[Paper Review] CLIPZyme: Reaction-Conditioned Virtual Screening of Enzymes

Peter G. Mikhael, Itamar Chinn+1 more|arXiv (Cornell University)|Feb 9, 2024

Advanced Nanomaterials in Catalysis7 citations

TL;DR

CLIPZyme learns a contrastive alignment between reaction representations and enzyme structures to rank enzyme candidates for a given reaction, outperforming EC-based screening and improving results when combined with EC predictors.

ABSTRACT

Computational screening of naturally occurring proteins has the potential to identify efficient catalysts among the hundreds of millions of sequences that remain uncharacterized. Current experimental methods remain time, cost and labor intensive, limiting the number of enzymes they can reasonably screen. In this work, we propose a computational framework for in-silico enzyme screening. Through a contrastive objective, we train CLIPZyme to encode and align representations of enzyme structures and reaction pairs. With no standard computational baseline, we compare CLIPZyme to existing EC (enzyme commission) predictors applied to virtual enzyme screening and show improved performance in scenarios where limited information on the reaction is available (BEDROC$_{85}$ of 44.69%). Additionally, we evaluate combining EC predictors with CLIPZyme and show its generalization capacity on both unseen reactions and protein clusters.

Motivation & Objective

• Motivate scalable, in-silico screening of naturally occurring enzymes to identify catalysts for novel reactions.
• Develop a reaction-encoder and a protein-encoder that can be aligned via a CLIP-style objective to rank enzymes by predicted activity.
• Evaluate performance against EC-based predictors and under scenarios with partial reaction information.
• Assess generalization to unseen reactions and protein clusters, including terpene synthases and unannotated enzymes.

Proposed method

• Encode reactions by combining reactant and product graphs with a pseudo-transition-state graph using atom-mapped data and DMPNN encoders; derive a reaction embedding by encoding the TS graph with a second DMPNN.
• Encode enzymes from AlphaFold-predicted structures using an E(n)-equivariant graph neural network (EGNN) initialized with ESM-2 sequence embeddings to produce protein embeddings.
• Train reaction and protein encoders with a contrastive learning objective to maximize cosine similarity between correct reaction-enzyme pairs and minimize it for negatives.
• Represent enzymes and reactions in a shared embedding space to enable retrieval-style ranking of enzymes for a given reaction (virtual screening).
• Use EnzymeMap-based datasets with atom-mapped reactions and AlphaFold structures; baselines include EC-prediction (CLEAN) and combinations of CLEAN with CLIPZyme for improved ranking.
• Evaluate with BEDROC (alpha=85) and enrichment factors, and compare across various reaction encodings and protein representations.

Experimental results

Research questions

• RQ1 Can CLIPZyme effectively rank enzymes for catalysis of a given reaction using reaction and protein structure embeddings?
• RQ2 Does reaction representation choice (structure-based vs. SMILES/CGR) impact virtual screening performance?
• RQ3 Does combining EC-prediction with CLIPZyme improve screening performance across different levels of EC information?
• RQ4 How well does CLIPZyme generalize to unseen reactions and to proteins dissimilar to the training set, including terpene synthases and unannotated enzymes?

Key findings

• CLIPZyme achieves BEDROC 85 of 44.69% on EnzymeMap test screening, outperforming EC-prediction baselines.
• Combining CLIPZyme with CLEAN (EC predictor) consistently improves retrieval performance across EC levels (e.g., BEDROC 85 increases from 0.96% to 57.03% at Level 1 when combining sources).
• Reaction representation based on the pseudo-transition state yields stronger performance (BEDROC 85 of 44.69%) than language-based SMILES representations.
• EGNN-based protein encodings leveraging structure outperform sequence-only baselines (ESM-2) for screening tasks.
• Terpene synthase dataset shows robust performance (BEDROC 85 of 72.46%), while unannotated EnzymeMap reactions yield meaningful rankings (BEDROC 85 of 42.94%).
• Performance declines when excluding proteins highly similar to training data, highlighting the importance of structural features; still retains ranking utility for top candidates.