Assembly Theory Overview

Updated 6 July 2026

Assembly Theory is a framework that defines an object's complexity by the minimum number of joining operations required to build it from elementary units, with applications in molecules, strings, and materials.
It aggregates the assembly index and copy number to distinguish objects produced by selection from those arising by undirected processes, underpinning biosignature and evolutionary analyses.
The theory is subject to debate and diverse formulations, with discussions on computational NP-hardness, context-free grammar correspondences, and substrate-specific empirical thresholds.

Searching arXiv for papers on Assembly Theory to support the article. {"query":"Assembly Theory assembly index copy number selection evolution arXiv", "max_results": 10} Here are some relevant arXiv results I found:

Assembly Theory and its Relationship with Computational Complexity (Kempes et al., 2024)
Mapping Evolution of Molecules Across Biochemistry with Assembly Theory (Pagel et al., 2024)
Assembly Spaces: Formal Definitions and Fast Methods for Approximating Assembly Indices (Siebert et al., 13 Jun 2026)
Quantifying the Complexity of Materials with Assembly Theory (Patarroyo et al., 13 Feb 2025)
Elucidating the Size of Chemical Space with Assembly Theory (Parra et al., 9 Jun 2026)
A Combinatorial Theory of Assembly Systems via Generating Functions (Ortiz-Muñoz, 18 Jan 2025)
Templated Assembly Theory: An Extension of the Canonical Assembly Index with Block-Compressed Template (Masierak, 26 Jan 2026) Need answer using same language as query: English. Need only article itself. Need 5-7 numbered h2 sections. Start with intro paragraph no heading. Must cite papers by arXiv id. Need avoid claims not in data. Need probably cover controversies objectively, including critique papers (Abrahão et al., 2024) and (Ozelim et al., 2024), and response (Kempes et al., 2024) and (Walker et al., 2024). But must maintain factual fidelity to provided data. We can say some papers argued equivalence; another paper presented counterexamples/proofs against equivalence. Need carefully attribute.

Need comprehensive article on "assembly-theory". Since audience technical. Need maybe mention multiple formulations of Assembly metric differ between papers. Need avoid inventing resolution. Could discuss formalism varies across literature. Need maybe note conflicting definitions of overall "Assembly" or measure A across papers: (Sharma et al., 2022) defines A=(1/N_T) sum d_i n_i. (Kempes et al., 2024) defines A=sum (n_i-1)a_i. (Pagel et al., 2024) defines A = sum n_i(a_i-1). (Siebert et al., 13 Jun 2026) defines A = sum_{i:a_i>1} a_i N_i / N_T. (Patarroyo et al., 13 Feb 2025) defines A=(1/N_T) sum (n_i e^{a_i}-1). Since user's principle says every concrete claim must appear verbatim in data. We can report that the literature uses multiple ensemble-level quantities under "Assembly"/"assembly measure", naming the explicit formulas and citing the papers. Need not reconcile.

Need structure 5-7 numbered h2 sections. Maybe 1. Core formalism and observables. 2. Assembly spaces and path representations. 3. Computational properties and approximations. 4. Experimental measurement and biosignatures. 5. Chemical and evolutionary applications. 6. Extensions to materials/chemical space and alternate formulations. 7. Debate and open questions.

Need intro paragraph define Assembly Theory. Mention central observables assembly index and copy number. Mention aims: quantify selection, complexity, biosignatures. Cite main papers maybe (Kempes et al., 2024, Sharma et al., 2022, Siebert et al., 13 Jun 2026).

Need use markdown only. Need sections with numbered h2 titles e.g., "## 1. Conceptual scope and central observables". Since user asked numbered h2.

Need maintain concise but comprehensive. Could include small table maybe formulas across papers? Tables allowed. Since helpful to organize different formulas of A. Use <=3 columns. Let's do one table with paper and ensemble-level definition. Need explanations outside table. Ensure brief cells.

Need mention object definitions substrate-independent (Q,T,J) and examples molecules, strings, mineral unit cells, atmospheric species. Mention assembly path, assembly index optimization, path hierarchy.

