🧬 biology

Reconstructability of evolutionary intermediates in generative epistatic landscapes

This paper demonstrates that reconstructing evolutionary intermediates from protein endpoints is a calibrated probabilistic problem where maximum-likelihood predictions often fail to capture plausible histories, and success depends on landscape topology and endpoint mutability rather than sequence divergence alone.

Original authors: Roberto Netti, Martin Weigt

Published 2026-06-29

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Roberto Netti, Martin Weigt

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a detective trying to solve a mystery: You have a "before" photo and an "after" photo of a protein (a tiny biological machine), but the middle photos—the steps the protein took to get from A to B—are missing. The question is: Can we reconstruct the missing middle steps just by looking at the start and the finish?

This paper, by Roberto Netti and Martin Weigt, explores this exact problem. They didn't just guess; they built a virtual laboratory to test how well different computer models can "fill in the blanks" of evolutionary history.

Here is the story of their findings, explained simply.

1. The Setup: A Virtual Time Machine

Since we can't travel back in time to watch proteins evolve in real life, the authors created a simulated universe.

They took a real family of proteins (called Chorismate Mutases) and used a computer model to map out all the possible shapes these proteins can take. Think of this map as a hilly landscape.
Low valleys represent stable, healthy proteins (high fitness).
High peaks represent unstable, broken proteins.
They then simulated millions of "evolutionary journeys" where a protein starts in one valley, wanders around, and ends up in another. They recorded the start, the middle, and the end of every single trip.

Now, they hid the "middle" photo and asked their computer models to guess what it looked like, using only the start and end photos as clues.

2. The Three Guessing Strategies

They tested three different ways to guess the missing middle:

Strategy A: The "Straight Line" Guess (Naive Baseline)
- The Idea: If the start is "Red" and the end is "Blue," the middle must be "Purple." It assumes the protein just changed one letter at a time in a straight line.
- The Result: This works okay for very short trips, but fails miserably for long ones. It ignores the fact that proteins are complex; changing one part often forces changes in other parts (like how changing a tire on a car might require adjusting the suspension). This method creates "impossible" proteins that would break in real life.
Strategy B: The "Perfect Score" Guess (Maximum Likelihood)
- The Idea: The computer calculates the single most probable sequence for the middle step. It picks the "best" amino acid for every spot to get the highest possible score.
- The Result: This guess is very accurate at the individual letter level. However, it's a bit of a trick. It produces a sequence that is statistically "too perfect." It's like a student who memorizes the textbook perfectly but has never actually lived the experience. The resulting protein looks like a "low-cost" path that nature rarely takes. It misses the randomness and variety of real evolution.
Strategy C: The "Realistic Crowd" Guess (Generative Sampling)
- The Idea: Instead of picking the single "best" answer, the computer rolls the dice based on the probabilities it learned. It generates many possible middle sequences.
- The Result: This is the winner. While it might get a few individual letters "wrong" compared to the specific ground truth, the overall picture is much more realistic. It captures the ensemble of possibilities. It understands that evolution is a lottery, not a straight line. If you want to know what a protein could have looked like, this method gives you the most realistic "family" of answers.

3. The Terrain Matters: Valleys vs. Plateaus

The most fascinating discovery is that not all parts of the journey are equally easy to reconstruct. The authors found that the "landscape" itself dictates how much information is lost.

The Deep Valleys (Constrained Regions):
Imagine a protein stuck in a deep, narrow canyon. It can't move much without hitting a wall. Because it's so constrained, there are very few ways to get from point A to point B.
- Result: If your journey stays in these valleys, the start and end points contain lots of information about the middle. You can reconstruct the path very well.
The Flat Plateaus (Permissive Regions):
Now imagine the protein climbs out of the canyon onto a wide, flat, foggy plain. Here, it can move in almost any direction without falling off a cliff. There are thousands of different paths to get from A to B.
- Result: If the journey crosses this "foggy plain," the memory of the specific path is erased. Even if you know the start and end, you can't tell which specific route the protein took because there were so many valid options. The landscape acts like an "information horizon"—once you cross it, the past becomes blurry.

4. The Time Trap: Distance vs. Time

Finally, the paper tackles a practical problem. In real life, we don't know how long two proteins have been evolving apart; we only know how different they look (their "distance").

The Trap: Usually, scientists assume that "more different" means "more time passed."
The Reality: This is wrong. A protein in a "foggy plateau" (high mutability) can change its look very quickly. A protein in a "deep valley" (low mutability) changes very slowly. Two proteins could look equally different, but one might have taken 1,000 years to get there, while the other took 100,000 years.
The Solution: The authors showed that to guess the time correctly, you can't just count the differences. You have to look at how easy it is to change the starting and ending proteins (using a metric called "Context-Dependent Entropy"). By combining the distance with the mutability of the endpoints, the computer can accurately guess the hidden "time" of the journey, allowing for a much better reconstruction.

The Big Takeaway

The paper concludes that we shouldn't try to find a single "true" missing sequence. Evolution is too random for that. Instead, we should use data-driven models to generate a realistic set of possibilities.

However, this only works if the "terrain" allows it. If the evolutionary path crossed a "foggy plateau" where many routes were possible, the start and end points simply don't contain enough clues to reconstruct the middle. The paper teaches us to recognize when we have enough information to solve the mystery, and when the fog is just too thick to see through.

