PAL-AI

A maturing oocyte is transcriptionally silent, so which proteins it makes is decided almost entirely by how long each mRNA's poly(A) tail becomes. Two short elements in the 3' untranslated region set that outcome: the cytoplasmic polyadenylation element (CPE, motif UUUUA), read by CPEB1, and the polyadenylation signal (PAS, motif AWUAAA), read by the CPSF complex. Counting those motifs is not enough to say what a given transcript will do — their spacing, copy number, flanking bases and distance from the 3' end all change the result, and a linear model built on motif counts systematically under-predicts the mRNAs whose tails lengthen most.

PAL-AI (poly(A)-tail length AI) replaces that hand-built feature set with a neural network that reads the 3' UTR directly and outputs a single number: the change in tail length over oocyte maturation. It was built by Kehui Xiang and David Bartel at the Whitehead Institute, with appointments at MIT and HHMI, posted as a preprint in November 2024 and published in Nature Communications in 2025. The architecture follows the convolution-plus-recurrence design used for mRNA half-life prediction in Saluki, simplified for a single regression target.

Because the output is a physical quantity measured the same way in every species, the same network can be pointed at frog, mouse and human oocyte data, and at synthetic reporter libraries — which is what turns it from a descriptive model into a variant-scoring tool.

#Key Features

• Sequence in, tail-length change out: input is one-hot ACGU encoding of the last 2000 nt of 3' UTR, with a fifth channel marking 3'-UTR positions against appended coding sequence and an optional sixth channel carrying per-base unpairing probability from RNAfold.
• Multi-output heads for heterogeneous data: PAL-AI-m and PAL-AI-ms attach dataset-specific output layers to a shared trunk, so endogenous mRNAs, injected reporter libraries and different species train together without one cohort degrading the others.
• Interpretable by in silico mutagenesis: scoring every single-nucleotide substitution recovers the CPE and PAS as the highest-impact 6-mers and quantifies the contextual features around them, rather than leaving the learned rules opaque.
• Ships pretrained ensembles: four checkpoint sets — frog endogenous, the reporter library, and the two multi-head variants — are distributed as 10-fold cross-validation ensembles of Keras files, applied to new sequences through a predict mode.

#Technical Details

The base model, PAL-AI-s, stacks an input convolution block, repeated 1D convolution / maxpool / dropout blocks, a GRU recurrent block and a dense head. Trained on 6054 frog (Xenopus laevis) oocyte mRNA isoforms under 10-fold cross-validation, it explains more than 67% of the variance in measured tail-length change (Rs = 0.82, Rp = 0.82). Input extent matters more than extra features: the last 100 nt alone reach average Rp = 0.64, performance saturates at 2000 nt, and adding coding sequence or RNAfold pairing probabilities gives no significant gain. A ResNet-based variant reached comparable accuracy less consistently while needing substantially more parameters, and was not pursued.

PAL-AI-m adds a second output head for an injected N60(LC)-PASmos reporter library of roughly 35,000 3'-UTR variants, holding endogenous accuracy at Rs = 0.82 while raising the reporter library to Rs = 0.81, Rp = 0.88. PAL-AI-ms extends the same trick to three species, reaching Rs = 0.63 on mouse and Rs = 0.68 on human oocyte mRNAs; the frog-only model transfers to mammals more modestly (Rs = 0.57 on mouse, 0.61 on human).

#Applications

The practical payoff is variant interpretation in a regulatory region that most tools ignore. Applied to roughly 8.4 million 3'-UTR variants from the All of Us Research Program — about 0.85 million of them in the last 100 nt — and to gnomAD v4.1 across 76,215 individuals, alleles predicted to disrupt tail lengthening are depleted at higher frequencies, the signature of negative selection, tying cytoplasmic polyadenylation to human female fertility. Predicted tail-length changes also track reported translational efficiency changes (Rs = 0.89), which the authors use to nominate genes important for oocyte maturation, and the same predictions can rank designed 3' UTRs before an injection experiment.

#Impact

PAL-AI is deliberately narrow: one biological process, one scalar output, and compact checkpoints of one to two megabytes each rather than a general-purpose RNA backbone. Within that scope it does something few sequence models manage — it connects a specific cis-regulatory grammar to population-scale human genetics, and it supplies released weights that let others score their own 3' UTRs. Running it takes a small pipeline rather than an API call: a FASTA of 3' UTRs, optional coding sequence and RNAfold output, driven by a YAML config. The code and checkpoints carry no license file, which limits reuse until the authors add one.