Need mention assembly depth d(O) from 2409, concurrent construction, adenine example a=7 d=5. Mention joint assembly space maybe.

Need mention computational complexity: ASM-INDEX NP-hard via Subgraph Isomorphism in (Kempes et al., 2024), while Huffman/LZW in P. Also (Siebert et al., 13 Jun 2026) says exact string AssemblyCPP exponential time in |s|; RePair upper bound near-linear, LZ lower bound O(n²) or O(n). Need mention CFG correspondence in strings, smallest CFG equals assembly index. Wait careful: (Siebert et al., 13 Jun 2026) says every assembly path in string space can be turned into CFG and size of smallest CFG that generates target string equals its assembly index. But (Kempes et al., 2024) said AT's assembly index not equivalent to compression length. Potential tension. Need explain carefully: in string assembly space, (Siebert et al., 13 Jun 2026) states smallest CFG size equals string assembly index; at the same time (Kempes et al., 2024) states assembly index is not equivalent to Shannon entropy, Huffman, or LZW, and not in same complexity class as those compression algorithms. This is coherent because smallest CFG problem differs from Shannon/Huffman/LZW. Need mention that.

Need maybe mention examples: strings x=zbzbzc and y=zzzbbc same entropy/Huffman length but assembly index 4 vs 5. Wait there is typo p(a)=0, p(b)=2/3, p(c)=1/3 but x contains z? Actually maybe alphabet inconsistency. But data says verbatim. Need not reproduce suspicious details maybe just say paper gives strings x and y with identical letter-frequency distribution and identical Huffman code lengths but different assembly indices. Avoid specifying counts if data suspect? Could specify as example. Since claim appears verbatim. But to avoid inconsistency, simply say explicit counterexamples are given. For LZW example s_n = z repeated 2^n-1, a(s_n)=n-1 while LZW output bit-length is O(n); the scalings differ exponentially. Need maybe phrase carefully.

Need mention experimental measurement: (Jirasek et al., 2023) gives MA estimated from IR, NMR, MS with correlations 0.75-0.87; combined 0.88 or 0.90. (Siebert et al., 13 Jun 2026) says for molecules measured using mass spec, IR, NMR. Biosignature threshold: Marshall et al. 2021 and Jirasek et al. 2024 showed no abiotic sample contained molecules with a>15, biological up to ≈25-30, copy number threshold >10⁴ defines empirical biosignature cutoff at a≈15. (Walker et al., 2024) emphasizes threshold is empirical and substrate-specific, requiring experimental measurement for each building-block set. Use that in controversy/open questions.

Need applications: mapping natural products (Pagel et al., 2024). Use specific numbers: ~211k natural products, ~70M PubChem, beta≈0.499, k≈0.12, peaks around a≈10-15 with tail beyond 20, d lower by 10-20%, Brefelamide a=11 d=10, fingerprint similarity drops from ≈0.8 to ≈0.15 when contingency loss increases. Drug discovery: contingency loss w=15, 10,000 molecules, handful QED 0.8-0.92. Should mention only if comprehensive.

Need materials/extensions: (Patarroyo et al., 13 Feb 2025) periodic solids, formula a_obj=a_uc + a_p, a_p ≃ k log2(n), HCP/FCC model, random plateau A≈10^8, engineered >10^8. (Parra et al., 9 Jun 2026) chemical space sizes: bounds log2 S ≤ a(O) ≤ S-1, maybe sharper with addition chains, N_{≤n} super-exponential and at most double-exponential, GDB-13 ~9.8×10⁸ molecules, fit p≈1.73 and N_{≤25}~10^117. Also (Ortiz-Muñoz, 18 Jan 2025) generating-function theory of assembly systems—linear polymers F=x+xyF. Might be more niche. Since 7 sections maybe can mention formal extensions and generalizations, including generating functions and templated assembly (Masierak, 26 Jan 2026). But comprehensive article could mention as later extensions. Yet need 5-7 sections. Maybe 6 sections: 1 Formal basis. 2 Assembly spaces and variants. 3 Relation to computational complexity and information theory. 4 Experimental measurement and biosignatures. 5 Applications across chemistry/materials. 6 Controversies, alternate formulations, and open problems. In section 6 mention extensions like generating functions and templated assembly? Or in 5 maybe "Applications and extensions". Need comprehensive but not too sprawling.

