Accelerating inference in genomic and proteomic foundation models via speculative decoding.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Genomic and protein foundation models (GFMs and PFMs) excel at learning biological sequence languages.
• Autoregressive decoding in these models is slow, limiting large-scale sequence generation due to high latency.
• The computational cost of generating long DNA and protein sequences is a significant bottleneck.

Purpose of the Study:

• To adapt and evaluate speculative decoding for accelerating GFMs and PFMs.
• To assess the impact of speculative decoding on sequence generation speed and prediction quality.
• To investigate the effectiveness of a probabilistic speculative decoding approach.

Main Methods:

• Adapted a probabilistic variant of speculative decoding for GFMs (DNAGPT) and PFMs (ProGen2, ProtGPT2).
• Employed a lightweight draft model to propose token spans, verified in parallel by a larger target model.
• Systematically studied effects of speculation window length, temperature, draft architecture, and prompt length.
• Benchmarked performance in tokens per second across multiple configurations.

Main Results:

• Speculative decoding achieved consistent speedups over standard decoding methods.
• Maximum speedup reached 100%, with average gains between 20% and 40% (1.2x-1.4x).
• The method preserved the target model's sampling distribution, ensuring prediction quality was maintained.

Conclusions:

• Speculative decoding is a practical and model-agnostic strategy for accelerating biological sequence generation.
• This approach effectively reduces latency in GFMs and PFMs without sacrificing predictive performance.
• Enables more efficient large-scale generation of genomic and proteomic sequences.