority

Low conservation + low fitness. Not evolutionarily constrained, but mutations don't help either. These are surface-exposed or disordered positions with little functional relevance.

Spearman ρ = −0.247, NDCG@10% = 0.70 (p = 6.8×10⁻⁵). The negative correlation confirms that conservation predicts mutational intolerance, but the scatter is wide — many positions deviate from the trend, creating engineering opportunities.

06

Conservation, Hotspots & Coldspots

Sequence Analysis

Conservation Analysis (Scorecons, 150 Orthologs)

Method: BLASTp search against UniRef90 (E-value < 1e-30, query coverage > 80%) yielded ~150 LCC homologs. Multiple sequence alignment via Clustal Omega, then conservation scored by Scorecons (Valdar 2002). Score range 0→1, where 1.0 = identical across all orthologs.

Rationale: Conserved positions are under evolutionary constraint — mutations are more likely to be deleterious.

Conservation vs DMS fitness (per-position, single-mutation mean): Spearman ρ = −0.247 (p = 6.8×10⁻⁵, n=255 positions). As expected, more conserved positions have lower mean fitness across all substitutions.

NDCG@10% = 0.70 (computed on per-position mean single-mutation fitness, ranking 259 positions by negated conservation score) — conservation alone identifies ~70% of top-performing positions correctly.

Coldspots — Do Not Touch

Position	WT AA	Mean Fitness	% Deleterious	Role
165	Ser	−0.69	83%	Catalytic nucleophile
167	Gly	−1.10	100%	Oxyanion hole
210	Asp	−0.47	80%	Catalytic acid
242	His	−0.80	100%	Catalytic base
246	Asn	−1.27	100%	Oxyanion hole
96	Thr	−0.71	100%	Buried structural
170	Gly	−0.74	100%	Near active site

Hotspots — Mutationally Tolerant

Position	Mean Fitness	% Beneficial	Max	Note
217	+0.76	67%	+1.60	allosteric handle ★ candidate
207	+0.50	60%	+1.83	allosteric core ★ candidate
286	+0.44	60%	+1.33	structural_essential C-term loop, ACI 60.1% ★ candidate
117	+0.64	100%	+1.05	Surface exposed
44	+0.63	86%	+1.70	N-terminus
229	+0.56	100%	+0.84	Surface loop

Candidate Positions — Conservation Status

Position	Scorecons	Median	Safe?	Status
217 (Q→P)	0.531	0.826	Yes	Hotspot
140 (N→D)	0.713	0.826	Yes	Neutral
207 (A→E)	0.845	0.826	Borderline	Hotspot
104 (W→L)	1.000	0.826	Conserved	Neutral
286 (R→P)	0.990	0.826	Conserved	Hotspot

Apparent contradiction: conserved yet hotspot? "Hotspot" means that specific substitutions at this position are beneficial (e.g. R286P = +1.33), while "conserved" means most organisms keep the wild-type residue. This happens when one or two specific mutations escape the evolutionary constraint — e.g. Pro rigidifies a loop in a way evolution didn't explore. ICCG similarly mutated the highly conserved N246 (scorecons 0.985) successfully.

07

Deep Mining Strategy — 5 Computational Tools

Methods Overview

Rationale

DMS fitness alone ranks mutations by observed performance, but doesn't explain why they work or predict how they'll combine. We integrate 5 orthogonal zero-shot tools + DMS epistasis analysis — each capturing a different aspect of protein function — to identify positions with convergent multi-signal support, maximizing confidence for wet lab validation.

Tool 1

OHM Allostery

allosteric paths

Tool 2

RINpy Network

hub residues

Tool 3

ESM-2 PLM

evolutionary fit

Tool 4

MULTI-evolve

combo prediction

Step 5

Benchmark

16 models × 65K combos

OHM — Why?

Identifies positions that participate in allosteric signal transduction to the active site. Mutations at allosteric positions can modulate activity through long-range effects — a mechanism distinct from direct fitness.

RINpy — Why?

Builds a residue interaction network from atomic contacts in the PDB. Identifies structural hub residues. Mutations at hubs are usually catastrophic — exceptions are exceptionally valuable engineering targets.

ESM-2 — Why?

Protein language model trained on millions of sequences. Captures evolutionary constraints beyond simple conservation — understands amino acid context and co-evolution patterns.

MULTI-evolve — Why?

Arc Institute framework (Science 2026). The only tool that directly predicts multi-mutant fitness from single/double data using a neural network. Tests if top positions remain top in combination.

Scoring Benchmark — Why?

Exhaustive search of all 65,535 subsets of 16 tools via rank averaging. Best combo: ESM-2 3B + ThermoMPNN + OHM ACI (Spearman=0.248, NDCG@10%=0.833) — outperforms any single tool.

08

Tool 1: OHM Allosteric Communication Analysis

Allostery

What OHM Computes

OHM (Ohm-based Allosteric Model) analyzes how perturbations at one residue propagate through the protein to the active site.

Output: one ACI score per position (not per amino acid) — it is a property of the position in the structure, not of specific mutations. ACI is a percentile (0–100%) measuring how strongly that position participates in signal transduction to the catalytic triad.

Higher ACI = stronger allosteric coupling to the catalytic triad. This is not simply "better" — it depends on context: high-ACI positions in the Allosteric Handle zone (high ACI + low conservation) are the preferred engineering targets, as mutations there can tune catalysis with manageable epistasis risk. High-ACI positions in the Allosteric Core (conserved) are risky to mutate. Low-ACI positions (Safe Target) produce additive, predictable effects.

