Can predictive tools like AlphaFold be trusted in multi-million dollar drug discovery programs?

In 2024 the Nobel Prize in Chemistry was awarded to David Baker for computational protein design (namely RFDiffusion) and Demis Hassabis and John Jumper for protein structure prediction (namely AlphaFold2). These days, AI-supported drug discovery seems to be on everyone’s mind. However, there still is a lot of discord on how it is perceived. Will it really be the magic bullet to fix everything? Or is it completely overhyped and not be trusted?

How much can we really trust tools like AlphaFold?

To really understand the impact of AlphaFold in particular on the field of protein engineering, let’s go back a few years.

In 2002, I remember standing in my first year PhD confirmation meeting when a senior academic asked me “Why would you go to all the effort of solving a protein structure when you can just model it?!”.

At the time I was outraged by the question for many reasons. Primarily, because at the time I would have had to model the protein domain from a different species and a different protein, which meant I would have gone forth and generated a suite of publications and potentially drug discovery work based on assumptions. The potential to waste time and money was incredibly high. Instead, I solved the protein structure experimentally with NMR and was quite pleased with myself and felt rather vindicated, when, at the end of 3 years, I could demonstrate biologically significant differences between the ‘actual’ structure I had solved, and the ‘modelled’.

Like my peers in the field, throughout my career as a structural biologist, I learnt the physics involved in solving protein structures, how proteins like to fold, what interactions are favourable or disfavourable and details about protein dynamics and energy landscapes. It may sound crazy, but over time, and with training, it is possible to just look at a structure and predict the impact a mutation is likely to have on the protein structure.

Fast-forward to 2025, asking this same question of ‘why don’t you just model it?’ is quite a different scenario.

Initiatives such as AlphaFold make is easier than ever to model protein structures and even interactions with DNA, RNA or small molecules.

But the killer question still remains - how do we know the model is correct?

I think of AlphaFold and ChatGPT along the same lines. We all know that ChatGPT can be a bit of fun and save us time, but can produce great sounding, non-factual, nonsense. We know it is nonsense based on prior knowledge of the topic or understanding of language. With this in mind, would you trust ChatGPT to write your PhD thesis or an important email to your boss unsupervised? Probably not.

So, the big question remains, can you base a multi-million dollar drug discovery or protein engineering campaign based on the output of AlphaFold?

Well actually, yes, it is possible, but with one important caveat.

I would trust AlphaFold’s output, if the AlphaFold structures are being generated by a trained structural or computational biologist who has a deep understanding of favourable protein interactions, energy landscapes, dynamics etc. A trained eye can quickly look at an AI generated model and tell, if it is nonsense or if there is value within the predicted model. To this day no AI can replicate that knowledge and training. But machine learning models like AlphaFold can be a fantastic tool to shortcut previously lengthy experimental discovery.

A few years ago, I moved out of academic research and found myself working as a program manager with a structural biology team at a biotechnology company based in the United States. The team were charged with engineering various protein targets under the constraints of limited funding and time. To overcome these constraints, the team were able to take known structures and further refine AlphaFold models using RosettaFold2 and RosettaFold2NA from the Baker lab, to model the impact of single point mutations through to more significant changes that opened up freedom-to-operate or IP space. In one project we modelled structures with AlphaFold and for every x number of models we would also solve the crystal structure, to align the known with the predicted. This gave us extra confidence that the machine learning models were generating structures that had true value.

Ultimately, this saved the program significant time and money by reducing the resourcing required to generate and characterise multiple recombinant proteins. Rather, the team tested directly in cell lines to make meaningful interpretations of the results and move on to identification of lead molecules.

It is not clear at this stage whether tools such as AlphaFold will advance to a point where they replace the art of structure determination, but they are meaningful tools to accelerate drug discovery and protein engineering programs. Given the current state of research funding in both academic and industry environments, any tools that can provide meaningful data that enables the program to move quicker and save money is a powerful one.

Yet, only trust structures generated from tools such as AlphaFold if they are interpreted by a trained structural biologist who can assess the value of the end product (and always cross-check the content of any ChatGPT generated texts you send out….).