Elucidating the Size of Chemical Space with Assembly Theory

This paper utilizes Assembly Theory, a first-principles measure of molecular complexity based on recursive bond-joining operations, to re-estimate the size of chemical space, revealing that under drug-like constraints (mass < 500 Da), the number of possible molecules reaches approximately 10^117 at an Assembly Index of 25, growing super-exponentially to double-exponentially with increasing complexity.

The Gist

Imagine you are trying to count every single possible Lego castle you could ever build. You might think, "Well, there are so many ways to snap bricks together that the number is basically infinite." Scientists have tried to guess this number before, often saying there are about 10 to the power of 60 (a 1 followed by 60 zeros) "drug-like" molecules. But those guesses had a flaw: they counted every possible combination of bricks, even the ones that would fall apart immediately or make no sense physically. They didn't ask, "How hard is it to actually build this?"

This paper introduces a new way to count the universe of possible molecules using a concept called Assembly Theory. Think of it not just as counting the final castle, but as counting the minimum number of steps required to build it.

Here is the breakdown of their findings using simple analogies:

1. The "Instruction Manual" Metric

Imagine you have a specific molecule. To build it, you need a set of instructions.

• The Old Way: Just count how many atoms are in the molecule.
• The New Way (Assembly Theory): Count the minimum number of "snap-together" moves needed to construct it from scratch.
  • If you have a long chain of identical beads, you can build it quickly by duplicating a small chunk over and over. This is a "low complexity" object.
  • If you have a molecule where every single part is unique and you have to attach them one by one, that takes many more steps. This is a "high complexity" object.

The authors call this number of steps the Assembly Index. It's like a "difficulty rating" for building a molecule.

2. The "Lego Universe" vs. The "Real World"

The paper distinguishes between two spaces:

• The Assembly Universe: This is the theoretical space of every possible shape you could make with Lego bricks, even if the shape is unstable or impossible to hold together in real life.
• Chemical Space: This is the "Real World" subset. It only includes molecules that are physically stable and can actually exist (like the ones in the GDB-13 database, which contains nearly 1 billion real-world drug-like molecules).

The researchers used the GDB-13 database as a map to see how big the "Real World" chemical space actually is.

3. How Fast Does the Space Grow?

The big question was: As the "difficulty rating" (Assembly Index) goes up, how fast does the number of possible molecules explode?

• The Finding: The number of possible molecules grows super-fast.
  • It grows faster than a standard exponential curve (like compound interest).
  • It grows at a rate somewhere between "super-exponential" and "double-exponential."
  • The Analogy: If you imagine the number of molecules as a balloon, standard growth is like blowing it up slowly. This paper suggests the balloon is inflating so fast it's practically exploding.

4. The "Filter" Effect

The paper also looked at what happens when you put "filters" on the Lego set.

• No Rings: If you only allow straight chains of atoms (no loops), the space grows in a specific way.
• With Rings: If you allow atoms to form loops (rings), the molecules tend to be more "symmetric" (easier to build by copying parts), which changes how the space grows.
• Specific Motifs: If you demand a molecule must have a specific shape (like a square ring), the space shrinks, but it's still astronomically huge.

5. The Final Count

When the researchers applied all the standard rules for "drug-like" molecules (things like: must be under a certain weight, must be stable, must have specific types of atoms) and looked at molecules with an Assembly Index of 25, they calculated the size of this space.

The Result: There are approximately 10 to the power of 117 possible molecules.

To put that in perspective:

• The previous estimate was 10^60.
• The new estimate is 10^117.
• That is a number so large it dwarfs the number of atoms in the entire observable universe.

Summary

The paper argues that the "universe of possible molecules" is not just big; it is mind-bogglingly vast, and it grows at a terrifyingly fast rate as complexity increases. By using a "step-counting" method (Assembly Theory) instead of just counting atoms, they found that even with strict rules for what makes a good drug, the number of possibilities is roughly 10^117. This suggests that finding a specific, useful molecule in this ocean of possibilities is an incredibly difficult task, simply because the ocean is so much bigger than we previously thought.

Technical Summary

Technical Summary: Elucidating the Size of Chemical Space with Assembly Theory

Problem Statement
Chemical space is traditionally estimated to be vast, with heuristic calculations suggesting approximately $10^{60}$ "drug-like" molecules below 500 Da. However, these estimates often fail to account for the structural and synthetic complexity of the molecules enumerated. Existing approaches to quantifying chemical space rely on explicit enumeration (e.g., GDB databases) or implicit generation (e.g., genetic algorithms, deep learning), but neither fully captures the fundamental constraints of constructability or the rate at which the space grows with increasing complexity. There is a need for a first-principles framework to partition chemical space based on the causal complexity required to construct molecules, rather than just their static properties.