Zone Classification

OHM classifies each position into one of 4 zones based on ACI and conservation:

Zone	n	Meaning
Allosteric Core	39	High ACI + conserved → signal relay backbone
Allosteric Handle	23	High ACI + not conserved → tunable modulators
Safe Target	97	Low ACI → mutations are additive, low epistasis risk
Structural Essential	94	Conserved but low ACI → structural integrity

ACI vs DMS single-mutation fitness (per-mutation, n=1,228): Spearman=0.127, NDCG@10%=0.822. ACI has modest Spearman but high NDCG — it excels at identifying the top-10% beneficial positions, even if overall rank ordering is weaker.

Engineering Insights from OHM

Insight 1: Q217 is the ONLY position that is simultaneously an allosteric handle (ACI 84.9%) AND DMS-beneficial (+1.60). Handle positions are ideal engineering targets because they modulate the catalytic relay without being structurally essential.

Insight 2: A207 has the highest ACI (98.4%) among all beneficial mutations — it sits directly on the catalytic relay connecting Ser165 to Asp210. A207E can electrostatically tune the catalytic acid.

Insight 3: ICCG mutations F243I and N246D sit on Path 2 (substrate access channel). Their negative individual fitness is compensated by allosteric pathway optimization in combination — OHM explains why ICCG works despite deleterious singles.

Insight 4: Mutations on different allosteric paths (e.g. Path 1 + Path 3) have orthogonal effects on catalysis → predicted low antagonistic epistasis when combined.

09

OHM: Allosteric Pathways — Two Perspectives

Results

From Ser165 (Nucleophile) Outward

ACI radiates outward through a highly conserved, mutationally intolerant core:

Pos	AA	ACI%	Cons	Fitness	Insight
164	His	99.6	1.00	−0.18	DO NOT TOUCH — relay backbone
168	Gly	99.2	1.00	−0.13	DO NOT TOUCH — relay backbone
167	Gly	98.1	1.00	−0.67	Oxyanion hole backbone — ABSOLUTELY IMMUTABLE
171	Thr	96.9	0.84	+0.03	Neutral — tolerable but no gain
169	Gly	95.0	1.00	−0.42	Path 3 junction — Q217P relay feeds through here
166	Met	93.8	1.00	−0.47	Conserved relay residue
170	Gly	92.2	0.97	−0.47	Relay backbone
163	Gly	91.9	1.00	−0.46	Relay backbone

Insight: The Ser165 relay core is entirely locked by conservation + negative fitness. Engineering must approach from the periphery (Path 3: Q217P → ... → Ser165).

From Asp210 (Acid) Outward

Pos	AA	ACI%	Cons	Fitness	Insight
207	Ala	98.4	0.85	+0.23	★ A207E — HOTSPOT on relay!
209	Ala	96.1	0.46	+0.04	Low conservation — handle zone
213	Ala	94.6	1.00	+0.13	Synergistic (A213T+T230A Δ=+1.22)
212	Val	94.2	0.84	−0.02	Neutral
208	Glu	91.1	0.90	−0.05	Conserved, relay core
217	Gln	84.9	0.53	+0.34	★ Q217P — HANDLE + HOTSPOT

Insight: Position 207 is the only high-ACI position near Asp210 that is also a DMS hotspot. All other high-ACI neighbors are conserved + deleterious.

From His242 (Base) Outward

Pos	AA	ACI%	Cons	Fitness	Insight
241	Ser	98.8	0.80	−0.38	Handle — substrate entry
244	Ala	97.3	0.75	−0.10	Handle zone
240	Ala	96.5	0.99	−0.27	Core relay, conserved
245	Pro	95.7	1.00	+0.15	Only mutable residue on channel
243	Phe	93.4	0.75	−0.29	ICCG F243I sits here
246	Asn	93.0	0.99	−0.75	Oxyanion hole — ICCG N246D
238	Asp	90.3	0.55	−0.20	ICCG D238C disulfide partner

Insight: The His242 channel is tightly optimized. ICCG mutated two positions here (F243I, N246D) — both individually harmful but combinatorially rescued.

10

Interactive 3D Structure

Interactive

Left-drag = rotate · Right-drag = pan · Scroll = zoom · Click residue = info popup. Use sidebar tabs within each viewer to select specific paths.

11

Tool 2: RINpy — Residue Interaction Network Analysis

Network

What RINpy Computes

RINpy (Residue Interaction Network in Python) builds a graph where each residue is a node and edges connect residues within 4.5 Å (non-bonded atomic contacts in PDB 4EB0). It then computes betweenness centrality (BC) — the fraction of all shortest paths in the network that pass through each residue.

High-BC residues are structural "hubs" — removing or modifying them disrupts the most communication pathways in the protein. Result: 258 nodes, 1,314 edges.

BC vs DMS Fitness

BC vs DMS single-mutation fitness (per-mutation, n=1,068): Spearman=+0.027 (near zero — BC alone barely predicts fitness), NDCG@10%=0.771. Hub residues tend to have lower fitness tolerance, but the relationship is weak at per-mutation level. BC contributes mainly through its inclusion in the best doubles combo.

