In a panel moderated by Brandon Anderson, Pushmeet Kohli (Google DeepMind) and Sal Candido (Biohub) examined why AlphaFold’s achievement was a major advance but far from the final answer on protein folding. They discussed scaling laws, data quality, handcrafted architectures, and what it will take to model dynamics, cells and clinically meaningful biology.
The “bitter lesson” applied to biology
The panel opened by asking whether the AI community’s “bitter lesson” — that scalable methods win — applies to biological data. Sal Candido argued that scaling alone is not sufficient: first you must find the regime where adding compute or data produces consistent gains. Much effort goes into identifying what data and what statistics are needed for a target capability. Practically, researchers often favor problems for which data is already available.
Pushmeet Kohli emphasized that the problem should determine the approach. One should not rigidly identify as “a modeler” or “an experimentalist.” If resources and expertise make model work the best path, invest in modeling; if new data are needed, generate it. In AlphaFold’s case, the available Protein Data Bank (PDB) resources made the modeling route the right choice at the time.
Data quality matters more than sheer volume
Both speakers repeatedly stressed that good, informative data with the right coverage is more valuable than simply more data. Pushmeet noted that curating data is itself an art — duplicating the same data does not advance understanding. Sal gave a concrete example: low‑quality metagenomic sequences, although imperfect and often incomplete, can still boost the performance of protein language models for protein design and understanding, but relying solely on easily generated data can misdirect effort.
Handcrafted architectures versus scale
AlphaFold 2 incorporated many biologically motivated design choices that made it data‑efficient: embedding residue interactions and other inductive biases reduced the amount of data needed. Sal pointed out that scaling is also a craft — scaling models and training corpora requires algorithmic and architectural work. We are not in a post‑Transformer era; rather, Transformers and derivative architectures are evolving and being specialized for biological data.
Why AlphaFold didn’t solve all of protein biology
Five years after AlphaFold 2 was announced, the field has made notable gains, but large questions remain. Pushmeet stressed that AlphaFold reproduced experimentally observed structures from the PDB — a useful and important achievement — but proteins are not static blocks. Their conformations can be disordered, context‑dependent, and distributed. AlphaFold’s successes do not mean we fully understand protein dynamics, ensembles of conformations, or how structure maps to function in all contexts. The model’s behavioral characteristics are well studied, yet its inner mechanisms remain hard to interpret in human terms.
Missing data regimes and next steps
Pushmeet suggested working directly with richer raw data sources, such as cryo‑EM micrographs, which may contain richer statistical and dynamical information than already‑processed structure records. He experimented with this idea but noted it requires substantial effort. Sal emphasized that protein language models likely already encode hidden biological knowledge — structure, function and motion signals — that we have not fully extracted; improving interpretability and probing those representations could unlock more utility.
From single proteins to cells and clinical impact
Both panelists argued for moving beyond models of individual proteins to models of interacting systems, cells and tissues. Sal used an analogy: current models often study a single spoke of a bicycle, but practitioners need models of the whole bicycle (or the vehicle) to design parts that function in context. Regarding clinical impact, Pushmeet said AI is already integrated across drug discovery today, but achieving 10×–100× accelerations in timelines will require deeper biological models and addressing the hard problems the new efforts aim to tackle. Sal urged thinking in terms of 10× breakthroughs rather than incremental 10% improvements to reach ambitious goals such as Biohub’s mission to cure all disease.
Trust, calibration and interpretability
Pushmeet highlighted that well‑calibrated uncertainty estimates are crucial: a model that gives good predictions but misreports its confidence can lead to harmful decisions. He argued that behavioral interpretability (knowing what a model can and cannot do) is essential for safe application, even if full mechanistic interpretability in human terms remains out of reach. He also allowed the possibility that future, larger models might be able to interpret other models better than humans can.
Key practical takeaways
- Define the scientific problem first; choose between modeling, data generation, and expertise accordingly.
- Prioritize good, well‑curated data with appropriate coverage rather than only increasing quantity.
- Use scientific inductive biases to gain data efficiency, but also explore what larger datasets can reveal.
- AlphaFold advanced structure prediction but did not close questions about dynamics, disorder, and system‑level behavior.
- AI is already used in drug discovery; transformative clinical accelerations require deeper biological understanding and potentially 10×‑level innovations.
Conclusion
The panel concluded that AlphaFold is a milestone, not the endpoint. Future progress requires coordinated investment in targeted data collection, model architectures that encode useful inductive biases, new raw data modalities (e.g., cryo‑EM micrographs), and system‑level biological models. Only by combining these efforts can AI deliver the deeper understanding and the dramatic accelerations in medicine that the field hopes for.
(Source: Public panel moderated by Brandon Anderson with Pushmeet Kohli — Google DeepMind and Sal Candido — Biohub; discussion topics and timestamps referenced from the published transcript.)
