Schrödinger speeds up molecular discovery 4x with AlphaEvolve
Quick answer
Schrödinger uses Google DeepMind's AlphaEvolve to evolve faster algorithms, achieving a 4x speedup in molecular discovery. Drug screening drops from months to days.
Computational chemistry has always been a swamp of trade-offs: fast but fuzzy classical force fields, or precise but painfully slow quantum methods. Schrödinger, a veteran in scientific software, decided to wade deeper and found a way to glide through the muck with AlphaEvolve, an evolutionary AI coding agent from Google DeepMind.
The bottleneck in the bayou
Schrödinger’s machine-learned force fields (MLFFs) were hitting a wall. Two algorithms—neighbor list computation and Ewald summation—were the gators blocking the channel. The Ewald sum, in particular, was a slow for-loop in PyTorch, dragging down training and inference. They needed a faster paddle.
AlphaEvolve to the rescue
AlphaEvolve is like a capybara that keeps refining its path through the reeds. It iteratively generates and tests code, evolving toward the most efficient algorithm. For Schrödinger, it replaced the Ewald sum’s for-loops with parallel batch matrix multiplication—a batched approach that let the GPU flex its muscles.
Rigorous testing in the swamp
Schrödinger didn’t just trust the new code blindly. They put it through a gauntlet:
- Inverse time: Throughput improved from a baseline of 7.9 to nearly 30.
- Functional correctness: All evolved programs passed regression tests on complex systems like disordered water.
- Success rate: Jumped from under 1% to over 60%—meaning most evolved code was both correct and faster.
4x speedup: from months to days
The result? A 4x acceleration in both MLFF training and inference. That means researchers can screen molecular candidates in days instead of months. Gabriel Marques, Schrödinger’s ML technical lead, says: “AlphaEvolve allows us to explore larger chemical spaces faster and more efficiently than ever before.” This directly impacts drug discovery, catalyst design, and materials development.
What’s next?
Schrödinger plans to turn AlphaEvolve loose on custom GPU kernels, testing if AI-generated code can outpace human engineering. If you’re curious, dive into the full technical paper or chat with the Google Cloud AI team.
For more on cloud platforms that power such breakthroughs, check out our Google Cloud review.
Original announcement published on Google Cloud.