Mutation	BC Rank	Degree	Fitness	Category
N140D	#5/258	14	+1.84	Beneficial Hub
S136Y	#8/258	16	+1.43	Beneficial Hub
W104L	#104/258	9	+1.39	Moderate
R286P	#99/258	10	+1.33	Moderate
Q217P	#253/258	6	+1.60	Low BC (Safe)

Engineering Insights from RINpy

Insight 1: N140D is a major structural hub (BC rank #5/258). N140D sits on many shortest paths between residues — a true connector node with 14 direct contacts. Yet it is DMS-beneficial (+1.84), making it a rare "Beneficial Hub." This makes it a uniquely valuable engineering target: hub mutations usually disrupt function, but N140D defies that trend.

Insight 2: Q217P has near-zero betweenness centrality (BC rank #253/258). Q217P carries almost no network load — it sits at the periphery of the structural communication graph. This makes it exceptionally safe to mutate from a network perspective. Its allosteric effect (ACI 84.9%) operates through OHM's allosteric relay (Path 3), not through direct structural shortest paths.

Insight 3: High BC (N140D #5) + Low BC (Q217P #253) = complementary pairing. N140D anchors the structural network as a hub; Q217P operates through allosteric relay orthogonally. Combining a hub mutation with a peripheral allosteric handle is predicted to have lower antagonistic epistasis than two hub mutations. Source: rinpy_dms_merged.csv pdb_pos column (correct PDB node mapping).

Column Definitions

BC Rank: Betweenness Centrality rank out of 258 residues. #1 = most central hub (most shortest paths pass through it). Degree: Number of direct residue contacts within 4.5 Å — higher degree = more packed neighbors in the structure. Categories: Beneficial Hub = BC top-10% AND DMS fitness > +0.3 (rare: hub residues usually can't be mutated). Safe Peripheral = BC bottom-10%, very few structural contacts, safe to mutate without disrupting the fold. Moderate = BC in middle range, some structural role but not a critical hub.

12

Tool 3: ESM-2 Protein Language Model

PLM

What ESM-2 Computes

ESM-2 (650M parameters) is a transformer-based protein language model trained on ~250M protein sequences. We use masked marginal scoring: for each position, mask the residue, and compute log P(mutant|context) − log P(wildtype|context). A positive score means the PLM considers the mutation more "natural" in this sequence context.

Why Use a PLM?

Unlike simple conservation (MSA counting), ESM-2 captures context-dependent co-evolutionary patterns. It can detect that a mutation is acceptable in this specific protein even if the residue is conserved across the family — because the surrounding context compensates.

Calibration Against DMS

Spearman ρ = 0.242 · NDCG@10% = 0.817 (n=1,246 single mutations scored by ESM-2 3B)

ESM-2 is a weak predictor of micro-droplet fitness. This is expected: DMS fitness integrates activity + stability + expression, while ESM-2 primarily captures evolutionary plausibility. ESM-2 is one input signal, not a standalone predictor.

Engineering Insights from ESM-2

Insight 1: ESM-2 correctly identifies the catalytic triad as unmutable — Ser165, Asp210, His242 all have strongly negative ESM-2 scores for any substitution, consistent with DMS coldspot status.

Insight 2: ESM-2 flags Q217P as mildly positive (evolutionary context accepts Pro at this position), consistent with its low conservation (scorecons=0.531) and DMS hotspot status.

Insight 3: ESM-2 3B is the best single zero-shot predictor (Spearman=0.242, NDCG@10%=0.817), but combining it with ThermoMPNN and OHM ACI via rank averaging reaches Spearman=0.248, NDCG=0.833 (see benchmark slide). We also tested SaProt-650M (Sp=0.204, NDCG=0.809) and ProstT5 (Sp=0.179, NDCG=0.773) — both moderate, not in the best combination.

Updated assessment: With full 1,050-mutation evaluation, ESM-2 3B achieves Spearman=0.242, NDCG@10%=0.817. Combined with ThermoMPNN (stability) and OHM ACI (allostery), the rank-averaged score reaches Spearman=0.248, NDCG=0.833 — a meaningful improvement over any single tool.

12

Structural Criteria: Where Are the Top Mutations?

Structure

Top-10 Singles — Structural Location

Mutation	Fitness	SASA	Min Dist†	Structural Zone	—	Location
N249H	+1.84	Surface	6.6Å	Surface	—	Surface, near C-terminus
N140D	+1.84	Surface	12.3Å	Surface	—	β-sheet edge, RINpy BC major hub (#5/258)
A207E	+1.83	Buried	3.0Å	2nd Shell	—	Secondary shell, allosteric core (3.0Å to binding pocket)
T192P	+1.81	Surface	3.8Å	2nd Shell	—	Secondary shell, surface loop
L142M	+1.79	Buried	14.9Å	Core	—	Core, buried
Q40P	+1.76	Surface	26.0Å	Surface	—	N-terminus, flexible
N44H	+1.66	Surface	25.5Å	Surface	—	N-terminus, flexible
N225K	+1.65	Surface	12.3Å	Surface	—	Surface loop
Y95F	+1.62	Surface	—	Binding site	—	Oxyanion hole (Tournier 2020) — direct substrate contact
Q217P	+1.60	Surface	7.1Å	Surface	—	Allosteric handle, surface-exposed

Pattern: 7/10 surface/distal; 2/10 (A207E, L142M) buried; 1/10 (Y95F) in binding site; 2/10 (A207E, T192P) in secondary shell. Binding site: 14 residues with direct PET/MHET contact from Tournier 2020 molecular docking (PDB 94, 95, 125, 127, 130, 164, 165, 166, 190, 210, 212, 242, 243, 246). Secondary shell: any residue with closest atom ≤5Å to any binding site residue (42 residues total, computed from PDB 4EB0). †Min Dist = minimum atom-to-atom distance to any binding site residue; "—" = residue IS in binding site.

