📄 evolutionary biology

Forecasting viral evolution from phylogenetic trees

The paper introduces antiGen, a machine learning model that leverages phylogenetic trees to accurately forecast future viral mutations across multiple pathogens, including SARS-CoV-2, influenza, and dengue, thereby enhancing pandemic preparedness and therapeutic development.

Original authors: Specht, I., Park, S., Chithrananda, S., Driscoll, C. L., Brixi, G., Palacios, J. A., Hie, B. L.

Published 2026-09-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Specht, I., Park, S., Chithrananda, S., Driscoll, C. L., Brixi, G., Palacios, J. A., Hie, B. L.

Original paper licensed under CC BY 4.0 (https://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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Viruses are relentless shapeshifters. As they multiply inside a host, they make tiny copying errors in their genetic code, creating new versions of themselves. Most of these changes do nothing, but occasionally, a mutation gives the virus a new advantage, such as the ability to spread faster or dodge the immune system. This constant evolution is why a flu shot from last year might not protect you this year, and why new variants of diseases like COVID-19 can emerge unexpectedly. To stay ahead of these threats, scientists need to predict which changes are likely to happen next. Traditionally, researchers have tried to guess these future mutations by studying how proteins behave in a lab or by using computer models that look at vast libraries of existing genetic sequences. However, these methods often struggle to see the full picture of how a virus actually moves through time and populations.

A team of researchers at Stanford University and the Arc Institute has developed a new approach that looks at the virus's family history to predict its future. They created a machine learning model called antiGen, which learns from the evolutionary path of a virus rather than just its current genetic code. Instead of simply memorizing sequences, the model studies the specific steps a virus took to get from one version to another, much like a historian studying the specific events that led to a major historical shift. By analyzing the "family tree" of a virus—a diagram that maps out how different strains are related and when they diverged—the model learns the rules that govern which mutations are likely to appear next. This method allows the system to anticipate changes that have never been seen before, offering a powerful tool for preparing vaccines and treatments before a new variant even spreads.

The researchers tested this system first on the SARS-CoV-2 virus, the cause of the COVID-19 pandemic. They fed the model millions of genetic sequences collected over the first year of the pandemic and asked it to predict mutations that would appear in the years following. The model performed better than any existing method at forecasting the next likely changes to the virus's spike protein, the part of the virus that allows it to enter human cells. Crucially, it also succeeded in predicting mutations that had never been observed in the training data, suggesting it had learned the underlying principles of viral evolution rather than just memorizing past events. To confirm these predictions were biologically real, the team created the predicted virus variants in a lab. They found that most of the computer-generated mutations produced viruses that could still infect cells, proving that the model was proposing viable, functional changes rather than random errors.

The success of antiGen was not limited to the coronavirus. The team applied the same model to three other common viruses: seasonal influenza, respiratory syncytial virus (RSV), and dengue virus. Even though there was far less genetic data available for these viruses compared to SARS-CoV-2, the model still outperformed traditional methods in predicting their future mutations. This suggests that the approach works across a wide range of viruses, regardless of how much surveillance data is available. The researchers also discovered that the model's accuracy improved as more data was collected, meaning it becomes a more powerful tool the longer we continue to track viral evolution. In some cases, combining the model's predictions with existing statistical methods created an even more robust forecasting system, capable of handling both data-rich and data-poor scenarios.

While the model is highly accurate at predicting which mutations are likely to occur, the researchers emphasize that it does not create dangerous new pathogens. The changes it predicts are small, single-step mutations that are already part of the natural evolutionary process. These predictions are intended to help public health officials and vaccine developers stay one step ahead. By knowing which mutations are likely to emerge, scientists can design vaccines that target not just the virus circulating today, but the versions that will likely dominate tomorrow. This proactive approach could transform how we manage infectious diseases, shifting the focus from reacting to outbreaks to preventing them before they gain a foothold. The study demonstrates that by listening to the history written in a virus's genetic code, we can begin to read its future.

Technical Summary: Forecasting Viral Evolution from Phylogenetic Trees

Problem Statement
Viral mutation forecasting is critical for pandemic preparedness, enabling the anticipation of novel variants and the design of proactive interventions. While evolutionary histories represented as phylogenetic trees offer insights into past strain emergence, their utility in predicting future sequence changes has remained largely unexplored. Existing computational methods, such as protein language models (PLMs) trained on multiple sequence alignments (MSAs) or deep mutational scanning (DMS) experiments, face limitations. DMS measures fitness independently of mutational accessibility and is time-intensive, while PLMs often struggle with pandemic variant prediction without fine-tuning and can be biased by the overrepresentation of certain genotypes in training data. Current leading approaches, including EVE and EVEscape, underperform basic statistical predictors on several evaluations.

