Designing highly potent compounds using a chemical language model

Abstract

Compound potency prediction is a major task in medicinal chemistry and drug design. Inspired by the concept of activity cliffs (which encode large differences in potency between similar active compounds), we have devised a new methodology for predicting potent compounds from weakly potent input molecules. Therefore, a chemical language model was implemented consisting of a conditional transformer architecture for compound design guided by observed potency differences. The model was evaluated using a newly generated compound test system enabling a rigorous assessment of its performance. It was shown to predict known potent compounds from different activity classes not encountered during training. Moreover, the model was capable of creating highly potent compounds that were structurally distinct from input molecules. It also produced many novel candidate compounds not included in test sets. Taken together, the findings confirmed the ability of the new methodology to generate structurally diverse highly potent compounds.

Knowledge-informed generation of organic structure-directing agents for zeolites using ChatGPT towards human-machine collaborative molecular design

“…This highlights the potential of LLMs in materials synthesis. Bajorath et al have developed a generative chemical language model to predict highly potent compounds from less potent ones as input for drug discovery 21 . Priyakumar et al enabled a transformer-decoder model named MolGPT inspired by GPT models to generate drug-like molecules 22 .…”

Generative design of compounds with desired potency from target protein sequences using a multimodal biochemical language model

“…It aimed at deriving models for predicting potent compounds for targets of interest without specifying numerical potency values across wide ranges, thereby circumventing some of the obstacles associated with benchmark compound potency predictions [ 11 ]. Previously, we derived transformer-based chemical language models (CLMs) for molecular string-to-string conversion conditioned on potency differences between pairs of structural analogues [ 14 , 15 ]. So-called conditional transformer models not only learn conditional probabilities for character sequence translation, but also for other context-dependent rules (such as molecular property constraints).…”

“…So-called conditional transformer models not only learn conditional probabilities for character sequence translation, but also for other context-dependent rules (such as molecular property constraints). Our rules included potency difference thresholds required for the formation of activity cliffs (i.e., analogue pairs having largest potency differences in compound activity classes) [ 14 ] or -in a generalized form- desired potency difference thresholds structural analogues [ 15 ]. In the latter case, transformer models were trained based on large numbers of analogue pairs with greatly varying potency differences.…”

“…In the latter case, transformer models were trained based on large numbers of analogue pairs with greatly varying potency differences. In both instances, conditional transformers consistently reproduced highly potent compounds from activity cliffs or other compound pairs for a variety of activity classes, thus providing proof-of-principle, and generated other structurally diverse candidate compounds [ 14 , 15 ]. On the basis of these findings, we extended this transformer architecture for generative modeling of potent compounds by a meta-learning framework for modeling in low compound data regimes [ 16 ].…”

Anatomy of Potency Predictions Focusing on Structural Analogues with Increasing Potency Differences Including Activity Cliffs

“…Furthermore, ML-based and randomized potency value predictions are often only separated by narrow error margins of 1 to 2 orders of magnitude, which leads to artificially favorable predictions in benchmark settings. At least in part, these limitations result from compound potency and similarity distributions in target-based compound sets (often termed activity classes) that are commonly used for benchmarking. , As a possible alternative, potency predictions might be focused on identifying highly potent compounds, taking into account that it might be difficult to precisely predict their potency values, given that their magnitude is statistically underrepresented in activity classes.…”