Binding Site & Secondary Shell in Published Variants

ICCG Y127G — IN BINDING SITE (aromatic clamp, direct substrate contact). ICCG/LCC-LANL F243I — IN BINDING SITE (binding pocket). LCC-LANL M125I — IN BINDING SITE (hydrophobic groove). WCCM T130M — IN BINDING SITE. Key insight: ICCG and LCC-LANL primarily engineer the binding pocket itself, not just the secondary shell.

Position 217 — Allosteric Hub

23 beneficial mutations within 10Å of position 217, forming a cluster in the 192–222 region. Key neighbors: T192P (+1.81), A213S (+1.10), F222C (+1.02), P214L (+0.75).
4 synergistic doubles involving Q217: Q217R+I252T (Δ=+2.45), G170D+Q217H (Δ=+2.17), Q217K+A250S (Δ=+1.96).
Position 217 is surface-exposed (46% rSASA), polar, 18Å from active site — ideal for engineering without disrupting the catalytic machinery.

Double Mutation Pattern by Structure

Allosteric+Allosteric pairs: worst epistasis (mean Δ=−1.63). Secondary shell+shell: also bad (−1.55). Best pairs: other+other (−0.50) and close_to_active+close_to_active (~39% positive Δ). Structural proximity to active site may increase synergy potential.

13

Tool 4: MULTI-evolve — Can We Predict Combinations?

ML

What MULTI-evolve Is

MULTI-evolve (Arc Institute, Science 2026) trains a FCNN on measured single + double mutant fitness to predict multi-mutant combinations. We have no experimental doubles for the top-20 singles, so we tried using zero-shot pseudo-doubles instead.

Our Attempt: Zero-Shot Pseudo-Doubles

Step 1: Use best doubles zero-shot combo (ESM-2 650M + BLOSUM62 + SaProt + OHM ACI, Sp=0.186 with real doubles) to score 190 pairwise doubles of top-20 singles.
Step 2: Normalize scores to DMS fitness scale using linear mapping learned from 2,467 real doubles (DMS_fitness = 0.0044 × zs_score − 3.87, R²=0.03).
Step 3: Train FCNN on 21 real singles + 190 normalized pseudo-doubles = 211 training points.

Validation on Real k=3 DMS (n=1,650)

Method	Sp	NDCG
ESM-2 3B additive (no training)	+0.177	0.703
ZS combo additive (no training)	+0.166	0.706
DMS additive (no training)	+0.105	0.667
FCNN w/ ZS pseudo-doubles	−0.011	0.646
FCNN w/ DMS additive pseudo	−0.023	0.620

Why FCNN Failed

One-hot features cannot generalize

The FCNN uses 5,180-dim one-hot features (259 positions × 20 AAs). It only sees the 20 positions in the top-20 singles during training. For k=3 variants involving any other position, the model has zero learned weights — it outputs random predictions.

Only 36 of 1,650 k=3 validation variants had even 1 mutation in the top-20. On those 36, the zero-shot FCNN achieves Sp=+0.215 — but this is too few to be reliable.

Alternative 1: 14-dim Features on All Doubles

Replace one-hot with 16 zero-shot model scores per mutation (sum for doubles). Train Ridge/GBR on 1,787 real doubles:

Method	k=3 Sp	k=3 NDCG
Ridge (16-dim, all doubles)	+0.129	0.681
ESM-2 3B additive (no training)	+0.177	0.703

Alternative 2: Train on Real Top-50 Doubles

Use top-50 DMS doubles (fitness 2.1–3.9) as training data. 154 samples: 87 singles + 66 doubles + WT. 16-dim features, FCNN (128-64):

Method	k=3 All (n=1650)	k=3 In-Training Positions (n=59)
FCNN 16-dim (top doubles)	Sp=+0.066	Sp=+0.132, NDCG=0.728
ESM-2 3B additive	Sp=+0.177	Sp=+0.241, NDCG=0.710

Key finding: Within the trained position set, 16-dim FCNN achieves the best NDCG (0.728) — it captures some epistasis signal from real doubles. But it doesn't generalize to unseen positions.

Conclusion: Training on real doubles with 16-dim features captures epistasis within the training positions (NDCG=0.728 > ESM-2 3B's 0.710). But no model generalizes to all positions. For a full combinatorial prediction, Phase 2 experimental doubles covering more positions remain essential.

14

Cross-Tool Consensus — Which Positions Do All Tools Agree On?

Consensus

Consensus Criteria (5 Independent Signals)

For each of 259 positions, we count how many tools independently flag it as "interesting": (1) DMS fitness in top 25% · (2) ACI above median · (3) BC above median · (4) Appears in top-100 multi-mutants · (5) Low conservation (scorecons < median). Only 13 of 259 positions (5%) have 4+ tools agreeing — these are the consensus positions.

