Charlotte Rochereau

Ahmed Elnaggar, Hazem Essam, Wafaa Salah-Eldin et al.

As opposed to scaling-up protein language models (PLMs), we seek improving performance via protein-specific optimization. Although the proportionality between the language model size and the richness of its learned representations is validated, we prioritize accessibility and pursue a path of data-efficient, cost-reduced, and knowledge-guided optimization. Through over twenty experiments ranging from masking, architecture, and pre-training data, we derive insights from protein-specific experimentation into building a model that interprets the language of life, optimally. We present Ankh, the first general-purpose PLM trained on Google's TPU-v4 surpassing the state-of-the-art performance with fewer parameters (<10% for pre-training, <7% for inference, and <30% for the embedding dimension). We provide a representative range of structure and function benchmarks where Ankh excels. We further provide a protein variant generation analysis on High-N and One-N input data scales where Ankh succeeds in learning protein evolutionary conservation-mutation trends and introducing functional diversity while retaining key structural-functional characteristics. We dedicate our work to promoting accessibility to research innovation via attainable resources.

Charlotte Rochereau, Benoît Sagot, Emmanuel Dupoux

Neural language models trained with a predictive or masked objective have proven successful at capturing short and long distance syntactic dependencies. Here, we focus on verb argument structure in German, which has the interesting property that verb arguments may appear in a relatively free order in subordinate clauses. Therefore, checking that the verb argument structure is correct cannot be done in a strictly sequential fashion, but rather requires to keep track of the arguments' cases irrespective of their orders. We introduce a new probing methodology based on minimal variation sets and show that both Transformers and LSTM achieve a score substantially better than chance on this test. As humans, they also show graded judgments preferring canonical word orders and plausible case assignments. However, we also found unexpected discrepancies in the strength of these effects, the LSTMs having difficulties rejecting ungrammatical sentences containing frequent argument structure types (double nominatives), and the Transformers tending to overgeneralize, accepting some infrequent word orders or implausible sentences that humans barely accept.