Technical Summary: Reconstructability of Evolutionary Intermediates in Generative Epistatic Landscapes

Problem Statement
While generative models have successfully learned the statistical constraints of protein families to design novel functional sequences, the problem of reconstructing the specific evolutionary paths connecting two distinct wild-type sequences remains a significant challenge. Evolutionary intermediates connect observed proteins, but the sequence of mutational steps producing them is rarely recoverable from extant data alone. The primary difficulties include the lack of ground-truth evolutionary trajectories in nature, the unrealistic assumptions of many simulation models (e.g., ignoring epistasis), and the intrinsic stochasticity of evolution, which makes deterministic reconstruction of deeply divergent sequences impossible. This paper addresses the question: what can and cannot be inferred about evolutionary intermediates from their endpoints?

Methodology
The authors employ a controlled in silico framework to benchmark data-driven reconstruction strategies against ground-truth simulated trajectories. The workflow consists of four main stages:

Landscape Inference: A generative probabilistic model, specifically Boltzmann Machine Direct Coupling Analysis (bmDCA), is trained on natural multiple sequence alignments (MSA) of protein families (primarily chorismate mutases, with supplementary results for $\beta$ -lactamase and RR domains). This model defines a statistical energy landscape $H(S)$ acting as a proxy for the epistatic fitness landscape.
Simulated Evolution: Evolutionary trajectories are simulated within this inferred landscape using a stochastic Markov Chain Monte Carlo (MCMC) process. Mutations are modeled at the nucleotide level to account for genetic code degeneracy, including insertions and deletions. This generates a ground-truth dataset of sequence triples: a starting sequence ( $S_{start}$ ), a temporal midpoint ( $S_{mid}$ ), and a final endpoint ( $S_{end}$ ), spanning evolutionary timescales $\tau$ from $10^1$ to $10^6$ Monte Carlo steps.
Model Training: An autoregressive generative model (arDCA) is trained on these triplets to learn the conditional probability distribution $P(S_{mid} | S_{start}, S_{end})$ . The model is conditioned on both endpoints to predict the intermediate.
Reconstruction Strategies: Three distinct approaches are benchmarked:
- Direct Path: A naive baseline assuming equal probability for variant amino acids at polymorphic sites, ignoring epistasis.
- Greedy Maximum Likelihood (ML): Deterministic decoding where the most probable amino acid is selected at each site sequentially.
- Generative Sampling: Stochastic sampling from the learned conditional distribution $P(S_{mid} | S_{start}, S_{end})$ .

Key Contributions and Results

Reconstruction Accuracy vs. Statistical Faithfulness: The study finds that the "best" point prediction (Maximum Likelihood) is not necessarily the most faithful evolutionary reconstruction. While greedy ML minimizes residue-wise reconstruction error, it produces atypical sequences with statistically low energies that do not reflect the ensemble of plausible histories. In contrast, conditional sampling better captures the global statistical features of the evolutionary path, reproducing the distributions of sequence divergence and statistical energy observed in the ground truth, even if it sacrifices a marginal amount of residue-level accuracy.
Landscape Topology as an Information Horizon: The predictability of intermediates is fundamentally limited by the topology of the fitness landscape, quantified by Context-Dependent Entropy (CDE).
- Constrained Basins (Low CDE): Regions of low statistical energy (high fitness) are mutationally constrained. Endpoints in these regions retain strong information about the path, allowing for reliable reconstruction even at longer timescales.
- Permissive Plateaus (High CDE): High-energy regions allow many alternative mutational routes. When a trajectory passes through a high-CDE region (a "barrier"), the memory of the specific path is erased. If an endpoint lies in a high-CDE plateau, it carries significantly less information about the intermediate than a low-CDE endpoint.
- Barrier Crossing: Even when both endpoints are in constrained basins, if the trajectory crosses a high-CDE barrier at the midpoint, the reconstructability of that midpoint drops substantially. This loss of information is intrinsic to the stochastic dynamics of crossing permissive regions, not merely a result of endpoint divergence.
Inferring Evolutionary Time from Sequence Data: The paper demonstrates that simple sequence divergence (Hamming distance) is an insufficient proxy for evolutionary time in heterogeneous epistatic landscapes. Trajectories starting from high-CDE regions diverge faster than those from low-CDE regions, meaning the same divergence can correspond to vastly different timescales.
- The authors propose inferring the latent evolutionary timescale $\tau$ by integrating the Hamming distance with the mutability estimates (CDE) of the endpoint sequences.
- Using multinomial logistic regression to predict $\tau$ from divergence and CDE allows the use of time-parameterized generative models for real-world applications where time is unknown. This approach yields significantly higher reconstruction accuracy and better-calibrated statistical properties compared to models trained directly on divergence bins.

Significance and Claims
The paper recasts intermediate reconstruction as a calibrated probabilistic problem rather than a search for a single "true" sequence. The authors claim that:

Data-driven models should prioritize identifying when endpoints contain sufficient evolutionary information to reconstruct a path, rather than forcing a deterministic prediction.
The appropriate output for such models is a realistic ensemble of plausible intermediates that reflect the intrinsic stochasticity of evolution, rather than a single maximum-likelihood sequence.
Topological features of the sequence landscape (specifically mutability/CDE) are essential for bridging the gap between artificial simulations and real-case applications, enabling the estimation of evolutionary rates and the accurate placement of intermediates without explicit temporal data.

The work establishes that while intermediate reconstruction is feasible within constrained regimes, it hits a fundamental limit when trajectories traverse permissive, high-entropy regions of the landscape, where path-specific memory is irretrievably lost.