Consensus Positions (4-5 / 5 tools agreeing)

PDB Pos	Structural Zone	n Tools	DMS Top25%	ACI > median	BC > median	In Top Multimuts	Low Conserv.	Candidate?
217	Surface	4/5	✓	✓	—	✓	✓	★ Yes (Q217P)
286	Surface	4/5	✓	✓	✓	✓	—	★ Yes (R286P)
104	2nd Shell	4/5	✓	✓	✓	✓	—	★ Yes (W104L)
140	Surface	4/5	✓	—	✓	✓	✓	★ Yes (N140D)
207	2nd Shell	3/5	✓	✓	—	✓	—	★ Yes (A207E)
197	Surface	5/5	✓	✓	✓	✓	✓	—
203	Surface	5/5	✓	✓	✓	✓	✓	—
192	2nd Shell	5/5	✓	✓	✓	✓	✓	—
193	Surface	4/5	✓	✓	—	✓	✓	—
136	Surface	5/5	✓	✓	✓	✓	✓	—

Q217P, R286P, W104L, N140D each reach 4/5 consensus. A207E reaches 3/5 (BC #132 below median, cons 0.845 above median). 4 of our 5 candidates meet ≥4/5 threshold. BC source: rinpy_dms_merged.csv pdb_pos column (correct mapping). Q217P BC #253/258 (below median — safe peripheral); N140D BC #5/258 (major hub). W104L and A207E are in 2nd Shell (≤5Å from binding site), not core. †R286P BC #99/258 (above median ✓).

15

Candidate Mutations Mapped onto LCC Structure

Structure

Orange = 5 candidate mutations. Red = catalytic triad. Grey = protein backbone. Drag to rotate, scroll to zoom.

Spatial Distribution

3D viewer labels use PDB 4EB0 residue numbers (= DMS position + 34).

Mutation	Viewer label	Location	Structural Role
Q217P	GLN217	Surface loop	Allosteric handle — remote from active site, modulates via Path 3
W104L	TRP104	Core helix	Allosteric core — buried, part of signal relay network
A207E	ALA207	Active site adjacent	On Path 1 relay — directly influences catalytic Asp210
R286P	ARG286	C-terminal loop	Structural essential — Pro rigidifies C-terminus
N140D	ASN140	β-sheet edge	RINpy BC #5/258 (major hub) — sits on many structural shortest paths; defies trend as a beneficial hub
Ser165	SER165	Active site	Catalytic nucleophile — DO NOT TOUCH
Asp210	ASP210	Active site	Catalytic acid — DO NOT TOUCH
His242	HIS242	Active site	Catalytic base — DO NOT TOUCH

Spatial logic: The 5 mutations are spread across the structure — not clustered in one region. Q217P is on the surface (~25 Å from active site), A207E is near the active site, N140D is on the β-sheet edge, W104L is in a core helix, R286P is at the C-terminus. This spatial distribution minimizes steric clash risk between mutations.

16

The Case for Q217P + W104L + A207E + R286P + N140D

Recommendation

Per-Mutation Multi-Signal Evidence

Mutation	Fitness	Deep Mining Rank†	ACI %	BC Centrality Rank (of 258 positions)	Scorecons	OHM Zone	Hotspot?	In Top Combos?	Consensus
Q217P	+1.60	#1	84.9%	#253	0.531	handle	Yes	✓	4/5
W104L	+1.39	#2	78.7%	#104	1.000	core	—	✓	4/5
A207E	+1.83	#3	98.4%	#132	0.845	core	Yes	✓	3/5
R286P	+1.33	#4	60.1%	#99	0.990	essential	Yes	✓	3/5†
N140D	+1.84	#5	39.9%	#5	0.713	hub	—	✓	4/5

Why This Specific Combination?

Orthogonal allosteric paths: A207E (Path 1) + Q217P (Path 3) → different relay networks, predicted low antagonism
Hub + Low-BC mix: N140D (hub #5/258) + Q217P (peripheral #253/258) → N140D anchors the structural network, Q217P is safe to engineer via allosteric relay without structural network disruption
All 5 appear in real top combos in DMS — not just theoretically combined
3 of 5 are DMS hotspots (Q217P, A207E, R286P) — positions where most mutations improve fitness
Epistasis +9.44 — additive prediction −2.47, observed +6.98; extreme positive epistasis drives this combination

Honest Risk Assessment

W104 conserved (1.0) and R286 conserved (0.99)higher risk of disrupting fold stability. Mitigation: ICCG also mutated conserved positions successfully.
No experimental double data for these specific pairs — epistasis is predicted, not measured.
Antagonistic epistasis is common (mean Δ = −0.70 in DMS) — even well-chosen combos may underperform additivity.

This is why we propose a phased approach with k=3 as a safer first test.

17

Experimental Recommendation — Phased Approach

Next Steps

Phase 1: Immediate — 5 Constructs (2 weeks)

Q217P+A207E+N140D (k=3) — Safest bet: 3 different allosteric paths, 2 hotspots, 1 hub
Q217P+W104L+A207E (k=3) — Pure allosteric: handle + core + core
Q217P+W104L+A207E+R286P+N140D (k=5) — Full top-1 candidate
N249H+N140D+A207E+T192P (k=4) — Top-4 by pure single fitness; N140D is a major structural hub (BC #5/258)
T121M+Y127N+A183V+F196S+A281T (k=5) — Positive control: this exact combination already measured at +6.98 in DMS, serves as experimental benchmark to validate assay conditions

Phase 2: Pairwise Doubles — 105 Constructs (4-6 weeks)

Select 15 top mutations → synthesize C(15,2)=105 pairwise doubles. Measure in micro-droplet or plate assay. Unlocks real epistasis data for MULTI-evolve retraining.