Methodology: The antiGen Model
The authors introduce antiGen, a machine learning model designed to forecast the evolutionary future of viruses by learning directly from their evolutionary past encoded in phylogenetic trees, rather than from raw sequence databases.

Data Generation and Tree Construction: The authors developed a high-performance pipeline to construct pandemic-scale phylogenetic trees with millions of leaves, inspired by the UShER algorithm. This process involves taking a "guide tree" with inferred ancestral states and iteratively attaching sequences from a large protein MSA to the tree at the location that introduces the fewest new mutations (parsimonious placement).
Training Objective: Unlike PLMs that train on sequences, antiGen is trained on "parent-child pairs" extracted from the phylogeny. Each training example consists of an inferred ancestral genotype and the set of single amino acid substitutions inferred to have altered that genotype. The model learns the joint probability distribution over the genomic position and the new amino acid state involved in a substitution, conditional on the starting sequence.
Architecture: antiGen utilizes a small, randomly initialized Transformer encoder (approx. 465k–565k parameters). It encodes amino acids via learned token and absolute positional embeddings, processes the sequence through two Transformer layers, and outputs per-position logits. A key design departure from standard language modeling is the application of the softmax function over the entire output matrix rather than per position, preserving each position's propensity to mutate.
Evaluation Metrics: The study introduces two primary metrics:
1. Next-mutation recall: Measures the proportion of true future mutations falling within the top $q$ -quantile of predicted mutations.
2. PSSM precision: Assesses the ability to predict the set of mutations observed in future population-wide genomic surveillance data (descendants of a specific root sequence) by scoring the probability that top-decile predictions exceed an observation threshold in a Position-Specific Scoring Matrix (PSSM).
Hybrid Approaches: The authors explore fine-tuning existing PLMs (EVE, Tranception, ESM-2) on the antiGen objective and a "gated" predictor that blends antiGen outputs with a PSSM baseline using learned, position-specific weights.

Key Results

SARS-CoV-2 Performance: Trained on the first year of pandemic data (323,950 sequences) and tested on subsequent years (9+ million sequences), antiGen achieved state-of-the-art performance in predicting next mutations, including de novo mutations never seen in training. It outperformed DMS experiments and existing PLMs in next-mutation recall and PSSM precision. For predicting any mutation to appear at a population frequency ≥1 downstream of BA.2, antiGen marginally surpassed the PSSM; however, for predicting mutations appearing at a frequency ≥100, the PSSM and EVE marginally outperformed antiGen.
Experimental Validation: The authors experimentally validated antiGen's predictions using pseudovirus infectivity assays. Most antiGen-forecasted SARS-CoV-2 spike mutants, including de novo variants, retained pseudoviral infectivity, confirming the functional plausibility of the proposed mutations.
Endemic Viruses: Despite significantly less available sequencing data, antiGen demonstrated leading predictive performance on surface proteins of Influenza (H3N2), Respiratory Syncytial Virus (RSV), and Dengue virus, outperforming baselines in both next-mutation recall and PSSM precision.
Data Efficiency and Hybrid Models: antiGen's performance improves with increased surveillance data. While antiGen alone excels in high-data regimes, a gated model combining antiGen with a PSSM baseline showed robust performance across varying data volumes, particularly in early pandemic (low-data) settings where the PSSM provided complementary signal. Fine-tuning EVE on the antiGen objective further improved performance on specific tasks, such as predicting de novo mutations downstream of the BA.2 variant.

Significance and Claims
The paper claims that antiGen represents a significant advance by explicitly learning from phylogenetic structure, aligning the training objective with the prediction task of forecasting evolutionary changes. The authors posit that this approach offers a valuable resource for applications ranging from epidemiological modeling to therapeutic development, specifically for anticipating mutations that escape antibody therapeutics or selecting vaccine strains.

The study emphasizes that while antiGen provides reliable predictions for future mutations and those maintaining in vitro infectivity, single-substitution variants are unlikely to pose unprecedented biosecurity risks, as they retain high sequence homology detectable by existing synthesis screens. The authors conclude that the model's improved performance with additional data underscores the importance of widespread genomic surveillance for public health decision-making and that the method could better equip authorities to preempt future variants posing serious risks to human health.

More like this