AI agents beating human made tools at molecular discovery

Intology's automated research system Locus developed scientific AI models that outperformed state-of-the-art expert solutions across molecular scoring and design tasks, the company said in a thread on X. Intology tested Locus on training molecular property prediction models for a proprietary task, fine-tuning Boltz-2 to predict small-molecule potency, and building antibody design workflows around BoltzGen, Boltz-2, ESMC and ESMFold2, reporting that Locus beat state-of-the-art baselines and LLM agents, generalized better with limited data, and found more sample-efficient and diverse solutions. Intology said Locus's antibody design workflows were far more sample efficient and cheaper than existing design workflows.

Intology on X: "Locus sets the frontier for automating molecular design Our automated research system Locus develops scientific AI models that outperform SOTA expert solutions Unlocking the potential of scientific AI requires leveraging proprietary data. Locus achieves this autonomously 🧵👇 --… / X Intology on X: "Locus sets the frontier for automating molecular design Our automated research system Locus develops scientific AI models that outperform SOTA expert solutions Unlocking the potential of scientific AI requires leveraging proprietary data. Locus achieves this autonomously 🧵👇 -- We test Locus across a number of tasks related to molecular scoring and design: training molecular property prediction models for a proprietary task, fine-tuning Boltz-2 to predict small-molecule potency, and building antibody design workflows around BoltzGen, Boltz-2, ESMC, and ESMFold2. Each of these settings are real-world AI for science use cases, where tasks are often out-of-distribution for existing foundation models and limited in data, while featuring expensive and noisy evaluators. Compared to state-of-the-art baselines and LLM agents set to optimize these tasks, Locus achieves higher performance. Additionally, Locus generalizes better despite limited data, finds more sample efficient solutions, and explores a higher diversity of solutions." Locus sets the frontier for automating molecular design Our automated research system Locus develops scientific AI models that outperform SOTA expert solutions Unlocking the potential of scientific AI requires leveraging proprietary data. Locus achieves this autonomously 🧵👇 -- We test Locus across a number of tasks related to molecular scoring and design: training molecular property prediction models for a proprietary task, fine-tuning Boltz-2 to predict small-molecule potency, and building antibody design workflows around BoltzGen, Boltz-2, ESMC, and ESMFold2. Each of these settings are real-world AI for science use cases, where tasks are often out-of-distribution for existing foundation models and limited in data, while featuring expensive and noisy evaluators. Compared to state-of-the-art baselines and LLM agents set to optimize these tasks, Locus achieves higher performance. Additionally, Locus generalizes better despite limited data, finds more sample efficient solutions, and explores a higher diversity of solutions. Locus sets the frontier for automating molecular design Our automated research system Locus develops scientific AI models that outperform SOTA expert solutions Unlocking the potential of scientific AI requires leveraging proprietary data. Locus achieves this autonomously 🧵👇 -- We test Locus across a number of tasks related to molecular scoring and design: training molecular property prediction models for a proprietary task, fine-tuning Boltz-2 to predict small-molecule potency, and building antibody design workflows around BoltzGen, Boltz-2, ESMC, and ESMFold2. Each of these settings are real-world AI for science use cases, where tasks are often out-of-distribution for existing foundation models and limited in data, while featuring expensive and noisy evaluators. Compared to state-of-the-art baselines and LLM agents set to optimize these tasks, Locus achieves higher performance. Additionally, Locus generalizes better despite limited data, finds more sample efficient solutions, and explores a higher diversity of solutions. AI R&D problems in scientific settings are often heavily data-constrained. For molecular property prediction and small-molecule potency prediction, Locus overcomes these problems’ limited data initialization to train models that generalize better. Locus designs solutions thatShow more Many in silico antibody design workflows generate large numbers of binder candidates, followed by expensive scoring and strict filtering steps. We find that Locus is able to devise solutions that are far more sample efficient, and cheaper, than existing design workflows. The solution space for problems in AI R&D for science is typically vast. For potency prediction, we find that Locus performs a more diverse set of experiments, and in doing so, achieves better performance than existing LLM agent baselines. Locus performs strategic experimentation that makes effective use of large-scale compute. Across various scaling axes, such as research compute, model size, and generated synthetic data, Locus discovers molecular property prediction solutions that perform better than baselines. Previously, we shared Locus’ state-of-the-art results in automated LLM post-training, demonstrating Locus’ ability to scale in an expensive domain. Locus’ post-training results were featured in Jack Clark’s blog below: Through our research program, we have improved Locus’ capabilities in dealing with large hypothesis spaces, evaluation costs, and experimental noise. From language model post-training to scientific AI, we see the challenges in our mission of scaling experimentation acrossShow more As always, a joy to read the technical report. Very interesting to see the role of synthetic data in all this. But you also offer strong insights on evals, etc. Thanks for sharing.