Need not mention unrelated A-theory assembly map results etc because topic is assembly-theory, and data includes irrelevant "A-theory assembly map" from algebraic topology. Avoid because different topic despite name. Similarly active colloids etc irrelevant.

Need maybe mention different definitions of Assembly and ensemble quantities. Because across papers there are multiple formulas. Could use table in section 1 or 2.

Need citations at most once per section per paper. Manage.

Need mark interpretations clearly: e.g., "This suggests..." when inferring. Most article can be factual. For controversies, say "Some papers argue..." with citations. Need not resolve beyond noting conflict and empirical focus.

Let's draft with 6 sections and intro.

Potential intro: Assembly Theory (AT) is an experimental and theoretical framework that defines complexity in terms of the minimal number of joining operations required to construct finite, distinguishable objects from elementary building blocks, and uses that quantity together with copy number to quantify selection in observed ensembles (Sharma et al., 2022, Kempes et al., 2024, Siebert et al., 13 Jun 2026). In the current literature AT is applied to molecules, strings, minerals, periodic solids, and atmospheres, with a stated emphasis on physically measurable observables and on detecting objects whose high assembly index and high abundance are difficult to explain by undirected generation alone (Siebert et al., 13 Jun 2026).

Section 1 formalism: define assembly space 𝒜=(Q,T,J), assembly path, assembly index. mention molecular graphs and string spaces. mention assembly depth d and concurrent construction. Then table of ensemble-level measures across papers. Need explain literature uses multiple formulas. Could say "The literature uses several non-identical ensemble-level quantities under the names Assembly, integrated assembly, or assembly measure." Then table with paper and formula. Need include formulas brief. Probably three columns: Paper | name | formula. Use 4-5 rows. Need maybe include exactly formulas. Ensure no multiple paragraphs in cells.

Section 2 assembly spaces etc: assembly universe A_U, A_P, A_C, A_O from (Sharma et al., 2022); joint assembly space, path hierarchy from (Siebert et al., 13 Jun 2026) (Assembly Path, Poset Path, Object Path, Pool Path); molecular and string instantiations. mention exploration dynamics parameter alpha and r(d) from 2409 or 2206. Could include. Maybe section titled "Assembly spaces, path hierarchies, and dynamics". Mention nested spaces, double-exponential etc. Need careful: 2206 says |A_U| roughly double-exponentially with maximal d, A_P exponential, A_C sub-exponential to polynomial, A_O small observed set. Mention selection modeled by P_a ∝ [N_a(t)]^α, α=1 undirected, α<1 directed. 2409 defines α and r(d)=N^{NP(d)/N^{M(d)≈ke^{-βd}.}} Mention.

Section 3 computational properties and info-theoretic relations: mention NP-hardness, strings and CFG, approximation methods RePair, LZ lower bound, performance numbers. mention counterexamples against Shannon/Huffman/LZW. Also mention critique papers that argued equivalence but that AT-complexity paper explicitly gives counterexamples and complexity proofs. Since controversy section later maybe save critique there. But at least describe AT paper claims. Maybe controversy section later can state debate. Here factual from 2406 and 2606. Could mention assembly index in strings equals smallest CFG size per 2606, but is not equivalent to Shannon entropy or standard compression lengths per 2406; Huffman/LZW in P, ASM-INDEX NP-hard. That's interesting.

Section 4 experimental measurement and biosignatures: molecules measured via MS² fragmentation, IR, NMR. Provide correlations and thresholds. Could include details from 2302 and 2606 and (Walker et al., 2024). Mention empirical separation at a≈15 and copy number >10⁴ in tandem MS. Need careful because (Walker et al., 2024) says threshold specifically covalent molecules and requires experimental measurement for each substrate. Could mention that as factual caution. Also mention "assembly index is treated as physical observable" from 2406.

