How many phage species remain undiscovered? Species sampling approaches to inform phage discovery

This study employs specialized mathematical and computational estimation approaches to demonstrate that classical non-parametric techniques outperform model-based methods in predicting undiscovered phage species, thereby providing a framework to optimize future isolation efforts for combating antimicrobial resistance.

The Gist

Imagine the world of bacteria as a massive, bustling city. For decades, we've been trying to fight the "bad guys" in this city (harmful bacteria) using antibiotics. But the bad guys are getting smarter, evolving defenses that make our weapons useless. This is the crisis of antimicrobial resistance.

Enter the heroes: Bacteriophages (or "phages" for short). Think of phages as tiny, specialized virus-hunters that only eat specific types of bacteria. They are nature's perfect police force. But here's the catch: to catch a specific criminal, you need the right hunter. If you only have one type of hunter, the criminal might just hide or evolve to escape. To win, we need a massive, diverse "police force" (a phage cocktail) with thousands of different hunters.

The Big Question:
We have a database of known phages, but the universe of phages is huge. The researchers asked: "How many more phage hunters are out there that we haven't found yet? And if we go looking for more, will we find a goldmine or just empty pockets?"

The Detective Work: Counting the Invisible

The authors of this paper are like statisticians playing detective. They didn't just guess; they used a mathematical tool called the "Species Sampling Problem."

Imagine you are at a party trying to guess how many different types of ice cream flavors exist in the whole world.

1. The Sample: You taste 100 scoops from a specific bowl. You find 20 unique flavors.
2. The Mystery: You see that some flavors (like Chocolate) appear 20 times, while others (like "Lava Lamp") appear only once.
3. The Prediction: If you taste 100 more scoops, how many new flavors will you find?

The researchers applied this logic to phages. They looked at the "party" of bacteria (specifically 8 common types like E. coli and Salmonella) and analyzed the "flavors" (phage species) they had already collected in their database (INPHARED).

The Tools: Guessing the Unknown

They tested four different "guessing machines" (mathematical models) to see which one was best at predicting the future:

• The Non-Parametric Guessers (ET & GT): These are like a smart observer who just looks at the pattern of what they've seen so far. "If I saw a lot of rare flavors once, there are probably many more rare flavors out there."
• The Model-Based Guessers (FPG & PYP): These are like a theorist who tries to fit the data into a perfect mathematical curve, assuming nature follows a specific rulebook.

The Verdict: The "smart observer" (specifically the Efron-Thisted or ET estimator) turned out to be the champion. It was the most accurate and didn't need to assume nature followed a rigid rulebook. It worked best when you already had a decent amount of data.

The Results: Who's Full, Who's Empty?

When they projected what would happen if scientists doubled their current collection efforts, the results were very different depending on the bacteria:

• The "Goldmines" (Klebsiella, Streptococcus, Vibrio): These bacterial hosts are like unexplored jungles. If we keep looking, we will find hundreds of new phage species. The "ice cream bowl" is still full of new flavors. We need to keep sampling these!
• The "Exhausted Mines" (Mycobacterium, Salmonella, Escherichia): These are like a small, well-stocked pantry. We've already found almost everything there is to find. If we keep looking, we'll mostly just find the same old flavors again. The "ice cream bowl" is nearly empty.
  • Why? For Mycobacterium, it turns out most of the phages found so far came from a very specific, narrow source (a single type of lab strain). We haven't looked at the wild, diverse versions yet, so the "new" phages we found in the database were mostly just repeats of what we already knew.

The Twist: Time Travel Didn't Work

The researchers tried to predict the future by looking at the past. They took data from 2024 and tried to predict what would be added in 2025.

• The Result: It failed. The models predicted we would find a few new species, but the database actually exploded with many more new species than expected.
• The Lesson: The "party" changed. In 2025, scientists started looking at different types of bacteria or in different environments. The "flavors" available changed. This teaches us that you can't predict the future if the rules of the game change. If we change how we look for phages, our math won't work.

The Bottom Line: What Should We Do?

This paper gives us a roadmap for the future of phage therapy:

1. Stop digging in the same hole: For bacteria like Mycobacterium, we have enough data. Instead of spending money finding more of the same, we should start mixing and matching the phages we already have to create powerful "cocktails" to fight infections.
2. Keep digging in the rich soil: For bacteria like Klebsiella, we are just scratching the surface. We need to keep hunting for new phages to build a stronger, more diverse army.
3. Change the strategy: If we want to find more phages for the "exhausted" bacteria, we can't just look harder in the same place. We need to change our strategy—look in new environments or target different bacterial strains.

In short: We have a map of where the treasure is. Some chests are empty; others are overflowing. The key to winning the war against superbugs isn't just finding more phages, but finding the right ones in the right places, and knowing when to stop digging and start building.

Technical Summary

1. Problem Statement

The emergence of antimicrobial resistance (AMR) necessitates the development of sustainable alternatives, such as bacteriophage (phage) therapy. Effective phage therapy often relies on "cocktails" (combinations of different phages) to prevent bacterial resistance. However, the efficacy of these cocktails depends on the availability of a diverse library of characterized phage species.

