On the rise of AI technologies in virtual screening

The paper demonstrates that the AI foundational model Boltz-2 significantly outperforms existing rescoring strategies in virtual screening, achieving twice the success rate on challenging datasets and enabling accurate, efficient ranking of millions of compounds on commodity hardware.

The Gist

Imagine you are a treasure hunter looking for a specific, rare gem (a drug) hidden inside a mountain so vast it contains billions of pebbles (chemical compounds). This is the daily reality of drug discovery.

For decades, scientists have used a "metal detector" called Virtual Screening to scan these mountains. The old metal detectors (traditional computer docking) are fast and can scan millions of pebbles in a day, but they are often clumsy. They might ring the alarm for a shiny rock that isn't a gem, or miss a real gem because it's buried under a weird layer of dirt (complex protein shapes).

Recently, a new, super-smart AI detective called Boltz-2 has entered the scene. This paper is the report card on how well this new detective performed when tested against the old metal detectors.

Here is the story of the findings, broken down simply:

1. The Challenge: The "Impossible" Test

The researchers didn't just test Boltz-2 on easy pebbles. They gave it a "final exam" using a dataset called ULVSH.

• The Setup: Imagine a pile of 943 pebbles that had already been picked by the old metal detectors. The problem? This pile was a mix of real gems (active drugs) and fake rocks that looked exactly like gems (inactive compounds).
• The Difficulty: In this specific pile, the fake rocks were so good at faking it that even the best old methods couldn't tell the difference. It was like trying to find a needle in a haystack where the hay was also made of needles.

2. The Showdown: Boltz-2 vs. The Old Guard

The researchers ran Boltz-2 against eight other popular methods (the "old guard").

• The Result: Boltz-2 was the clear winner. It correctly identified the real gems twice as often as the next best method.
• The Score: If you imagine a test where 0.7 is a "passing grade," Boltz-2 got a 0.70 average score. The other methods mostly failed, scoring around 0.60 or lower. In fact, Boltz-2 passed the test on 7 out of 10 targets, while the others passed on 0 to 3.

3. The Speed vs. Accuracy Trade-off

Here is the catch: Boltz-2 is slower.

• The Old Metal Detectors: Can scan a pebble in a fraction of a second. They are like a high-speed train.
• Boltz-2: Takes about 100 seconds to analyze a single pebble. It's more like a thoughtful detective who stops to examine every detail.
• The Verdict: You can't use Boltz-2 to scan the entire mountain of billions of pebbles; it would take too long. However, it is perfect for the second step.

4. The New Strategy: The "Filter and Refine" Approach

Since Boltz-2 is too slow for the whole mountain but too smart to ignore, the authors suggest a new workflow:

1. The Sweep: Use the fast, old metal detectors to scan the billions of pebbles and pull out the top 1,000 most promising candidates.
2. The Deep Dive: Hand those top 1,000 to Boltz-2. Let the AI detective take its time, look closely, and re-rank them.
3. The Payoff: This process "rescues" real gems that the fast detector missed. In their tests, this boosted the success rate by 4 to 5 times.

5. The Glitches and Limitations

Even the best detectives make mistakes.

• The Blind Spots: Boltz-2 failed on two specific targets (CNR1 and MTR1A). The researchers tried everything to fix it—changing settings, using better computer chips, or giving the AI more hints—but it still struggled. This suggests that for some very specific types of proteins, the AI just doesn't have enough experience yet.
• The "Look-Alike" Problem: Sometimes, Boltz-2 guessed the right answer (the gem) but got the position wrong (it thought the gem was buried in the wrong spot). Surprisingly, this didn't stop it from identifying the gem correctly, which is a bit of a mystery.

6. The Big Picture: A Paradigm Shift

The paper concludes that we are witnessing a paradigm shift.

• Before: We had to choose between "Fast but inaccurate" (Docking) or "Slow and accurate but too expensive" (Physics simulations).
• Now: Boltz-2 offers a "Goldilocks" solution. It is accurate enough to be trusted, and fast enough to be useful on a standard computer cluster.

The Analogy:
Think of drug discovery like hiring a team to find a lost child in a massive stadium.

• Old Method: You hire 1,000 security guards with walkie-talkies to run laps. They cover the whole stadium quickly but might miss the child hiding in the shadows.
• Boltz-2: You hire one highly trained, super-intelligent tracker. They can't run the whole stadium alone, but if you give them the top 100 spots where the guards think the child might be, they will find the child with incredible certainty.

Summary

This paper proves that AI is ready to join the drug discovery team. It's not replacing the fast scanners yet, but it is becoming the ultimate "second opinion" that can turn a list of "maybe" candidates into a list of "definite" drug leads, saving time and money in the long run.

Technical Summary

1. Problem Statement

Virtual screening (VS) is a cornerstone of drug discovery, yet traditional methods face significant limitations:

• Traditional Docking: Struggles with receptor flexibility, solvent effects, and entropic contributions, leading to poor binding affinity predictions.
• Physics-Based Methods (ABFE/RBFE): While highly accurate, they are computationally prohibitive for large-scale screening (millions of compounds).
• Current Rescoring Gap: Non-rigorous rescoring methods (e.g., empirical scoring, MM/GBSA, semi-empirical QM) have historically failed to reliably distinguish true binders from inactives in "ultra-large" datasets, particularly when the dataset contains high hit rates and chemically similar compounds.
• Specific Challenge: The authors focused on the ULVSH dataset, a difficult benchmark comprising 943 virtual hits from ten ultra-large screening campaigns (targeting GPCRs, kinases, etc.) that underwent experimental validation. Previous studies (Sindt et al.) showed that no standard rescoring function could reliably classify actives vs. inactives in this specific dataset.