Section 5 applications across chemistry, evolution, materials: natural products mapping, chemical space sizes, materials, HCP/FCC, maybe atmospheres if from 2606. Could cover multiple applications. Need perhaps use two paragraphs: molecules/natural products and materials/chemical space. Include numeric findings. Could cite 2409, (Parra et al., 9 Jun 2026, Patarroyo et al., 13 Feb 2025) once each. That's fine.

Section 6 controversies, critiques, and open directions: mention papers (Abrahão et al., 2024) and (Ozelim et al., 2024) argued equivalence to Shannon/LZ and redundancy; (Kempes et al., 2024) argues not equivalent and not same complexity class as Huffman/LZW; (Walker et al., 2024) argues biosignature thresholds are empirical and substrate-dependent. Mention ongoing issues: energetic barriers, reaction kinetics, environmental constraints ignored presently; temporal ordering absent; future merging with quantum-chemical bond formation costs, autocatalytic sets, exobiology; fast approximation, generating functions, templated assembly. Need cite 2409 for limitations, (Ortiz-Muñoz, 18 Jan 2025) and (Masierak, 26 Jan 2026) for extensions maybe once. Could mention "Later formalisms extend AT to generating functions for equilibrium polymer ensembles and templated assembly with wildcard templates that strictly generalize canonical string-based assembly indices." This is factual from 2501 and 2602.

Need ensure each section doesn't cite same paper multiple times. Could use up to one cite per section per paper.

Let's ensure intro no heading. Section headings h2 with numbers. Need article only. Need not say "Based on." Keep neutral.

Let's produce polished prose. Assembly Theory (AT) is an experimental and theoretical framework in which the complexity of a finite, distinguishable object is defined by the minimum number of recursive joining operations required to construct it from elementary building blocks, and the significance of that complexity is evaluated together with copy number in an observed ensemble (Sharma et al., 2022, Kempes et al., 2024, Siebert et al., 13 Jun 2026). Across the recent literature, AT is presented not merely as a description-length formalism but as a substrate-grounded scheme for quantifying selection, with applications to molecules, strings, minerals, periodic solids, and atmospheres; its central claim is that objects with simultaneously high assembly index and high abundance record a deep causal history that is difficult to obtain by undirected generation alone (Siebert et al., 13 Jun 2026).

1. Core formalism and observables

In the most general recent formulation, an assembly space is a triple $\mathcal{A}=(Q,T,J)$ , where $Q$ is the set of objects, $T\subseteq Q$ is the set of units, and $J\subseteq Q\times Q\times Q$ is a ternary relation encoding one-step joining operations, with $J(x,y,z)=1$ iff $x$ and $y$ can be joined in one step to yield $z$ (Siebert et al., 13 Jun 2026). An assembly path leading to a target set $X\subseteq Q$ is a finite sequence of such joins in which every input is either a unit or an earlier product, and the assembly index $a(X)$ is the minimum path length over all valid paths producing the non-unit elements of $Q$ 0; for a single object $Q$ 1, this is written $Q$ 2 (Siebert et al., 13 Jun 2026). In molecular instantiations, the object is typically a finite undirected graph whose nodes are atoms and edges are bonds, while in string instantiations the units are alphabet symbols and joining is concatenation (Pagel et al., 2024, Siebert et al., 13 Jun 2026).

A closely related literature defines the assembly index as the length of the shortest recursive assembly pathway from a chosen set of elementary building blocks, often using bond types for molecules or symbols for strings (Sharma et al., 2022, Jirasek et al., 2023). For molecular graphs with bond set $Q$ 3, a naïve one-bond-at-a-time construction gives $Q$ 4, whereas reuse of duplicated subgraphs reduces the count; one exact expression given for molecular assembly is

$Q$ 5

where $Q$ 6 is the multiset of duplicated subgraphs reused in the shortest pathway (Jirasek et al., 2023).