The central scientific question addressed is: How many phage species remain undiscovered? More practically, the authors ask: Given an existing dataset of isolated phages, how many new species can be expected if $m$ additional phages are sampled? This is a classic "Species Sampling Problem" (SSP), where the goal is to estimate the number of unseen species ($u(n, m)$) in a future sample based on the frequency distribution of species observed in a current sample ($n$).

2. Methodology

Data Source

The study utilized the INPHARED database, a curated repository of phage genomes. Two temporal snapshots were analyzed:

• DB24: September 2024 version (29,410 phages, 10,934 species).
• DB25: May 2025 version (36,288 phages, 14,604 species).
  The analysis focused on eight common bacterial host genera: Escherichia, Klebsiella, Mycobacterium, Pseudomonas, Salmonella, Staphylococcus, Streptococcus, and Vibrio. Species were defined based on 95% identity over 85% target coverage.

Estimation Models

The authors applied and compared four statistical estimators to predict the number of unseen species:

1. Good-Toulmin (GT): A non-parametric estimator.
2. Efron-Thisted (ET): A smoothed, non-parametric correction of GT, suitable for cases where the future sample size $m$ exceeds the current sample size $n$.
3. Fisher-Poisson-Gamma (FPG): A parametric estimator assuming species abundance follows a Poisson-gamma mixture (negative binomial distribution).
4. Pitman-Yor Prior (PYP): A semi-parametric estimator using a two-parameter prior distribution for set partitions.

Validation Strategy

• Internal Validation: Random subsets (25% to 80% of the data) were used as training sets to predict the number of new species in the remaining "test" set. Performance was measured using Normalized Absolute Error (NAE).
• Temporal Validation: The species distribution from DB24 was used to predict the number of actual new species observed in DB25.
• Diversity Metrics: Hill numbers ($^qD$) were calculated for orders $q=0$ (species richness), $q=1$ (Shannon diversity), and $q=2$ (Simpson diversity) to characterize the underlying biodiversity.

3. Key Results

Estimator Performance

• Non-parametric superiority: For the specific context of phage data, the Efron-Thisted (ET) estimator generally outperformed the parametric (FPG) and semi-parametric (PYP) models, particularly when the training set was large relative to the prediction set.
• Parametric utility: The FPG and PYP estimators showed better performance only when the training set was very small ($m > n$), suggesting that parametric assumptions help when data is sparse but become limiting as data volume increases.
• Error Rates: Internal validation showed that prediction errors were manageable, typically staying within 22% of the true number of unseen species.

Temporal Discrepancy

When predicting the actual growth of the database from DB24 to DB25, the models significantly underestimated the number of new species for most hosts (negative normalized mean errors).

• Exception: Mycobacterium and Vibrio showed different patterns.
• Interpretation: The authors attribute this to a shift in sampling strategy over time. The DB25 snapshot likely included samples from a broader range of host genotypes than DB24, violating the assumption that new samples are drawn from the same underlying distribution. Mycobacterium data, largely derived from the SEA-PHAGES program (narrow host range), showed consistent distribution, leading to accurate predictions.

Future Sampling Efficiency (Saturation Analysis)

Using the ET estimator on the DB25 data, the authors projected the number of new species discoverable if the sample size were doubled:

• High Saturation (Low Discovery Potential): Escherichia, Salmonella, and Mycobacterium are approaching saturation. Doubling the sample size is predicted to yield only ~200, ~50, and ~30 new species, respectively.
• Low Saturation (High Discovery Potential): Klebsiella, Streptococcus, Staphylococcus, and Vibrio show little to no saturation. Doubling the sample size is predicted to yield 300–400+ new species for Klebsiella and Streptococcus.
• Intermediate: Pseudomonas shows a slow approach to saturation.

4. Key Contributions

1. Methodological Benchmarking: The study provides a rigorous comparison of SSP estimators specifically for phage genomics, demonstrating that simple non-parametric methods (ET) are often more robust and cost-effective than complex parametric models for this data type.
2. Quantitative Roadmap: It offers specific, quantitative estimates of "discovery potential" for eight major bacterial pathogens, moving beyond qualitative statements about diversity.
3. Sampling Strategy Insight: The paper highlights that the "law of diminishing returns" applies unevenly across bacterial hosts. It identifies that current sampling efforts for some genera (e.g., Mycobacterium) are exhaustive, while others (e.g., Klebsiella) remain under-sampled.

5. Significance and Implications

• Optimizing Phage Therapy: The results guide researchers on where to focus isolation efforts. For genera nearing saturation (Mycobacterium), the focus should shift from discovering new species to developing cocktails from existing genomes. For under-sampled genera (Klebsiella, Streptococcus), continued isolation is critical to build diverse therapeutic libraries.
• Database Management: The study suggests that the rapid growth of phage databases (as seen in DB25) is partly driven by changes in sampling breadth (e.g., targeting new host strains) rather than just increased sampling depth of the same strains.
• Limitations: The authors note that predictions assume a static sampling strategy. If future sampling targets different host genotypes or environments, the current models cannot predict the exact number of new species, only the efficiency of the current strategy.

In conclusion, the paper establishes that mathematical species sampling approaches are viable tools for optimizing the hunt for novel phages, providing a data-driven basis for deciding whether to continue current isolation protocols or pivot to new sampling strategies.