Methodology
The authors employ Assembly Theory, a framework that quantifies the causal complexity of an object by measuring the minimum number of recursive bond-joining operations required to construct it. This metric is defined as the Assembly Index ($a$).

1. Theoretical Framework:

   • Molecules are treated as molecular graphs ($O$) constructed from building blocks ($BB$).
   • The Assembly State is defined by the pair $(a, S)$, where $S$ is the size (number of bonds) and $a$ is the Assembly Index.
   • Theoretical bounds are established for the Assembly Index: $\log_2(S) \leq a \leq S-1$. Improved bounds are derived based on the number of building blocks ($|BB|$) and size $S$, denoted as $a_{max}(S, BB)$.
   • The paper utilizes set inclusions to bound the number of molecules up to a given Assembly Index $n$: $|Set(a_{max} \leq n)| \leq |Set(a \leq n)| \leq |Set(l(S) \leq n)|$.

2. Data and Computation:

   • Baseline: The GDB-13 database (containing ~977 million drug-like molecules with up to 13 heavy atoms: C, N, O, S, Cl) is used as a proxy for "Assembly Possible" (physically possible molecules).
   • Universe Proxy: Simple Connected Graphs (SCG) are used to represent the "Assembly Universe" (all combinatorial possibilities).
   • Algorithm: The authors compute the exact Assembly Index for all molecules in the GDB-13 database and for SCGs up to 18 edges using specialized algorithms (e.g., the nauty library).
   • Constraints: The analysis applies various constraints to the GDB-13 data, including:
     • Bond type restrictions (e.g., single bonds only, single/double bonds, specific heteroatom combinations).
     • Structural constraints (e.g., presence/absence of rings, specific ring sizes like cyclobutane).

Key Contributions

• Partitioning by Complexity: The paper introduces a method to partition chemical space into levels defined by the Assembly Index, creating a constructive ordering of molecules based on the minimal causal steps required for their formation.
• Growth Rate Bounds: The study establishes that chemical space grows at least super-exponentially and at most double-exponentially with respect to the Assembly Index.
• Quantification of Constraints: The authors demonstrate how specific physical and structural constraints (atom types, bond types, ring presence) alter the distribution of molecules within the Assembly State Space and affect the growth rate of the space.
• Numerical Validation: The work provides the first numerical confirmation, using the GDB-13 database, that the molecular Assembly Possible grows super-exponentially, supporting previous theoretical conjectures.

Results

• Growth Dynamics: Analysis of Simple Connected Graphs (SCG) shows that the accumulated number of graphs grows close to the double-exponential upper bound. Most graphs in this universe are "duplicate-symmetric" (constructed via recursive duplication).
• Chemical Space (GDB-13):
  • The distribution of GDB-13 molecules in the Assembly State Space lies between the theoretical lower and upper bounds.
  • The space grows super-exponentially up to Assembly Index 4 (the limit of complete data in the database).
  • Constraint Effects:
    • Molecules with only single carbon-carbon bonds exhibit growth close to the double-exponential bound.
    • Introducing multiple bond types or heteroatoms shifts the growth rate away from the double-exponential bound, though the majority of molecules remain duplicate-symmetric.
    • Ring Constraints: The presence of rings increases duplicate symmetry, causing the distribution to lean toward the symmetric bound. Conversely, acyclic (non-ring) molecules behave more like strings, leaning toward the duplicate-asymmetric bound.
• Size Estimation: Under constraints comparable to standard drug-like estimates (molecular mass < 500 Da), the analysis projects a chemical space of approximately $10^{117}$ molecules at Assembly Index 25.

Significance
The paper claims to provide a first-principles estimate of the size of chemical space that accounts for the causal complexity of construction, rather than just combinatorial enumeration. By using the Assembly Index, the authors offer a measurable metric to navigate the vastness of chemical space, distinguishing between regions that are structurally simple (high symmetry) and those that are complex. This approach allows for the quantification of how structural constraints (such as those found in drug discovery) shrink the accessible space. The authors posit that this framework yields insights not only into drug and material design but also into the nature of other combinatorial spaces, such as genetic and protein spaces, by defining the fundamental limits of constructability.