Phase 3: MULTI-evolve R2 — ~50 Constructs (2-4 weeks)

Retrain FCNN on 105 real doubles + 15 singles. Predict k=3..10 with epistasis-aware model. Expected: identify variants exceeding ICCG (Tm > 84°C, >90% PET in 10h).

Phase 1: 2 weeks → Phase 2: 4-6 weeks → Phase 3: 2-4 weeks · Total: ~3 months to optimized LCC variant

18

DMS Observation: The Epistasis Paradox in LCC

Critical Finding

What We Computed

For 2,467 double mutants where both singles are measured: Δ = f_observed(AB) − [f(A) + f(B)]. Positive Δ = synergy (better than expected). Negative Δ = antagonism (worse than expected). We then grouped pairs by whether each single is beneficial (>+0.3), neutral, or deleterious (<−0.3).

Epistasis by Pair Type — The Counter-Intuitive Result

Pair Type	n	Mean Δ	% Synergy	% Antagonism	Mean Observed
Beneficial + Beneficial	192	−1.724	12.5%	87.5%	−0.425
Beneficial + Neutral	606	−1.276	19.3%	80.7%	−0.610
Beneficial + Deleterious	424	−0.737	31.4%	68.6%	−0.826
Neutral + Neutral	429	−0.653	33.8%	66.2%	−0.635
Deleterious + Neutral	589	−0.245	45.5%	54.5%	−0.963
Deleterious + Deleterious	227	+0.496	63.4%	36.6%	−0.995

The Paradox

Combining two beneficial mutations is the worst strategy. 87.5% of beneficial+beneficial pairs show antagonism (mean Δ = −1.72). The average observed fitness of two beneficial mutations combined is −0.425 — worse than wild-type, despite an additive prediction of +1.30.

Conversely, two deleterious mutations combined synergize 63.4% of the time (mean Δ = +0.50). This explains why the best k=5 in DMS (T121M+Y127N+A183V+F196S+A281T = +6.98) uses 5 individually deleterious mutations.

What This Means for Our Strategy

Naively combining top-fitness singles is almost guaranteed to fail (87.5% antagonism rate)
The best existing DMS combo exploits compensatory epistasis among deleterious mutations — a strategy we cannot replicate computationally
No zero-shot tool (OHM ACI, RINpy BC, ESM-2, ThermoMPNN) can predict which pairs will synergize (all Sp ≈ 0 with Δ)
Phase 2 (measuring 105 pairwise doubles) is not optional — it's the only way to identify synergistic pairs for rational combination design

Honest conclusion: Any k≥3 candidate we propose based on additive singles fitness has an ~87% probability of antagonism per pair. We should present candidates with this caveat and prioritize Phase 2 experiments.

19

Zero-Shot Scoring Benchmark — 16 Models × 65,535 Combinations

Benchmark

Method

We tested 16 zero-shot scoring functions (no DMS data used for training), including the recently published FAMPNN (Full-Atom MPNN, ICML 2025). Each scores every single mutation independently. We then exhaustively searched all 2¹⁶−1 = 65,535 subsets, combining scores via rank averaging (convert each model's scores to ranks, then average). Evaluated by Spearman ρ and NDCG@10% against DMS single-mutation fitness (n=1,228 mutations with all 16 scores available).

Individual Models — Singles

#	Model	Type	Spearman	NDCG@10%
1	ESM-2 3B	PLM	+0.242	0.817
2	ESM-2 650M	PLM	+0.215	0.789
3	SaProt 650M	Structure-PLM	+0.204	0.809
4	ESM-1v	PLM	+0.191	0.789
5	ESM-2 150M	PLM	+0.186	0.764
6	ThermoMPNN	ddG/Structure	+0.179	0.803
7	MSA log-odds	Evolution	+0.162	0.807
8	ProstT5	Structure-PLM	+0.179	0.773
9	Conservation	MSA/Scorecons	+0.132	0.793
10	OHM ACI	Allostery	+0.127	0.822
11	MSA mut-freq	Evolution	+0.132	0.819
12	BLOSUM62	Substitution	+0.123	0.807
13	ProFAM	Autoregressive PLM	+0.154	0.816
14	FAMPNN	Full-Atom Design	+0.115	0.800
15	ProteinMPNN	Structure	+0.140	0.803
16	RINpy BC	Network	+0.027	0.771

Best Combination by Size (Rank Average, Spearman)

Size	Best Combination	Spearman	NDCG@10%
1	ESM-2 3B	+0.242	0.817
2	ESM-2 3B + OHM ACI	+0.248	0.811
3	ESM-2 3B + ThermoMPNN + OHM ACI	+0.248	0.833
4	+ ESM-2 650M	+0.254	0.808
5	+ SaProt + Conservation	+0.252	0.799
6	+ BLOSUM62	+0.249	0.801
...	Spearman decreases monotonically as more features are added
15	All 15 features	+0.213	0.793

Best Combination by NDCG@10% (top mutations)

Size	Best Combination	Spearman	NDCG@10%
3	ThermoMPNN + FAMPNN + OHM ACI	+0.200	0.823
5	ThermoMPNN + BLOSUM62 + FAMPNN + OHM ACI + RINpy	+0.200	0.826

Best Spearman (overall ranking): ESM-2 3B + ThermoMPNN + OHM ACI (Sp=0.248, NDCG=0.833) — unchanged after adding FAMPNN.

Best NDCG (finding top mutations): ThermoMPNN + BLOSUM62 + FAMPNN + OHM ACI + RINpy BC (Sp=0.200, NDCG=0.826) — FAMPNN improves top-mutation identification. Note: ESM-2 3B is NOT in the best NDCG combo — structure-based models dominate here.

