In a rapidly evolving scientific landscape, computational approaches are transforming how we optimize drugs. During my recent invited short talk at the CDD Community Meeting in Basel on February 27, 2025, I shared insights into leveraging free AI and ML resources for computational drug design (CADD). Here’s a recap of the key takeaways, real-world case studies, and the future potential of AI-driven molecular design.
Free AI & ML Resources for CADD
A growing number of high-quality, freely available AI and ML tools are revolutionizing molecular modeling and cheminformatics. Some of the most valuable resources I regularly incorporate into my work include:
- AlphaFold2: Revolutionizing homology modeling with high-accuracy protein structure predictions.
- Public ML Models & Datasets: Including GitHub-hosted scripts and datasets from ChEMBL and other sources.
- Large Language Models for Biology: ESM-2 and EVO-2 enable deep understanding of protein function.
- Antibody-Specific Models: Ablang2 helps refine antibody sequence prediction.
- Molecular Representation & Prediction Tools: ChemProp for property prediction, RDKit for cheminformatics, and DataWarrior for visualization.
- ChatGPT: Enhancing research workflows through automation and insight generation.
Case Studies: AI in Action
To highlight the real-world impact of AI and ML, I presented three case studies where these tools have been successfully applied.
Case Study 1: AlphaFold2 for Homology Modeling
- Challenge: Modeling a protein target with only 38% sequence identity to known structures.
- Key Benefits: Deep alignment capabilities, reduced human error, and rapid predictions.
- Success Story: AlphaFold2 predicted a structure closely resembling the inhibitor-bound conformation, crucial for rational drug design.
- Limitations: Required additional modeling for co-factors, metals, and ligands, highlighting the need for hybrid AI-physics approaches.
Case Study 2: ChemProp for Molecular Property Prediction
- Objective: Predicting peptide membrane permeability using graph neural networks (GNNs).
- Findings: ChemProp achieved similar predictive accuracy (RMSE 0.75) compared to alternative deep learning models but was more user-friendly.
- Impact: Provided a fast, accessible alternative for cheminformatics researchers without deep ML expertise.
Figure S14 from [1] depicting ChemProp-predicted log P against experimentally measured log P for 425 test set molecules from public AZ dataset (2538 mols).
References
[1] Clemens Isert, Jimmy C. Kromann, Nikolaus Stiefl, Gisbert Schneider, Richard A. Lewis, ACS Omega. 2023 Jan 4;8(2):2046-2056. Machine Learning for Fast, Quantum Mechanics-Based Approximation of Drug Lipophilicity.
[2] Jianan Li, Keisuke Yanagisawa, Yutaka Akiyama, Brief Bioinform. 2024 Jul 25;25(5). CycPeptMP: enhancing membrane permeability prediction of cyclic peptides with multi-level molecular features and data augmentation.
[3] https://github.com/cisert/rescoss_logp_ml
[4] CycPetpMPDB: http://cycpeptmpdb.com
Case Study 3: AI-Boosted Antibody Modeling
- Tools Used: AF2-multimer, Ablang2, and ESM-2.
- Breakthroughs: AI models enhanced affinity prediction, stability assessment, and reduced immunogenicity.
- Challenges: While core scaffolds were well-modeled, variable loops like CDR_H3 showed lower confidence scores, highlighting areas for further improvement.
References
1] My favorite Jupyter notebook implementation of AF2: https://colab.research.google.com/github/sokrypton/ColabFold/
[2] https://github.com/oxpig/AbLang2
[3] Lin Z. et al, Science. 2023 Mar 17;379(6637):1123-1130. Evolutionary-scale prediction of atomic-level protein structure with a language model.
[4] Brian Hie et al, Nat Biotechnol. 2024 Feb;42(2):275-283. Efficient evolution of human antibodies from general protein language models.
[5] Brian Hie and co-workers, bioRxiv, Feb 19, 2025.DOI. Genome modeling and design across all domains of life with Evo 2.
AI in Daily Research: Benefits & Challenges
Beyond case studies, AI and ML are integral to my everyday research activities:
- Enhanced Efficiency: AI-assisted coding and email automation improve productivity.
- Modeling Accuracy: AI reduces errors in structure predictions but requires manual refinement for small molecule interactions.
- Open-Source Value: The accessibility of public models and datasets accelerates innovation without high computational costs.
Future Directions & Open Questions
As AI continues to evolve, several key areas warrant further exploration:
- Integration of Small Molecules with AlphaFold: How can AI better model ligand interactions within proteins?
- New Molecular Representations: Can novel AI-driven descriptors enhance predictive power?
- Merging Physics with ML: How can we combine classical simulation techniques with AI for more accurate predictions?
Conclusion
The fusion of AI, ML, and computational chemistry is unlocking new possibilities in drug discovery. By leveraging free, high-quality AI resources, researchers can accelerate innovation while reducing costs. If you’re interested in implementing these tools into your own projects or exploring collaborative opportunities, let’s connect!
For more insights on AI-driven drug discovery and CADD consulting, visit CADD Consulting GmbH or reach out directly.