The literature also distinguishes assembly index from assembly depth. In the natural-products formulation, the depth $Q$ 7 is defined recursively by

$Q$ 8

so that $Q$ 9, with equality only for strictly linear shortest constructions (Pagel et al., 2024). Adenine is given as an example with $T\subseteq Q$ 0 and $T\subseteq Q$ 1, illustrating that concurrent construction of independent substructures lowers depth relative to path length (Pagel et al., 2024).

The second central observable is copy number: if an observed sample contains distinguishable object types $T\subseteq Q$ 2 with abundances $T\subseteq Q$ 3, then $T\subseteq Q$ 4 is the number of observed instances of type $T\subseteq Q$ 5 (Kempes et al., 2024, Siebert et al., 13 Jun 2026). AT treats copy number as indispensable because complexity without recurrence is not, in this framework, sufficient evidence of selection (Sharma et al., 2022).

The ensemble-level quantity called “Assembly,” “integrated assembly,” or “assembly measure” is not uniform across the literature. Several explicit formulas are in use.

Paper	Name	Formula
(Sharma et al., 2022)	Assembly $T\subseteq Q$ 6	$T\subseteq Q$ 7
(Kempes et al., 2024)	Assembly $T\subseteq Q$ 8	$T\subseteq Q$ 9
(Pagel et al., 2024)	total “Assembly” $J\subseteq Q\times Q\times Q$ 0	$J\subseteq Q\times Q\times Q$ 1
(Siebert et al., 13 Jun 2026)	assembly measure $J\subseteq Q\times Q\times Q$ 2	$J\subseteq Q\times Q\times Q$ 3
(Patarroyo et al., 13 Feb 2025)	Assembly $J\subseteq Q\times Q\times Q$ 4	$J\subseteq Q\times Q\times Q$ 5

These differences show that AT is unified most clearly at the level of the assembly index and copy number, whereas ensemble aggregation remains formulation-dependent.

2. Assembly spaces, path hierarchies, and selection dynamics

AT describes object formation in terms of structured spaces of causal possibility. One formulation distinguishes the Assembly Universe $J\subseteq Q\times Q\times Q$ 6, containing all combinatorially possible objects under the chosen joining rules, from Assembly Possible $J\subseteq Q\times Q\times Q$ 7, which respects known physical constraints, Assembly Contingent $J\subseteq Q\times Q\times Q$ 8, which is the subset actually reachable in a given history, and Assembly Observed $J\subseteq Q\times Q\times Q$ 9, the subset detected with nonzero copy number (Sharma et al., 2022). In that account, $J(x,y,z)=1$ 0 grows roughly double-exponentially with maximal assembly index, $J(x,y,z)=1$ 1 remains exponential though physically filtered, $J(x,y,z)=1$ 2 can shrink to sub-exponential or polynomial scale under strong selection, and $J(x,y,z)=1$ 3 is the small realized residue (Sharma et al., 2022).

A generalized path-hierarchy formalism further organizes representations of an assembly history into four projections: the fully ordered Assembly Path, the partially ordered Poset Path, the Object Path recording only the sequence of products, and the Pool Path recording the unordered set of products (Siebert et al., 13 Jun 2026). These form a commutative “forgetful” lattice in which increasingly coarse representations discard order or join-structure information (Siebert et al., 13 Jun 2026). This is significant because different computational methods and experimental proxies recover different levels of this hierarchy rather than the full ordered path.

Selection is represented in AT as biased exploration of assembly space. In one dynamical model, if $J(x,y,z)=1$ 4 is the number of distinct objects with assembly index $J(x,y,z)=1$ 5 at time $J(x,y,z)=1$ 6, then the probability of choosing an object of index $J(x,y,z)=1$ 7 scales as

$J(x,y,z)=1$ 8

with $J(x,y,z)=1$ 9 corresponding to undirected exploration and $x$ 0 to directed, selection-driven dynamics (Sharma et al., 2022). The resulting growth law for new unique objects at index $x$ 1 is written

$x$ 2

Numerical polymer-chain simulations reported in that work indicate that undirected dynamics rapidly fill local neighborhoods of assembly space, whereas directed dynamics suppress exploration ratio $x$ 3 but reach larger maximum assembly index more efficiently for the same number of steps (Sharma et al., 2022).