Why these 3 (Spearman)? ESM-2 3B = evolutionary plausibility, ThermoMPNN = stability, OHM ACI = allostery. Three orthogonal signals.
Why FAMPNN helps NDCG? FAMPNN is a full-atom design model — it excels at identifying the best mutations (top-10%) but doesn't distinguish well among average ones.

65,535 subsets exhaustively searched — these are global optima, not cherry-picked.

20

Double-Mutation Benchmark — Additive Zero-Shot Prediction

Doubles

Method

For each double mutant A+B with both singles measured (n=1,787), predict fitness as score(A)+score(B). Same 16 zero-shot features (including FAMPNN), exhaustive subset search (32,767 combos). Also tested with DMS additive f(A)+f(B) included as a 16th feature.

Individual Models — Doubles (top-10)

#	Model	Spearman	NDCG@10%
1	ESM-2 650M	+0.159	0.674
2	ESM-2 3B	+0.154	0.669
3	ESM-1v	+0.141	0.677
4	SaProt	+0.140	0.664
5	ProstT5	+0.137	0.668
6	ESM-2 150M	+0.131	0.668
—	DMS additive f(A)+f(B)	+0.127	0.680
7	MSA log-odds	+0.128	0.663
8	OHM ACI	+0.103	0.678
9	ThermoMPNN	+0.066	0.676

Note: ESM-2 650M beats 3B for doubles. ThermoMPNN drops significantly (stability ≠ combinatorial fitness). All PLM additive scores outperform DMS additive f(A)+f(B).

Best Combinations — Doubles

Category	Best Combination	Spearman	NDCG
Best Sp (zero-shot)	ESM2-650M + BLOSUM62 + SaProt + OHM ACI	+0.186	0.676
Best NDCG (zero-shot)	ESM2-3B + 650M + ProtMPNN + MSA-lo + MSA-mf + Cons	+0.145	0.696
Best Sp (+DMS add)	ESM2-650M + BLOSUM62 + OHM ACI + DMS_add	+0.202	0.674
Best NDCG (+DMS add)	ProstT5 + Cons + OHM ACI + RINpy + DMS_add	+0.166	0.703

Key findings for doubles:
1. Different best model: ESM-2 650M (not 3B) wins for doubles. Larger models may overfit to single-position context.

2. Different best combo: ESM2-650M + BLOSUM62 + SaProt + OHM ACI (Sp=0.186) — SaProt and BLOSUM62 enter the top combo for doubles but not singles. BLOSUM62's substitution matrix may capture pairwise compatibility.

3. OHM ACI appears in nearly all top combos for both singles and doubles — allosteric information is consistently valuable.

4. All methods are still weak (Sp < 0.20) — epistasis (mean Δ=−0.70) makes double-mutation fitness fundamentally hard to predict from single-mutation scores alone. Phase 2 experimental doubles remain essential.

21

Summary: Multi-Signal Evidence at a Glance

Summary

Tool	What It Measures	Sp / NDCG	Key Finding for Our Candidate
DMS Fitness	Direct experimental activity+stability+expression	— / —	All 5 mutations are beneficial singles (+1.33 to +1.84)
OHM ACI	Allosteric communication to active site	0.127 / 0.822	A207E on Path 1 (ACI 98.4%), Q217P on Path 3 (handle). Orthogonal paths
RINpy BC	Structural hub identification (betweenness centrality)	0.027 / 0.771	N140D = structural hub #5/258 (rare beneficial hub). Q217P = low BC #253/258 (safe peripheral, allosteric effect via OHM relay not structural network)
ESM-2 3B	Evolutionary constraint (masked marginal, 1246 singles)	0.242 / 0.817	Best single zero-shot predictor. Correctly flags catalytic triad
SaProt 650M	Structure-aware PLM (AA + 3Di tokens)	0.204 / 0.809	Adds structure signal; enters best doubles combo but not singles
ProstT5	Structure-aware PLM (3Di conditional LLR)	0.179 / 0.773	Moderate; enters best doubles NDCG combo
ThermoMPNN	Stability prediction (ddG from PDB)	0.179 / 0.803	Captures stability — orthogonal to PLM. In best singles combo
ProFAM	Autoregressive protein family LM (251M params)	0.154 / 0.816	Moderate Spearman, good NDCG. Family-specific autoregressive model.
FAMPNN	Full-atom protein design (ICML 2025)	0.115 / 0.800	Low Spearman but high NDCG — best at finding top-10%. In best NDCG combo.
Conservation	Evolutionary constraint (150 orthologs, Scorecons)	0.132 / 0.793	Q217P & N140D below median (safe). W104L & R286P conserved — like ICCG
MULTI-evolve	Multi-mutant fitness prediction (FCNN)	— / —	Converges on same top positions. Needs real doubles to unlock
Epistasis (DMS obs.)	Observed non-additive interactions from 2,467 doubles (not a predictor)	— / —	Mean Δ=−0.70. Core×core pairs synergize. Data-derived, not zero-shot.
Best Singles Combo	ESM-2 3B + ThermoMPNN + OHM ACI (rank avg)	0.248 / 0.833	65,535 subsets searched. 3 orthogonal signals: evolution + stability + allostery
Best Doubles Combo	ESM2-650M + BLOSUM62 + SaProt + OHM ACI (rank avg)	0.186 / 0.676	Different optimal combo for doubles. OHM ACI appears in both.

