Matt Gordon

Describe a protein’s job. AI proposes its structure and sequence. A scientist might want an enzyme that breaks down plastic without knowing what its molecular structure should look like. Many protein-design tools require the scientist to supply some of that structural information first. Dai and colleagues developed Pinal, a generative AI model that proposes proteins from written descriptions of their intended function. Its Transformer-based architecture first generates a compact description of a possible protein structure. A second model uses that description, together with the original request, to generate the amino acid sequence. Structure provides a useful bridge: proteins with quite different sequences can still adopt similar shapes. One experimental test went beyond designing an isolated enzyme. The team requested an H-protein, which carries reaction intermediates between enzymes in a CO₂-fixing pathway. The best design supported 1.7 times the glycine production of the native E. coli protein in a laboratory-reconstituted pathway, without repeated rounds of experimental optimization. Getting there still required selection: 50,000 generated candidates were narrowed to 20 for testing, and five were functional. The written request opens the search. Computational screening and experiments establish which designs actually work. Dai et al., Nature (2026). Journal license. nature.com/articles/s41586-0…

Why autonomous discovery is stalling on hardware, not algorithms The algorithms are ready. The hardware is not. That is the argument of this review, and it reframes where the bottleneck in autonomous discovery actually sits. Richard Canty and Milad Abolhasani trace self-driving labs from 1970s liquid handlers to today's closed-loop platforms. The numbers are not subtle. Artificial Chemist identified eleven distinct perovskite quantum dot formulations in 30 hours using under 210 ml of reagents. AlphaFlow navigated a 40-parameter reaction space with less than 0.2% of the reagents an equivalent manual search would need. Fast-Cat mapped a two-objective Pareto front for rhodium-catalysed hydroformylation in 60 experiments, replacing screens of over 400 runs. So why is this not everywhere? Optimization was never the hard part. AI suites already propose experiments and estimate uncertainty faster than any rig generates data. Meanwhile most platforms are bespoke assemblies of pumps, valves and reactors tuned to one narrow envelope. Retooling from electrochemical screening to organometallic catalysis means downtime and safety revalidation. The consequence for anyone training models on experimental data is sharp. Identical protocols give divergent results across labs because reactor geometry, optical alignment and sensor calibration never make it into the digital record. No amount of model capacity fixes an uncalibrated sensor. The authors argue the real barrier is missing process-complete metadata: provenance of raw materials, how precursors were made, the apparatus itself. Until hardware is described as rigorously as software is shared, reproducibility caps out below algorithmic potential. Two points I keep returning to. First, co-design: putting device-level figures of merit directly in the objective, so the agent optimizes solar cell efficiency rather than synthetic yield, with multifidelity surrogates carrying uncertainty across the synthesis to performance hierarchy. Second, where agents belong. Planning and supervision, yes. Real-time actuation, interlocks and safety-critical control should stay deterministic and formally verifiable. The near-term move is unglamorous and high return: instrument your existing platforms for full process capture before scaling autonomy. The compounding asset is not the robot, it is the dataset another lab can trust. Also worth noting, in a survey of 54 researchers, 74% did not want total automation, with hypothesis generation the function they most wanted to keep. Paper: Canty & Abolhasani, Nature Reviews Chemistry (2026) — journal license | doi.org/10.1038/s41570-026-0…