The natural-products literature recasts this bias in terms of an exploration ratio between nested spaces. If $x$ 4 denotes the observed natural-product space within a larger physically plausible molecular space, then at depth $x$ 5 the ratio

$x$ 6

quantifies how the observed subspace thins relative to the larger space, with $x$ 7 interpreted as the strength of selection (Pagel et al., 2024). A separate temporal selectivity parameter $x$ 8 is also used there, where $x$ 9 denotes random-walk-like exploration and $y$ 0 indicates drift toward higher-depth objects (Pagel et al., 2024). This suggests a family resemblance between the earlier kinetic model and later empirical depth-ratio analyses, even though the exact observables differ.

3. Computational complexity and relation to compression

A persistent theme in AT is that exact assembly-index computation is combinatorially difficult. A decision formulation,

$y$ 1

is stated explicitly, and deciding whether $y$ 2 is shown to be NP-hard by reduction from Subgraph Isomorphism (Kempes et al., 2024). In the same treatment, Huffman coding and Lempel–Ziv–Welch (LZW) length computation are shown to be in $y$ 3, yielding a formal complexity-theoretic distinction between assembly index and those standard compression lengths (Kempes et al., 2024).

The literature also gives explicit counterexamples to the claim that assembly index is equivalent to standard information-theoretic quantities. One paper constructs strings with the same letter-frequency distribution and the same Huffman code lengths but different assembly indices, and also gives a family $y$ 4 of repeated-symbol strings for which $y$ 5 while the LZW output bit-length scales as $y$ 6, concluding that assembly index is not a function of Shannon entropy, is not a restricted case of Huffman length, and is not in the same computational complexity class as those compression algorithms (Kempes et al., 2024).

At the same time, recent formal work on string assembly spaces establishes a correspondence between string assembly and grammar compression. In that setting, every assembly path of length $y$ 7 can be turned into a context-free grammar in Chomsky normal form with $y$ 8, and the size of the smallest grammar generating the target string equals its assembly index (Siebert et al., 13 Jun 2026). This does not identify assembly index with Shannon entropy or with practical compressors such as LZW; rather, it places canonical string-based AT close to smallest-grammar problems, which are themselves computationally hard (Siebert et al., 13 Jun 2026).

Because exact computation is expensive, approximation algorithms are emphasized. For strings, RePair yields an upper bound with near-linear runtime and empirically lies within $y$ 9 of the true $z$ 0, while classical LZ77 parsing gives a lower bound $z$ 1, computable in $z$ 2 time or $z$ 3 with suffix trees and usually within $z$ 4 of the true value (Siebert et al., 13 Jun 2026). Exact computation via AssemblyCPP is exponential in $z$ 5 and in practice handles $z$ 6 within seconds (Siebert et al., 13 Jun 2026). For molecules, AssemblyGo and AssemblyCpp implement best-first search over partial fragments, using hash-based canonical labeling, priority queues, and early exit conditions to limit fragment explosion in a directed acyclic hypergraph of feasible joins (Pagel et al., 2024).

4. Experimental measurement and biosignature interpretation

A distinctive feature of AT is the claim that assembly index is experimentally measurable. For molecules, the literature treats $z$ 7 as a physical observable that can be estimated from tandem mass spectrometry, infrared spectroscopy, and nuclear magnetic resonance rather than only inferred from full structure elucidation (Jirasek et al., 2023, Siebert et al., 13 Jun 2026). In molecular AT, experimentally motivated recursive bond-cutting pathways are said to correlate with fragmentation patterns, and molecular copy number is the literal abundance of indistinguishable molecules in the sample (Kempes et al., 2024).