2. Methodology

The authors evaluated Boltz-2, an AI foundational model capable of co-folding proteins and ligands and predicting binding affinity, against the ULVSH dataset.

• Dataset: 943 ligands (427 true positives, 516 false positives) across 10 targets (7 GPCRs, 1 kinase, 1 membrane receptor, 1 transporter).
• Workflow:
  1. Input Preparation: Automated pipeline using ligand MOL2 coordinates and protein FASTA sequences.
  2. Prediction: Boltz-2 generated 3D protein-ligand complexes and predicted binding affinities.
  3. Hardware: Executed on a single NVIDIA RTX 4500 Ada GPU (approx. 100 seconds per ligand).
  4. Comparison: Performance was benchmarked against eight existing rescoring strategies (Docking, HYDE, Δvina, Gnina, MMPBSA, MMGBSA, GFN-FF, PM6) and end-point free-energy simulations.
  5. Metrics: Primary metric was ROC-AUC (Receiver Operating Characteristic - Area Under Curve). A successful classification was defined as ROC-AUC > 0.7.
• Robustness Checks: The authors tested various hypotheses to explain failures (CNR1 and MTR1A targets), including:
  • Ensemble predictions (mean/best/worst of 20 runs).
  • Template conditioning (using experimental PDB structures).
  • Single-domain inputs (removing multi-domain noise).
  • Increased inference parameters (more sampling steps/recycling).
  • Structural validation (comparing predicted poses to experimental co-crystals).
• Secondary Experiment: A "drug discovery-like" scenario using the LSD database (20M–468M compounds). The top 1,000 docking hits were re-ranked by Boltz-2 to test enrichment capabilities.

3. Key Contributions

• Benchmarking AI on Hard Data: This is one of the first rigorous evaluations of a generative AI co-folding model (Boltz-2) on a challenging, experimentally validated ultra-large screening dataset where traditional methods failed.
• Performance Validation: Demonstrated that Boltz-2 significantly outperforms all tested non-rigorous rescoring methods and standard docking in distinguishing actives from inactives.
• Throughput Analysis: Quantified the computational cost, establishing that while Boltz-2 is slower than docking, it is fast enough for "mid-scale" rescoring (millions of compounds) on commodity hardware.
• Failure Mode Analysis: Systematically investigated why the model failed on two specific targets (CNR1, MTR1A) and found that standard parameter tuning or structural conditioning did not resolve the issue, suggesting inherent limitations in the model's domain of applicability for certain targets.
• Correlation Insight: Highlighted a critical finding: Binding affinity prediction accuracy is not correlated with structural pose accuracy. Even when the predicted pose had a high RMSD (>2 Å) compared to the crystal structure, the affinity prediction could still be accurate enough for classification.

4. Key Results

• Classification Performance (ULVSH):
  • Boltz-2 Average ROC-AUC: 0.70 (Standard Error: 0.05).
  • Success Rate: 7/10 targets achieved ROC-AUC > 0.7.
  • Comparison: The next best method (PM6) had an average ROC-AUC of 0.64 with a success rate of only 2/10. Boltz-2's success rate was more than double that of the best alternative.
  • Specific Targets: Boltz-2 excelled on targets like SC6A4 (0.86), ROCK1 (0.84), and DRD3 (0.81). It failed on CNR1 (0.40) and MTR1A (0.45).
• Enrichment (LSD Database):
  • When re-ranking the top 1,000 docking hits from the LSD database, Boltz-2 provided a 4–5 fold enrichment of true actives over docking alone in the top-ranked fractions (top-100 to top-750).
• Computational Efficiency:
  • Speed: ~100 sec/ligand on a single GPU.
  • Throughput: ~1,000 ligands/day (single GPU) or ~500,000 ligands/month (20-GPU cluster).
  • Limitation: This is 3–4 orders of magnitude slower than ultra-fast docking (e.g., Uni-Dock at 0.1 sec/lig), making it unsuitable for screening billions of compounds directly but ideal for rescoring the top 0.1–1% of a library.
• Structural vs. Affinity:
  • In 5 out of 10 cases, Boltz-2 predicted poses with RMSD > 2 Å compared to experimental structures.
  • Despite poor structural predictions, the affinity predictions remained robust for classification, indicating the model learns interaction patterns independent of perfect geometric alignment.

5. Significance and Conclusion

• Paradigm Shift: The study suggests a shift in virtual screening workflows. Instead of relying solely on docking followed by physics-based rescoring, AI co-folding models like Boltz-2 offer a "middle ground" that bridges the gap between ultra-large library screening and lead optimization.
• Practical Workflow: The authors propose a hybrid workflow:
  1. Ultra-fast docking to filter billions of compounds down to ~100,000.
  2. Boltz-2 Rescoring to refine the top ~1,000–10,000 compounds (achieving 2–5x enrichment).
  3. ABFE (Absolute Binding Free Energy) calculations on the top hits for final validation.
• Caveats: The authors acknowledge that recent independent studies (Bret et al., Shen et al., Kim et al.) confirm Boltz-2's superiority in classification but warn about its domain of applicability. The model may struggle with targets absent from training data or specific mutation scenarios, and its predictions are not always correlated with structural correctness.
• Conclusion: Boltz-2 represents a robust, efficient, and accurate tool for ligand classification, capable of rescuing true binders that traditional docking misses, provided the target falls within the model's learned chemical space. It opens the door to accurate virtual screening on commodity hardware for pharmacologically relevant targets.