🧵1/n LLMs significantly improve Evolutionary Algorithms for molecular discovery! For 18 different molecular optimization tasks, we demonstrate how to achieve SOTA performance by incorporating different LLMs! Learn more in our new paper! Website: Code)

🧵2/n Evolutionary Algorithms (EAs), general-purpose methods to traverse chemical spaces via random mutations and crossovers, demonstrate good performance for molecular optimization. However, finding an "optimal" molecule via random operations might be prohibitively expensive.

🧵3/n We propose the use of chemistry-aware LLMs inside mutation and crossover operations to propose new molecules. We demonstrate that it yields better objective values in a fewer number of calls than baseline crossovers and mutations, which randomly edit molecules.

🧵4/n For both commercial and open-source LLMs, we validate the performance & sample efficiency of MolLEO on the popular practical molecular optimization (PMO) benchmark and show that we beat the top baselines in PMO (25 total!) on 12 single-objective molecular optimization tasks

🧵5/n We also test our models on 6 more challenging structure-based drug design and multi-objective optimization tasks and show that we beat baseline models. Because the perturbations of MolLEO are better than random edits...

🧵6/n ..., we hope that it will become more feasible to use oracles that are more expensive but more accurate in representing the target property, generating candidates that are more promising for real-life applications.

🧵7/n This was a great collaboration with (* co-first) @Haorui_Wang123*, @martoskreto*, @chertianser, @WenhaoGao1, @konglingkai_AI, @felix_s_k, @chenru_duan, @yuchen_zhuang, @yue___yu, @Zhu_Yanqiao, († co-senior) @YuanqiD†, @A_Aspuru_Guzik†, @k_neklyudov†, @chaozhangcs†

@martoskreto @chertianser @WenhaoGao1 @konglingkai_AI @felix_s_k @chenru_duan @yuchen_zhuang @yue___yu @Zhu_Yanqiao @A_Aspuru_Guzik @k_neklyudov @chaozhangcs Tagging some people who might be interested in this work @gabepgomes @pschwllr @andrewwhite01 @JeffGuo__ @WengongJin @drecmb @json_yim @guolin_ke @austinjtripp @ShuiwangJi @KyleWSwanson @charlieharris01 @rneschneuing @igashov @kchonyc @hengjinlp @ysu_nlp @zhiling_zheng @TianfanFu

@martoskreto @chertianser @WenhaoGao1 @konglingkai_AI @felix_s_k @chenru_duan @yuchen_zhuang @yue___yu @Zhu_Yanqiao @A_Aspuru_Guzik @k_neklyudov @chaozhangcs @gabepgomes @pschwllr @andrewwhite01 @JeffGuo__ @WengongJin @drecmb @json_yim @guolin_ke @austinjtripp @ShuiwangJi @KyleWSwanson @charlieharris01 @rneschneuing @igashov @kchonyc @hengjinlp @ysu_nlp @zhiling_zheng @KexinHuang5 @GrantRotskoff @CanalJiananLi @Masatosh @HannesStaerk @luost26 @chence_shi @Ryan__Rhys @ZhuoranQ @olexandr @jchodera @bwood_m @lucky_pattanaik @wpwalters @RGBLabMIT @marwinsegler @gklambauer @rbhar90 @jmhernandez233

Cool work @YuanqiD, just had the time to look at it now. It makes sense that LLMs might be able to suggest better-mutations than hand-crafted heuristics. On the other hand, it's probably much slower, so it might struggle in settings with a high eval budget. Did you try this?

Thanks for reading it! The rate-limiting factor is the call to GPT API call here if we use an open-source LM it’s pretty fast. Usually the bottleneck is the oracle evaluation I would say, e.g. even calculate docking score is way slower then GPT API call.