Three orthogonal molecular proxies for assembly index have been developed. In the infrared fingerprint region, counting peaks above threshold and binning to $z$ 8 resolution yields an empirical linear predictor; simulated data on $z$ 9 molecules gave $X\subseteq Q$ 0 with $X\subseteq Q$ 1, and experimental data on $X\subseteq Q$ 2 compounds gave $X\subseteq Q$ 3 with $X\subseteq Q$ 4 (Jirasek et al., 2023). For $X\subseteq Q$ 5 NMR with DEPTQ classification, multivariate regression over $X\subseteq Q$ 6 simulated spectra gave

$X\subseteq Q$ 7

with $X\subseteq Q$ 8, and the same model yielded $X\subseteq Q$ 9 on $a(X)$ 0 experimental compounds without refitting (Jirasek et al., 2023). For tandem mass spectrometry, a recursive fragmentation-tree estimator achieved $a(X)$ 1 against computed MA on $a(X)$ 2 experimental molecules up to $a(X)$ 3 (Jirasek et al., 2023). Combined proxies improve performance, reaching $a(X)$ 4 on simulated IR+NMR data and $a(X)$ 5 on a set of $a(X)$ 6 molecules with all three measurements (Jirasek et al., 2023).

This measurement program underlies the biosignature interpretation of AT. A central empirical result summarized in the recent formal review is that no abiotic sample contained molecules with $a(X)$ 7, biological samples contained molecules with $a(X)$ 8 up to approximately $a(X)$ 9– $Q$ 00, and a tandem-MS copy-number threshold of $Q$ 01 copies defines an empirical biosignature cutoff at $Q$ 02 (Siebert et al., 13 Jun 2026). A related paper emphasizes that this threshold is not a theorem of AT but an experimental observation specific to covalent organic chemistry; extending AT to inorganic or other substrates requires new experimental calibration before any threshold can be asserted (Walker et al., 2024).

The resulting biosignature logic is conjunctive rather than purely structural. High assembly index measures deep causal history, while high copy number indicates repeated, directed production; their conjunction is taken as evidence for selection or evolution (Siebert et al., 13 Jun 2026). This framing also explains why AT literature repeatedly contrasts isolated random complexity with abundant assembled complexity (Patarroyo et al., 13 Feb 2025).

5. Molecular evolution, chemical space, and materials

AT has been used to analyze natural products as records of evolutionary contingency beyond genes. In a large-scale study, the natural-product assembly space $Q$ 03 was built from approximately $Q$ 04 natural products in COCONUT, while the broader molecular space $Q$ 05 used approximately $Q$ 06 PubChem molecules (Pagel et al., 2024). The depth-wise exploration ratio followed an exponential decay with $Q$ 07 and $Q$ 08, quantifying how Earth’s biochemical selection narrows accessible chemical space (Pagel et al., 2024). In these data, the assembly-index distribution peaks at moderate depths, roughly $Q$ 09– $Q$ 10, but has a long tail beyond $Q$ 11, and assembly depth is systematically lower than $Q$ 12 by $Q$ 13– $Q$ 14, reflecting concurrent construction of repeated or independent substructures (Pagel et al., 2024). A contingency-reconstruction case study on brefelamide, with $Q$ 15 and $Q$ 16, showed that as contingency-loss parameter $Q$ 17 increased from $Q$ 18 to $Q$ 19, mean fingerprint similarity to brefelamide fell from approximately $Q$ 20 to approximately $Q$ 21 across reconstructed libraries of $Q$ 22 molecules at each $Q$ 23 (Pagel et al., 2024). The same framework was proposed as a route to drug discovery, and under PAINS filtering a reconstruction from fragments of depth $Q$ 24 yielded examples with QED scores from $Q$ 25 to $Q$ 26 (Pagel et al., 2024).

