Antibody-antigen complexes are a little like Bach fugues: interlocking motifs, precise, and endlessly varied. However, for drug developers, these complexes remain one of the most challenging puzzles in drug design today.
When it comes to targeting antigens, antibodies offer distinct advantages over small molecules and other therapeutic modalities, combining high target specificity with the ability to bind more effectively to a broader range of complex previously “undruggable” targets.
Despite this, the structural complexity and variability inherent in these biomolecular complexes present profound technical challenges. Antibodies have protean qualities, adopting multiple conformations which permit their incredible recognition repertoire. Antigen surfaces may also vary widely, making accurate prediction and analysis of structure exceptionally difficult in some cases. Finally, the two partners may profoundly influence each other, making the final pairing far more than a mere 3D jigsaw puzzle affair.
As in many areas of science and drug discovery, AI modelling has emerged with the potential to resolve this sometimes intractable challenge. By learning from all the accumulated knowledge we have of structures and sequences, AI can generate a number of potential antibody structures that resolve the dance of coordinated dynamic flexibility and find solutions to the puzzle.
The question is, does the AI rhetoric live up to reality? What practical improvements can it deliver? And what challenges remain unresolved by AI technology as it stands?
Beyond the hype: AI’s potential and practical limitations
There has been much talk about the transformative effect of AI. Yet, these advances in technology have genuinely opened previously locked doors in biomolecular prediction. AI has enabled the development of more accurate, flexible, and biologically relevant models, creating the conditions for greater therapeutic innovation. Some of the most striking developments include:
Unlocking structure prediction: Models like AlphaFold 3 now predict protein and other biomolecular structures with sometimes uncanny accuracy. Using an innovative transformer-based architecture, it can capture complex interactions among proteins, nucleic acids, and small molecules, outperforming earlier tools for drug design.
Designing new proteins: Models such as RFdiffusion can generate entirely novel protein folds and binding scaffolds, expanding the universe of possible therapeutics far beyond what evolution has sampled.
Accelerating antibody discovery: Advances in integrated AI frameworks now allow the simultaneous modelling of antibody sequence, structure, and binding specificity. By linking sequence-level variation to molecular interaction profiles, these models can prioritise candidates with high potential for stable and specific target engagement early in discovery.
Despite these breakthroughs, challenges remain. One of the most striking is how these models’ performance on antibody-antigen complexes trails that of other proteins. Another related challenge is the benchmarking of AI models that predict antibody-antigen complexes. These tools can be enormously helpful to serious drug hunters, but many users are left on their own to judge when the results are reliable or not.
Benchmarking existing methods: Mapping future breakthroughs.
Benchmarking may not sound like the most exciting topic, but measuring the best computational approaches, including AI methods, is crucial to identifying priority areas for future breakthroughs. This will enable us to develop a detailed blueprint for targeted improvements in predicting antibody-antigen complexes. And that brings a certain satisfaction for any scientist working in the field
In this sense, benchmarking acts as a compass, charting the path for future discovery and clinical translation based on reliable, tried and tested methods
Despite the historic limitations, there have been significant improvements in benchmarking in recent years, evolving alongside the predictive models. Comprehensive datasets have been curated, such as the ABAG-docking dataset, consisting of over 170 diverse antibody-antigen complexes categorised by docking difficulty and structural features. These curated resources facilitate the identification of methodological gaps and prioritisation of challenges, thus helping direct research toward developing next-generation algorithms capable of overcoming current predictability limits.
This kind of benchmarking gives Antiverse a reality check on whether its models are genuinely learning antibody-antigen biology, rather than just overfitting to narrow datasets. It turns abstract “better scores” into concrete evidence that the platform will find more, and better, drug candidates in the lab.
Making predictions with confidence: The need for better metrics
One of the most exciting opportunities for breaking this predictive deadlock lies in confidence metrics for AI predictions. Thanks to the comprehensive datasets and metrics, such as ipTM and ipLDDT, we are able to produce by the models themselves to highlight which parts of the predicted structure are more or less confidently predicted. The models learn these metrics from comparing their predictions to the ‘ground truth’, usually an experimentally-derived structure, during the training process, and these metrics are then predicted alongside the structure itself for cases when the ground truth is unknown.
These confidence metrics are a huge boon to structural analysis, but they can lack precision in distinguishing between highly valuable, accurate complex conformations and red herrings - the plausible but misleading predictions. This limitation poses challenges for researchers who sometimes have to make critical decisions without an experimental structure.
There is strength in diversity, and researchers will commonly generate many structures from a model and indeed use several different structure prediction models when seeking a high-quality structure. This is known to be more likely to find the right structure, but at the cost of needing to find a needle in a haystack. Antiverse developed a novel composite scoring system, integrating multiple confidence parameters, such as interface precision and energetic stability, to better discriminate correct antibody-antigen complexes from incorrect ones. This proposed multi-faceted approach combines multiple angles of structural confidence, providing a more robust and nuanced confidence measure than single metrics alone.
By empowering researchers to prioritise highly reliable AI-generated models, we can therefore reduce experimental trial and error and accelerate the design and optimisation of impactful antibody therapeutics.
What’s next: The evolution in AI drug design
We’ve experienced AI ushering in a new era for biomolecular prediction and drug design, representing an analogue-digital switch in the industry, particularly in understanding antibody-antigen complexes. Advancements like AlphaFold 3 and RFDiffusion offer unprecedented accuracy and dynamic insights, enabling streamlined simulations and smarter antibody design. However, significant roadblocks remain to realising AI's full potential.
Overcoming these hurdles through improved data collection, rigorous benchmarking, advanced iterative refinement techniques, and improved confidence metrics is not merely a scientific aspiration but an ethical imperative. One that holds the potential to unlock novel therapies for previously untreatable diseases.
Newsletters
Receive the latest pharmaceutical news, personalities, education, and career development – weekly to your inbox.
