BLOSUM Is All You Learn - Generative Antibody Models Reflect Evolutionary Priors

Technical Summary: BLOSUM Is All You Learn - Generative Antibody Models Reflect Evolutionary Priors

Problem Statement
Generative models have become prominent tools for designing antibody sequences, with recent observations suggesting that a sequence's log-likelihood score from these models correlates with its binding affinity. This correlation has led to the use of log-likelihoods as ranking metrics in antibody design. However, the fundamental biochemical basis for why generative model log-likelihoods should correlate with binding affinity remains unclear. This work seeks to investigate the underlying mechanisms driving this relationship.

Methodology
The authors investigate the biochemical basis of model-derived log-likelihoods by comparing them against classical evolutionary similarity metrics. The study focuses on:

• Comparative Analysis: Evaluating the correlation between log-likelihood scores from state-of-the-art generative models (specifically diffusion-based models) and measured binding affinity.
• Metric Substitution: Replacing generative model scores with classical substitution matrix scores, specifically BLOSUM (BLOcks SUbstitution Matrix) similarity scores, calculated between designed antibody sequences and their parental (known binder) sequences.
• Controlled Comparisons: Assessing alternative metrics, including Position Weight Matrices (PWMs) and Position-Specific Scoring Matrices (PSSMs), which do not rely on knowledge of the parental sequence.
• Contextual Variation: Testing the impact of using consensus sequences versus parental sequences as the reference point for calculating BLOSUM scores.

Key Contributions and Results

1. Parity with Generative Models: The study finds that BLOSUM similarity scores between designed and parental antibody sequences correlate strongly with measured binding affinity. This performance is on par with the predictive capabilities of state-of-the-art diffusion-based generative models.
2. Alignment of Metrics: BLOSUM scores align closely with the log-likelihoods generated by multiple generative models. This suggests that these complex generative models are implicitly learning evolutionary priors that are already encoded in classical substitution matrices like BLOSUM.
3. Mechanism of Correlation: The authors propose that when computed relative to a known binder, both BLOSUM scores and log-likelihoods function as approximate measures of sequence distance from that reference. As this evolutionary distance increases, the probability of a candidate sequence being a functional binder decreases, thereby explaining the observed correlation with binding affinity.
4. Limitations of Other Metrics: In contrast, similarity scores based on PWMs and PSSMs, which lack specific knowledge of the parental sequence, demonstrate weaker and less consistent alignment with binding affinity. Their performance is shown to depend heavily on the background sequence data used.
5. Context Specificity: When consensus sequences are used instead of parental sequences to compute BLOSUM scores, the observed correlation with affinity largely disappears. This underscores that the correlation is context-specific and relies on the relationship to the specific parental binder.

Significance
The paper concludes that interpretable, evolution-inspired metrics, such as BLOSUM scores, can effectively complement generative modeling in antibody design. These findings provide a biochemical explanation for the behavior of generative models, indicating that their success in ranking candidates may stem from their ability to capture evolutionary priors rather than novel, complex binding rules. Consequently, these simpler metrics offer valuable insights into both the operational behavior of generative models and the biological relevance of the sequences they produce.