AT has also been used to estimate the size of chemical space by partitioning molecules according to assembly index. For molecular graphs of bond count $Q$ 27, one paper gives the bounds

$Q$ 28

and, with addition-chain refinement,

$Q$ 29

(Parra et al., 9 Jun 2026). On this basis, chemical space is studied through level sets $Q$ 30, with cumulative count $Q$ 31 shown to grow at least super-exponentially and at most double-exponentially with $Q$ 32 (Parra et al., 9 Jun 2026). Using GDB-13, which contains roughly $Q$ 33 molecules, a three-parameter fit in a constrained drug-like region gave

$Q$ 34

and extrapolation to $Q$ 35 yielded $Q$ 36 molecules, far above heuristic estimates based only on atom counts (Parra et al., 9 Jun 2026).

The extension of AT to periodic solids and materials introduces nested assembly levels. A crystal is decomposed into assembly of a unit cell with index $Q$ 37 and recursive assembly of periodic repetition with index $Q$ 38, so that

$Q$ 39

(Patarroyo et al., 13 Feb 2025). For an approximately cubic crystal with $Q$ 40 cells along each axis, $Q$ 41 with $Q$ 42 for one-, two-, or three-dimensional repetition (Patarroyo et al., 13 Feb 2025). In a one-dimensional model of HCP-to-FCC transformation, random stacking-fault realizations showed a material Assembly $Q$ 43 that rose with fault density $Q$ 44 up to about $Q$ 45– $Q$ 46 and then plateaued around $Q$ 47, whereas engineered periodic faulting could exceed that plateau; the paper therefore interprets crystalline samples with $Q$ 48 in that model as signatures of selection or technology rather than abiotic randomness (Patarroyo et al., 13 Feb 2025).

6. Debate, limitations, and recent extensions

AT has generated an explicit methodological controversy. Two critical papers argue that assembly index is equivalent to Shannon entropy, LZ-family compression, or the Block Decomposition Method, and conclude that AT adds no explanatory power beyond classical information theory or algorithmic complexity (Abrahão et al., 2024, Ozelim et al., 2024). By contrast, the computational-complexity paper responds with explicit counterexamples showing that assembly index is not a function of Shannon entropy, Huffman code length, or LZW compression length, and proves that the decision problem for assembly index is NP-hard while Huffman and LZW lengths are computable in polynomial time (Kempes et al., 2024). The disagreement is therefore not merely rhetorical; it concerns both formal equivalence claims and the ontological status of assembly index as a physical observable.

A second, more empirical limitation concerns substrate dependence. The molecular biosignature threshold near $Q$ 49 is repeatedly presented as an experimentally established fact for covalent organic molecules, not as a universal constant of the theory (Walker et al., 2024, Siebert et al., 13 Jun 2026). This implies that cross-substrate comparisons require care: changing the building blocks, joining rules, or measurement protocol changes the relevant assembly space and may change any observed threshold.

The most frequently acknowledged internal limitation is that current AT treatments are largely topological and typically ignore energetic barriers, reaction kinetics, and environmental constraints (Pagel et al., 2024). Temporal ordering of fragment emergence is also absent from standard formulations, though later work suggests that a time-dependent selectivity parameter $Q$ 50 could be defined if intermediate species were historically dated (Pagel et al., 2024). A plausible implication is that present assembly indices quantify accessible causal structure under an idealized rule set rather than full mechanistic feasibility.

Recent extensions broaden the framework rather than settling these debates. A combinatorial generating-function treatment formulates assembly systems in terms of bond structures and valid assemblies, deriving recursion relations such as

$Q$ 51

for linear polymers and using $Q$ 52 to study equilibrium polymer ensembles (Ortiz-Muñoz, 18 Jan 2025). A separate string-theoretic extension, templated assembly theory, augments canonical concatenative assembly with wildcard-bearing block-compressed templates, defines a templated assembly index $Q$ 53, and proves $Q$ 54, with strict separation possible in concrete examples (Masierak, 26 Jan 2026). These developments indicate that “Assembly Theory” now refers to a growing family of related formalisms centered on recursive construction, reuse, and causal histories, rather than a single universally fixed equation.