      The multi-signal approach works: Rank averaging ESM-2 3B + ThermoMPNN + OHM ACI (Spearman=0.248, NDCG=0.833) outperforms any single tool. 65,535 subsets exhaustively searched. Our candidate is supported by convergent evidence from DMS fitness, allosteric pathways, network topology, and this zero-shot scoring benchmark.
    

22

Conclusions

0.248

Spearman · NDCG 0.833
Best 3-tool rank avg (65,535 subsets)

4/5

Candidate positions with
4+ tool consensus

+9.44

Positive epistasis
(combo +6.98 vs additive −2.47)

3

Allosteric paths
covered by candidate

LCC is unusually mutationally tolerant (30% beneficial rate), with the best measured 5-mutant reaching +6.98 (3.8× best single)
OHM reveals 3 distinct allosteric paths to the active site; combining mutations from different paths provides orthogonal effects with predicted low epistasis
RINpy identifies rare "beneficial hub" mutations (N140D, S136Y) — positions where network restructuring improves function
Rank averaging ESM-2 3B + ThermoMPNN + OHM ACI (Sp=0.248, NDCG=0.833) outperforms any single tool — validated by exhaustive search of 65,535 subsets of 16 scoring functions. For doubles: ESM2-650M + BLOSUM62 + SaProt + OHM ACI (Sp=0.186, NDCG=0.676)

4/5 candidate positions have 4+ tool consensus — only 5% of all positions achieve this level of multi-signal agreement
ICCG proves the principle: individually deleterious mutations can be combinatorially powerful in LCC — our candidates build on this with more data-driven evidence
MULTI-evolve is limited by lack of experimental doubles — Phase 2 (105 doubles) will unlock its full predictive power
Phase 1 (5 constructs, 2 weeks) is a low-cost, high-information experiment that tests our computational predictions directly

Q217P + W104L + A207E + R286P + N140D — the highest-confidence multi-mutant candidate from 5 computational tools and 8,179 DMS variants.

LCC Structure, Activity & Published Variants

α/β-Hydrolase Fold

LCC WT vs ICCG — Literature Activity

ICCG Mutations in DMS

Key Variants (by improvement over ICCG) + DMS Singles Sum†

LCC Structural Zone Definitions — Binding Site & Secondary Shell

Published Variants vs DMS Dataset — Coverage & Alignment

No published variant combination exists in DMS

Key Observation: D238C and Y127G Missing Everywhere

LCC-LANL: F243I+P38L Antagonism

DMS Fitness Distribution — Single Mutations

Distribution Shape

DMS Dataset Overview — Multi-Site Variants

What is Micro-Droplet DMS?

DMS Fitness Landscape — Top Performers

Key Observation

Conservation vs DMS Fitness — Identifying Engineering Targets

✓ Top-Left: Ideal Targets

⚠ Top-Right: High-Risk High-Reward

✗ Bottom-Right: Avoid

Bottom-Left: Low Priority

Conservation, Hotspots & Coldspots

Conservation Analysis (Scorecons, 150 Orthologs)

Candidate Positions — Conservation Status

Deep Mining Strategy — 5 Computational Tools

Rationale

OHM — Why?

RINpy — Why?

ESM-2 — Why?

MULTI-evolve — Why?

Scoring Benchmark — Why?

Tool 1: OHM Allosteric Communication Analysis

What OHM Computes

Zone Classification

OHM: Allosteric Pathways — Two Perspectives

Path 1: Core Catalytic Relay · ACI > 95%

Path 2: Substrate Access · ACI > 90%

Path 3: Distal Handle · ACI 85-95%

From Ser165 (Nucleophile) Outward

From Asp210 (Acid) Outward

From His242 (Base) Outward

Interactive 3D Structure

Tool 2: RINpy — Residue Interaction Network Analysis

What RINpy Computes

BC vs DMS Fitness

Column Definitions

Tool 3: ESM-2 Protein Language Model

What ESM-2 Computes

Why Use a PLM?

Calibration Against DMS

Structural Criteria: Where Are the Top Mutations?

Binding Site & Secondary Shell in Published Variants

Position 217 — Allosteric Hub

Double Mutation Pattern by Structure

Tool 4: MULTI-evolve — Can We Predict Combinations?

What MULTI-evolve Is

Our Attempt: Zero-Shot Pseudo-Doubles

Validation on Real k=3 DMS (n=1,650)

One-hot features cannot generalize

Alternative 1: 14-dim Features on All Doubles

Alternative 2: Train on Real Top-50 Doubles

Cross-Tool Consensus — Which Positions Do All Tools Agree On?

Consensus Criteria (5 Independent Signals)

Candidate Mutations Mapped onto LCC Structure

The Case for Q217P + W104L + A207E + R286P + N140D

Why This Specific Combination?

Honest Risk Assessment

Experimental Recommendation — Phased Approach

Phase 1: Immediate — 5 Constructs (2 weeks)

Phase 2: Pairwise Doubles — 105 Constructs (4-6 weeks)

Phase 3: MULTI-evolve R2 — ~50 Constructs (2-4 weeks)

DMS Observation: The Epistasis Paradox in LCC

What We Computed

The Paradox

What This Means for Our Strategy

Zero-Shot Scoring Benchmark — 16 Models × 65,535 Combinations

Method

Double-Mutation Benchmark — Additive Zero-Shot Prediction

Method

Summary: Multi-Signal Evidence at a Glance