How our principles helped define the release of Alphafold

Reflections and lessons on sharing one of our biggest breakthroughs with the world

Our mission to resolve intelligence to promote science and benefit humanity comes with important responsibility. To have a positive impact on society, the ethical implications of research and its applications must be positively evaluated in a rigorous and careful manner. We also know that all new technologies can be harmful and take long-term risks seriously. We pioneered the foundation responsible from the start, focusing on responsible governance, research and impact.

This starts by setting clear principles that help realize the benefits of AI (AI) and mitigating its risks and potential negative consequences. Because responsible and pioneering efforts are collective efforts, they have contributed to many AI community standards, including Google, the AI ​​Partnership, and the OECD (Organisation for Economic Cooperation and Development).

Our operating principles have now defined both our commitment to prioritizing broad benefits and areas of research and application that we refuse to pursue. These principles have been at the heart of our decision-making since DeepMind was founded, and continue to be refined as the AI ​​landscape changes and grows. They are designed for our role as a research-driven science company and are consistent with Google’s AI principles.

From principles to practice

The written principles are just part of the puzzle. It is important how they are practiced. This poses an important challenge as complex research is being conducted in the AI ​​frontier. How can researchers predict potential benefits and harms that may occur in the distant future? How can we develop better ethical foresight from a broader perspective? And what is needed to explore difficult questions along with scientific advances in real time to prevent negative consequences?

Over the years, we have developed unique skills and processes for responsible governance, research, and impact, from creating internal toolkits on social issues to supporting efforts to deliberate and foresight across the AI ​​field. To help the Deep Mind Team responsibly empower and get harmed, our Interdisciplinary Institutional Review Board (IRC) meets every two weeks to carefully evaluate Deep Mind Projects, Dissertations and Collaborations.

Responsible pioneers are collective muscles and every project is an opportunity to strengthen our joint skills and understanding. As machine learning researchers, ethicists and safety experts sit together with engineers, security experts, policy experts, and more, the review process was carefully designed to include rotation experts from a wide range of fields. These diverse voices regularly identify ways to expand the benefits of technology, suggesting changes or delays in areas of research and application, and highlight projects that require external consultation.

We have made a lot of progress, but many of these aspects are in unknown territory. We don’t get it right every time, and we’re committed to continuous learning and iteration. I hope sharing the current process will help others who are working on responsible AI. And we hope that we encourage feedback as we continue to learn. So there are detailed reflections and lessons from one of the most complicated and rewarding projects, Alphafold. Our Alphafold AI system solved the challenges of protein structure prediction 50 years ago. And we are excited to see scientists use it to accelerate progress in areas such as sustainability, food security, drug discovery, and basic human biology since it was released to wider communities last year.

Focusing on protein structure prediction

A team of machine learning researchers, biologists and engineers have long considered the problem of protein folding as an incredible and unique opportunity for AI learning systems to make a major impact. In this field, there are standard measures of success or failure, and there are clear boundaries for what AI systems need to support the work of scientists – predicting the three-dimensional structure of proteins. And like many biological systems, protein folding is too complicated to write rules about how it works for anyone. However, AI systems may be able to learn those rules in their own way.

Another important factor was the biennial assessment known as CASP (Critical Evaluation of Protein Structure Prediction). It was founded by Professors John Molt and Professor Kurtzishtov Fidelis. With each gathering, CASP provides a very robust assessment of progress, and participants must predict structures recently discovered by the experiment. The results are a great catalyst for ambitious research and scientific excellence.

Understand practical opportunities and risks

As I was preparing for the CASP assessment in 2020, I realized that Alphafold had shown great potential to solve the challenges I had at hand. We spend a considerable amount of time and effort and ask questions analyzing the actual meaning. How can AlphaFold accelerate biological research and applications? What are the unintended consequences? And how can you share your progress in a responsible way?

This presented a wide range of opportunities and risks to consider. Many of these were areas where we didn’t necessarily have strong expertise. We therefore sought external input from over 30 field leaders, focusing on expertise and background diversity, including biology research, biosecurity, bioethics, and human rights.

Through these discussions, many consistent themes emerged:

Balance broad benefits with risk of harm. We began with a careful thinking about the risk of accidental or intentional harm, including how AlphaFold interacts with both future advancements and existing technologies. Through discussion with external experts, it has been revealed that Alphafold does not significantly facilitate harm to proteins given many practical barriers to this, but future progress must be assessed with caution. Many experts have argued that Alphafold will provide the greatest benefit through free and widespread access, as an advancement associated with many areas of scientific research. Experimental biologists explained how important it is to understand and share appropriately tuned and available reliability metrics for each part of Alphafold’s prediction. By indicating which of Alphafold’s predictions is likely to be accurate, users can estimate that predictions can be trusted and used in their work, and that alternative approaches should be used in their research. We initially considered omitting predictions about whether Alphafold has low confidence or high predictive uncertainty, but the outside experts we consulted have demonstrated why this is particularly important for retaining these predictions in our release and advised on the most useful and transparent way to present this information. There has been a lot of debate about how to avoid accidentally increasing disparities within the scientific community. For example, so-called neglected tropical diseases disproportionately affect poorer parts of the world, often less likely to receive original research funding. We were strongly encouraged to prioritize practical support and actively partner with groups working in these areas.

Establish a release approach

Based on the inputs above, the IRC has approved a set of Alphafold releases to address multiple needs such as:

Peer-reviewed publications and open source code, including two papers in nature, with open source code, to make it easier for researchers to implement and improve Alphafold. Immediately, you can add Google Colab and anyone can enter a protein sequence and receive the predicted structure instead of running the open source code itself. As a public institution, EMBL-EBI allows anyone to look at protein structure predictions as easily as Google searches. The initial release contained predicted shapes of all proteins in the human body, and the latest update included predicted structures of almost every cataloged protein known to science. This total of over 200 million structures are all freely available on the EMBL-EBI website, with open access licenses, and will be accompanied by support resources such as webinars on the interpretation of these structures. It builds 3D visualizations into a database, with high confidence and prominent labeling of low-compact areas, aiming for a prominent purpose leading up to forecasting. We also designed the database to be as accessible as possible, taking into account the needs of people with a lack of color vision, for example. It guides deeper partnerships with research groups working in underfunded areas, including neglected diseases and topics that are essential to global health. These include DNDI (drugs for neglected disease initiatives), DNDI, which is researching Chagas disease and leishmaniasis, and Enzyme Innovation Centre, which is developing plastic erodontic enzymes that help reduce plastic waste in the environment. Our growing public team continues to work on these partnerships to support more collaborations in the future.

How we are building this work

Since its first release, hundreds of thousands of people in over 190 countries have visited the Alphafold Protein Structure Database and have used AlphaFold open source code since its launch. It is an honor to hear how Alphafold’s predictions accelerate important scientific efforts and work to tell some of these stories in the unfolded projects. So far, we have not recognized any misuse or harm related to Alphafold, but we continue to pay close attention to this.

Although Alphafold was more complicated than most deep research projects, we use elements of what we learned and incorporate this into other releases.

We are building this work as follows:

It increases the range of input from external experts at every stage of the process and explores the mechanisms of participatory ethics on a larger scale. Beyond general projects and breakthroughs, it encourages understanding of AI in biology and develops a stronger view of opportunities and risks over time.

Like our research, this is a continuous learning process. The development of AI for widespread benefits is a community effort that goes far beyond deep minds.

We make every effort to be aware of how we partner with others and how we are responsible for the future and how much we put in there.